cognitive computing

July/August 2017

Vol. 19, No. 4

Te c h n o l o g y S o l u t i o n s f o r t h e E n t e r p r i s e

COGNITIVE COMPUTING Defining Digital Intelligence, p. 3 Blockchain for the Internet of Things, p. 68

www.computer.org/itpro

Technology Solutions for the Enterprise

CALL FOR PAPERS Cyberthreats and Security Publication: May/June 2018

Submission deadline: 1 October 2017

The landscape of cybersecurity continues to change rapidly, with new threats and attack paths that seemingly appear almost daily. At the same time, advances on the technology side have accelerated. In particular, the Internet of Things (IoT) and miniature cyber-physical systems are bringing information security concerns to every aspect of life, including clothing, kitchen appliances, and automobiles. The corporate world is also facing challenges in how to apply new technology such as the IoT and distributed ledger systems—technology that even many computer security professionals don’t fully understand. This special issue of IT Professional seeks to provide readers with an overview of current issues and advances in information and computer security. We seek high-quality contributions from industry, government, business, and academia that present recent developments in IT security, showcase successfully deployed security solutions, or discuss challenging security issues that deserve further study. Topics of interest include, but are not limited to, the following:

• • • • • • • • • •

Security in the IoT and cyber-physical systems Blockchain and distributed ledger systems Authentication and access control Physical and behavioral biometrics Applications of cryptography Cloud and distributed system security Computer forensics Privacy preservation in data mining and data analytics Intrusion-detection systems Malware detection

• • • • • • • • •

Mobile app and device security Network security Legalities and ethics associated with security and privacy policies Protection of personal information Security architecture Usability and human factors issues Wearable device security Web application security Security of online transactions

Submissions Feature articles should be no longer than 4,200 words and have no more than 20 references (with tables and figures counting as 300 words). Illustrations are welcome. For author guidelines, including sample articles, see www.computer.org/portal/web/peerreviewmagazines/acitpro. Submit your article at https://mc.manuscriptcentral.com/itpro-cs.

Questions? For more information, please contact the guest editors: Tim Weil, [email protected] Morris Chang, [email protected] Rick Kuhn, [email protected]

computer.org/ITPro

3

Cognitive Computing IN THIS ISSUE

16 Guest Editors’ Introduction: Cognitive Computing Haluk Demirkan, Seth Earley, and Robert R. Harmon

21 Enterprise Cognitive Computing Applications: Opportunities and Challenges Monideepa Tarafdar, Cynthia M. Beath, and Jeanne W. Ross

In this article, the authors provide an overview of cognitive computing applications for the enterprise. In particular, they classify opportunities for developing enterprise cognitive computing applications and describe implementation challenges.

28

36

45

Cognitive Compliance for Financial Regulations

Building a Cognitive Application Using Watson DeepQA

Arvind Agarwal, Balaji Ganesan, Ankush Gupta, Nitisha Jain, Hima P. Karanam, Arun Kumar, Nishtha Madaan, Vitobha Munigala, and Srikanth G. Tamilselvam

Christopher Asakiewicz, Edward A. Stohr, Shrey Mahajan, and Lalitkumar Pandey

Cognitive Communication in Rail Transit: Awareness, Adaption, and Reasoning

Compliance with regulations is getting increasingly hard because of complex documents and the sheer volume of regulatory change. The authors’ Cogpliance platform uses a cognitive approach to help users achieve regulatory compliance.

Academic advisors assist students in academic, professional, and personal matters. The authors’ cognitive advising system uses IBM Watson’s cognitive intelligence to identify question categories and then answer questions accordingly.

Cheng Wu and Yiming Wang

Applying cognitive radio in railway wireless communication systems is a cutting-edge research field. The authors’ rail cognitive radio model uses a Bayesian network with channel contextual features to probabilistically infer the likelihood of spectrum accessibility.

July/August 2017 Volume 19, Number 4

Te c h n o l o g y S o l u t i o n s f o r t h e E n t e r p r i s e

COLUMNS AND DEPARTMENTS

3 From the Editors What Is Digital Intelligence? Sunil Mithas and F. Warren McFarlan

7 IT Trends How Much to Trust Artificial Intelligence? George Hurlburt

12 IT in Emerging Markets Teaching Computer Sciences in Morocco: An Overview Yousef Farhaoui

63 Data Analytics The Problem With AI Seth Earley

68 Securing IT Can Blockchain Strengthen the Internet of Things? Nir Kshetri

55 Cognitive Gaming Wei Cai, Yuanfang Chi, and Victor C.M. Leung

As intelligent networked computing becomes pervasive, an emerging trend is to apply a cognitive computing paradigm to video game design and development. This article describes the concept of cognitive gaming and discusses its enabling architecture.

20 CS Info Inside Front Cover

Call for Papers: Cyberthreats and Security

On the Web: computer.org/itpro For more information on computing topics, visit the Computer Society Digital Library at www.computer.org/csdl.

FROM THE EDITORS

What Is Digital Intelligence? Sunil Mithas, Robert H. Smith School of Business F. Warren McFarlan, Harvard Business School

D

igital intelligence—the ability to understand and utilize the power of IT to our advantage, is becoming a critical skill for all managers in today’s economy,1 partly because of significant changes in the business environment in the last 50 years. The IT world has changed remarkably since the 1960s, when IT was largely a back-office function focused on automation and reducing costs, was not well integrated with business functions, and did not matter as much strategically.2 Much IT work was done in-house at that time by IT departments, and there were few external service providers.

IT Evolution and Digital Intelligence A lot has changed since then. Around 2010, upward of 50 percent of firms’ capital spending was going to IT, compared to less than 10–15 percent back in the 1960s. IT matters a lot today because of its revenue role and strategic potential; it is much more 1520-9202/17/$33.00 © 2017 IEEE

integrated with business functions, with many more options for business and functional units to configure IT themselves rather than rely on an internal IT department. The Strategic Impact Grid, introduced by F. Warren McFarlan in 1983, has been a useful tool in assessing IT changes over time and preparing a firm to respond to them (see the “About the Strategic Impact Grid” sidebar). Despite significant progress on the technology front since then and a manifold rise in the digitization of business operations, products, and services, many organizations fail to synchronize their IT and business strategies.3 The tension between the standards and controls that IT departments champion and the fast responses that businesses need still remains. McFarlan’s examples of companies such as William Carter, Li & Fung, Otis, Kodak, and Uber explain why digital intelligence should be a part of boardroom discussions in today’s information economy.4

Published by the IEEE Computer Society

Digital intelligence is more than being able to work with computers or IT; it involves an understanding of how to synchronize business and IT strategies, govern IT, and execute IT projects and enterprise systems. We next discuss some key elements of digital intelligence to gain competitive advantage and sustain it in the rapidly changing digital age.

Synchronize Business and IT Strategies Synchronizing business and IT strategies requires that managers envision IT, integrate IT with strategy, and explore new IT on a continuous basis. We prefer the word “synchronization” to “alignment” because alignment implies that either IT or strategy is preordained, whereas synchronization implies a continuous, two-way interaction between IT and strategy. In other words, synchronization better captures a mindset that is open to new possibilities enabled by technology and at the same time ensures that the use of IT is consistent with strategic needs.

computer.org/ITPro

3

FROM THE EDITORS About the Strategic Impact Grid The Strategic Impact Grid highlights the impact of IT on a firm’s competitiveness.1,2 The vertical dimension represents the firm’s exposure to real losses as a result of IT vulnerabilities or security breaches. The horizontal dimension represents the overall impact of the firm’s application development portfolio on its competitiveness (see Figure A). The grid can be used to illustrate how different firms, or parts of firms, are affected in different ways by IT. It can also facilitate a dialog among business and IT professionals regarding the position of the company as a whole or that of a firm’s business units or IT applications. The grid was originally used to assist with IT planning efforts, and more recently its use was extended to shape IT governance and spending decisions at the board level.

High Need for reliable IT, or avoid loss as a result of IT vulnerabilities

Factory

Strategic

Support

Turnaround

Low Low

Need for new IT, or gain competitive advantage using IT

High

Figure A. McFarlan’s Strategic Impact Grid.

References 1. F.W. McFarlan, J.L. McKenney, and P. Pyburn, “The Information Archipelago—Plotting a Course,” Harvard Business Rev., Jan./Feb. 1983, pp. 145–156.

Envision IT First, all managers need to have a vision for embracing IT’s potential and realize that IT can have a significant—even makeor-break—impact on an organization. When we say “all managers,” we refer to both business and IT managers. CEMEX, Zara, Capital One, and Amazon all demonstrate how IT and information-based capabilities helped firms create sustainable value in widely differing industries and ways. Conversely, companies such as FoxMeyer Drug, Blockbuster, and Borders had significant difficulties managing IT and dealing with IT-enabled transformations. Xerox’s failure to capitalize on the innovations of its PARC lab demonstrates the importance of this point.

Integrate IT Second, IT should be an integral part of any strategy discussion. It is the responsibility of senior leaders to develop inclusive but robust 4

IT Pro July/August 2017

2. R.L. Nolan and F.W. McFarlan, “Information Technology and the Board of Directors,” Harvard Business Rev., vol. 83, no. 10, 2005, pp. 96–106.

strategy development processes that are informed by the capabilities of IT but also stretch these capabilities for long‐term organizational sustainability. Senior leaders must understand the duality inherent in IT before they can choose an appropriate digital business strategy and an offensive or defensive posture. The dualities of IT refers to the idea that technology can be both sustaining and disruptive; enable adaptation to and shape competition; provide new competitive advantages, even if such advantages are highly visible and replicable; enable aggregation (horizontally) and disaggregation (vertically); and create tremendous digital uncertainties even while providing tools to manage them. Leaders need to question their conventional strategy concepts, which focus on tradeoffs, because IT can, at times, help overcome these trade offs altogether. For example, IT can help firms pursue both revenue growth and cost reduction, or

higher quality and lower costs— combinations that might not initially be visualized. An easy way to understand IT’s role in creating competitive advantage is to remember the acronym ADROIT. This acronym parses the value created by IT into six components: • Add revenues. IT can help to add revenues through inorganic or organic means that might involve increasing sales to existing or new customers through existing or new channels by selling existing or new products. • Differentiate. IT can help to differentiate or enhance nonprice attributes such as perceived quality or convenience that often increase customer satisfaction. • Reduce costs. IT can help a company reduce its overall costs through selective outsourcing while also investing in internal capabilities. Benchmarking on IT costs alone can be

counterproductive if IT investments can help to reduce nonIT costs substantially. • Optimize risks. IT can help to optimize risks (not necessarily reduce them). Managers must try to reduce downside risks from not investing in IT by engaging in counterfactual reasoning. One way to reduce downside risk is to split IT projects into must-do and may-do components and manage IT projects as having real options to resolve technical or market uncertainties. Managers should consider the effect of IT investments on intangibles such as customer satisfaction that can reduce downside or idiosyncratic risk. • Innovate. IT can help firms pursue IT-embodied or IT-enabled innovations by making R&D more effective and scalable, and by using innovation from outside the firm, as Lego, P&G (through Connect + Develop), and SAP have tried to do. • Transform business models and processes. IT can help transform business models and processes by replacing or complementing atoms with bits. Dealing with transformations requires that managers calibrate their response to the triggers that are causing transformation; protect their current revenue streams to the extent possible while finding ways to develop or grow new ones; and develop capabilities for dealing with change and transformation without being blinded by the rush to outsource key capabilities that might be necessary for future competitive advantage. This acronym can help managers think about IT’s role in a comprehensive manner to synchronize IT and strategy.

IT Pro Welcomes New Editorial Board Members G.R. Gangadharan is an associate professor in the Institute for Development & Research in Banking Technology (IDRBT), Hyderabad, India. His research interests include energy informatics, cloud computing, and enterprise information systems. Gangadharan received a PhD degree in information and communication technology from the University of Trento, Italy. He is a senior member of IEEE and ACM. For further details, see http:// www.idrbt.ac.in/grganga.html. Contact him at [email protected]. Charalampos Z. Patrikakis is an associate professor in the Department of Electronics Engineering at Piraeus University of Applied Sciences. He has 20+ years’ experience in international research projects, having participated in more than 32 national, European, and international programs, in 16 of which he has been involved as a technical/scientific coordinator or principal researcher. Patrikakis has more than 100 publications in book chapters, international journals, and conferences, and has two contributions to national legislation. He is a senior member of IEEE and a counselor of the IEEE student department at Piraeus University of Applied Sciences. Contact him at [email protected].

Explore New IT Third, managers and entrepreneurs need to repeatedly scan new technologies to assess their significance and use them to stay relevant and transform their organizations. This should not be a one-time exercise; these actions should become part of a manager’s routine because exploration of newer technologies can often facilitate new and more effective ways of doing business. Experimentation to gain insight into applications, technology, and change is key. To avoid making sense of newer technologies on an ongoing basis is to avoid change; this rarely pays off, as the failures of Kodak and Borders demonstrate. Just scanning new technologies and recognizing their significance is not enough. Leadership matters

when it comes to transforming organizations. Although frameworks or methodologies such as Baldrige Criteria, Design Thinking, or Agile can act as triggers, unless leaders empower organizations and monitor the progress made on these opportunities for improvement, they are unlikely to achieve success. Transformations, whether technology-enabled or otherwise, need leadership, management continuity, rigor, discipline, and eschewing of the pursuit of management fads. Sustained performance requires persistence, the refining of technologies, and their integration with incentive systems and business processes to yield desired outcomes. More than relying on the charisma of leaders, organizations must focus on creating processes that focus on long-term

computer.org/ITPro

5

FROM THE EDITORS thinking, where continuous improvement, scanning of newer technologies, and agile transformations to stay relevant become routine.

Govern IT Fourth, formulating strategy is rarely enough; deployment is equally important. Successful deployment needs attention to the governance of IT decisions, departments, dollars, and delivery in a way that is synchronized with the company’s strategy to avoid the “two-culture problem” that business and IT often struggle with. There are no simple solutions, and because of their structures, staff capabilities, and so on, different organizations will come to different answers regarding governance configuration. Not tackling governance issues in a systematic way or following through on them is an abdication of managerial responsibility because solid governance provides a platform for integrating various initiatives, just as an operating system allows a variety of applications to be built by leveraging a common platform. The governance failures at firms such as Blockbuster show that there’s significant room for improvement in governance.

Execute IT Fifth, IT projects need to be managed carefully, with attention to technology evolution, firm strategy, business processes, business value, and bottom-line benefit, while ensuring buy-in and business sponsorship whenever possible. It is the responsibility of business and IT managers to be aware of technology evolution, make informed decisions regarding technology upgrades, and understand how they should help to adopt, diffuse, and exploit IT 6


systems. Managers must also understand what risk-management strategies they should adopt when it comes to implementing various enterprise projects consistent with their broader strategy. Finally, it is not just IT systems, artificial intelligence, or big data that on their own can provide desirable business outcomes. Analytics and metrics matter; organizations suffer if they do not have metrics, but they also suffer if they focus on the wrong or narrow metrics to measure success. It is the job of managers to ask critical questions related to data definitions, do some upfront thinking about how the data will be analyzed and used to inform business decisions, and then ensure that, over time, such data-driven decision making becomes the norm.

I

n summary, managers need to care about IT because IT-induced technological advances affect most industries and functional areas. Smart managers can use IT as a lever to enhance their personal and professional competitive advantage. Because IT is so embedded with business processes and new initiatives, sooner or later, most managers will be involved in an IT project. Given how risky and important these projects are, we must invest the necessary effort to understand how to manage them to ensure success. Together, these are good reasons for managers to invest in their own digital intelligence and that of the people or organizations they supervise.

Acknowledgments This article is extracted and adapted from Digital Intelligence, by Sunil Mithas, with a foreword from F. Warren McFarlan (published by Finerplanet, and Penguin India for the Indian subcontinent).

We thank San Murugesan for helpful comments and suggestions.

References 1. S. Mithas, Digital Intelligence: What Every Smart Manager Must Have for Success in an Information Age, Finerplanet, 2016; http://a.co/hxsPEJv. 2. J. Dearden and F.W. McFarlan, Management Information Systems: Text and Cases, Richard D. Irwin, 1966. 3. R. Roberts and J. Sikes, “McKinsey Global Survey Results: IT’s Unmet Potential,” McKinsey Q., Nov. 2008, pp. 1–9. 4. F.W. McFarlan, “IT and Management 1960–2020,” Digital Intelligence: What Every Smart Manager Must Have for Success in an Information Age, Finerplanet, 2016, foreword; http://a.co /hxsPEJv.

Sunil Mithas is a professor at the Robert H. Smith School of Business at the University of Maryland. His research interests include strategies for managing innovation and excellence for corporate transformation, focusing on the role of technology and other intangibles. Mithas is the author of the books Digital Intelligence: What Every Smart Manager Must Have for Success in an Information Age (Finerplanet, 2016) and Dancing Elephants and Leaping Jaguars: How to Excel, Innovate, and Transform Your Organization the Tata Way (2014). He is a member of IT Professional’s editorial board. Contact him at sunil.mithas@ gmail.com. F. Warren McFarlan is the Albert H. Gordon Professor Emeritus of Business Administration at the Harvard Business School. He currently teaches in several short Executive Education programs. McFarlan has had a significant role in introducing materials on management information systems to all major programs at the Harvard Business School since the first course on the subject was offered in 1962. Contact him at [email protected].

IT TRENDS EDITOR: Irena Bojanova, NIST, [email protected]

How Much to Trust Artificial Intelligence? George Hurlburt, STEMCorp

T

here has been a great deal of recent buzz about the rather dated notion of artificial intelligence (AI). AI surrounds us, involving numerous applications ranging from Google search, to Uber or Lyft ride-summoning, to airline pricing, to Alexa or Siri. To some, AI is a form of salvation, ultimately improving quality of life while infusing innovation across myriad established industries. Others, however, sound dire warnings that we will all soon be totally subjugated to superior machine intelligence. AI is typically, but no longer always, software dominant, and software is prone to vulnerabilities. Given this, how do we know that the AI itself is sufficiently reliable to do its job, or—put more succinctly— how much should we trust the outcomes generated by AI?

Risks of Misplaced Trust Consider the case of self-driving cars. Elements of AI come into play in growing numbers of self-driving car autopilot 1520-9202/17/$33.00 © 2017 IEEE

regimes. This results in vehicles that obey the rules of the road, except when they do not. Such was the case when a motor vehicle in autonomous mode broadsided a turning truck in Florida, killing its “driver.” The accident was ultimately attributed to driver error, as the autonomous controls were deemed to be performing within their design envelope. The avoidance system design at the time required that the radar and visual systems agree before evasive action would be engaged. Evidence suggests, however, that the visual system encountered glare from the white truck turning against bright sunlight. This system neither perceived nor responded to the looming hazard. At impact, however, other evidence implicated the “driver,” who was watching a Harry Potter movie. The driver, evidently overconfident of the autopilot, did not actively monitor its behavior and failed to override it, despite an estimated seven-second visible risk of collision.1 The design assurance level was established,


but the driver failed to appreciate that his autopilot still required his full, undivided attention. In this rare case, misplaced trust in an AI-based system turned deadly.

Establishing a Bar for Trust AI advancement is indeed impressive. DARPA, sponsor of early successful autonomous vehicle competitions, completed the Cyber Grand Challenge (CGC) competition in late 2016. The CGC established that machines, acting alone, could play an established live hacker’s game known as Capture the Flag. Here, a “flag” is hidden in code, and the hacker’s job is to exploit vulnerabilities to reach and compromise an opponent’s flag. The CGC offered a $2 million prize to the winning team that most successfully competed in the game. The final CGC round pitted seven machines against one another on a common closed network without any human intervention. The machines had to identify vulnerabilities in an opponent’s system, fix them on their own system, and exploit them in

computer.org/ITPro

7

IT TRENDS opponents’ systems to capture the flag. Team Mayhem from Carnegie Mellon University was declared the winner.2 John Launchbury, director of DARPA’s Information Innovation Office, characterizes the type of AI associated with the CGC as handcrafted knowledge. Emerging from early expert systems, this technology remains vital to the advancement of modern AI. In handcrafted knowledge, systems reason against elaborate, manually defined rule sets. This type of AI has strength in reasoning but is limited in forms of perception. However, it possesses no ability to learn or perform abstraction.3 While building confidence that future reasoning AI can indeed rapidly diagnose and repair software vulnerabilities, it is important to note that the CGC was intentionally limited in scope. The open source operating system extension was simplified for purposes of the competition,4 and known malware instances were implanted as watered-down versions of their real-life counterparts.5 This intentionally eased the development burden, permitted a uniform basis for competitive evaluation, and reduced the risk of releasing competitors’ software into the larger networked world without requiring significant modification. The use of “dirty tricks” to defeat an opponent in the game adds yet another, darker dimension. Although the ability to reengineer code to rapidly isolate and fix vulnerabilities is good, it is quite another thing to turn these vulnerabilities into opportunities that efficiently exploit other code. Some fear that if such a capability were to be unleashed and grow out of control, it could become a form of “supercode”—both exempt from common vulnerabilities and 8


Figure 1. Some prevalent AI machine learning algorithms.

capable of harnessing the same vulnerabilities to assume control over others’ networks, including the growing and potentially vulnerable Internet of Things (IoT). This concern prompted the Electronic Frontier Foundation to call for a “moral code” among AI developers to limit reasoning systems to perform in a trustworthy fashion.4

Machine Learning Ups the Trust Ante Launchbury ascribes the term statistical learning to what he deems the second wave of AI. Here, perception and learning are strong, but the technology lacks any ability to perform reasoning and abstraction. While statistically impressive, machine learning periodically produces individually unreliable results, often manifesting as bizarre outliers. Machine learning can also be skewed over time by tainted training data.3 Given that not all AI learning yields predictable outcomes, leading to the reality that

AI systems could go awry in unexpected ways, effectively defining the level of trust in AI based tools becomes a high hurdle.6 At its core, AI is a high-order construct. In practice, numerous loosely federated practices and algorithms appear to compose most AI instances—often crossing many topical domains. Indeed, AI extends well beyond computer science to include domains such as neuroscience, linguistics, mathematics, statistics, physics, psychology, physiology, network science, ethics, and many others. Figure 1 depicts a less than fully inclusive list of algorithms that underlie second-wave AI phenomena, often collectively known as machine learning. This myriad of potential underlying algorithms and methods available to achieve some state of machine learning raises some significant trust issues, especially for those involved in software testing as an established means to assure trust. When the AI becomes associated with mission criticality, as

is increasingly the case, the tester must establish the basis for multiple factors, such as programmatic consistency, repeatability, penetrability, applied path tracing, or identifiable systemic failure modes. The nontrivial question of what is the most appropriate AI algorithm goes as far back as 1976.3 The everyday AI practitioner faces perplexing issues regarding which is the right algorithm to use to suit the desired AI design. Given an intended outcome, which algorithm is

One high-level AI test assesses the ability to correctly recognize and classify an image. In some instances, this test has surpassed human capability to make such assessments. For example, the Labeled Faces in the Wild (LFW) dataset supports facial recognition with some 13,000 images to train and calibrate facial recognition machine learning tools using either neural nets or deep learning. The new automated AI image recognition tools can statistically outperform human facial

The everyday AI practitioner faces perplexing issues regarding which is the right algorithm to use to suit the desired AI design. the most accurate? Which is the most efficient? Which is the most straightforward to implement in the anticipated environment? Which one holds the greatest potential for the least corruption over time? Which ones are the most familiar and thus the most likely to be engaged? Is the design based on some form of centrality, distributed agents, or even swarming software agency? How is this all to be tested? These questions suggest that necessary design tradeoffs exist between a wide range of alternative AI-related algorithms and techniques. The fact that such alternative approaches to AI exist at all suggests that most AI architectures are far from consistent or cohesive. Worse, a high degree of contextually-based customization is required for both reasoning and learning systems. This, of course, extends to AI testing, because each algorithm and its custom implementation brings its own unique deep testing challenges, even at the unit level.

recognition capability using this dataset.7 The task at hand, however, is fundamentally perceptual in nature. These tasks functionally discriminate through mathematically correlated geometric patterns but stop short of any form of higher-order cognitive reasoning. Moreover, while it compares selective recognition accuracy against human ability, other mission-critical aspects of the underlying code base remain unchecked under this test.

Beyond the Code Testing machine learning becomes further complicated as extensive datasets are required to “train” the AI in a learning environment. Not only should the AI code be shown to be flawless, but the data used in the training should theoretically bear the highest pedigree. In the real world, however, datasets often tend to be unbalanced, sparse, inconsistent, and often inaccurate, if not totally corrupt. Figure 2 suggests that information often results from resolving ambiguity. Even

under controlled conditions, significant differences result between the use of single or multiple wellvalidated datasets used to train and test classifiers. Thus, even controlled testing for classifiers can become highly complicated and must be approached carefully.8 Other trust-related factors extend well beyond code. Because coding is simultaneously a creative act and somewhat of a syntactic science, it is subject to some degree of interpretation. It is feasible that a coder can inject either intentional or unintentional cultural or personal bias into the resulting AI code. Consider the case of the coder who creates a highly accurate facial recognition routine but neglects to consider skin pigmentation as a deciding factor among the recognition criteria. This action could skew the results away from features otherwise reinforced by skin color. Conversely, the rates of recidivism among criminals skews some AI-based prison release decisions along racial lines. This means that some incarcerated individuals stand a better statistical chance of gaining early release than others—regardless of prevailing circumstances.9 Semantic inconsistency can further jeopardize the neutrality of AI code, especially if natural language processing or idiomatic speech recognition are involved. Some suggest that all IT careers are now cybersecurity careers.10 This too has a huge implication for the field of AI development and its implementation. The question of “who knows what the machine knew and when it knew it” becomes significant from a cybersecurity standpoint. What a machine learns is often not readily observable, but rather lies deeply encoded. This not only affects newly internalized data, but—in

computer.org/ITPro

9

IT TRENDS the IoT—these data can trip decision triggers to enliven actuators that translate the “learning” into some sort of action. Lacking concrete stimulus identity and pedigree, the overall AI-sparked IoT stimulus-response mechanism becomes equally uncertain. Nonetheless, the resulting actions in mission-critical systems require rigorous validation.

The Third Wave Launchbury foresees the need for a yet-to-be-perfected third wave of AI, which he names contextual adaptation. This technology, requiring much more work, brings together strengths in perception, learning, and reasoning and supports a significantly heightened level of cross-domain abstraction.3 The 2017 Ontology Summit, aptly entitled “AI, Learning, Reasoning, and Ontologies,” concluded in May 2017. Reinforcing Launchbury’s observation, the draft summit communique concluded that, to date, most AI approaches, including machine learning tools, operate at a subsymbolic level using computational techniques that do not approximate human thought. Although great progress has been achieved in many forms of AI, the full treatment of knowledge representation at the symbolic level awaits maturity (bit.ly/2qMN0it). Correspondingly, the utility of ontology as a formal semantic organizing tool offers only limited advantages to AI and its ultimate test environment. The semantic network involves graph representations of knowledge in the form of nodes and arcs. It provides a way to understand and visualize relationships between symbols, often represented by active words, which convey varying meanings when 10


Figure 2. Information provenance can often be unclear.

viewed in context. AI, largely subsymbolic today, will need to deal with applied semantics in a far more formal sense to achieve third-wave status. Under such circumstances, AI becomes nonlinear, in which cause and effect are increasingly decoupled via multiple execution threads. This leads to the establishment of complex adaptive systems (CAS), which tend to adhere to and be influenced by nonlinear network behavior. In a CAS, new behaviors emerge based on environmental circumstance over time. Here, there can be multiple self-organizing paths leading to success or failure, all triggered by highly diversified nodes and arcs that can come, grow, shrink, and go over time. Such networks defy traditional recursive unit testing when composed using embedded software, which is interrelated to data. This is because in a CAS, the whole often becomes far more than merely the sum of the parts.11 Rather, new approaches,

emerging from applied network science, offer a better means of assessing dynamic AI behavior that emerges over time. This becomes increasingly true as the temporal metrics associated with graph theory become better understood as a means of describing dynamic behaviors that fail to follow linear paths to achieve some desired effect.12

U

ntil some reliable methodology is adopted for the assessment of assured trust within AI, the watchword must be caution. Any tendency to put blind faith in what in effect remains largely untrusted technology can lead to misleading and sometimes dangerous conclusions.

References 1. N.E. Boudette, “Tesla’s Self-Driving System Cleared in Deadly Crash,” New York Times, 19 Jan. 2017. 2. D. Coldewey, “Carnegie Mellon’s Mayhem AI Takes Home $2 Million

from DARPA’s Cyber Grand Challenge,” TechCrunch, 5 Aug. 2016; tcrn.ch/2aM3iS7. 3. J. Launchbury, “A DARPA Perspective on Artificial Intelligence,” DARPAtv, 15 Feb. 2017; www.youtube .com/watch?v5-O01G3tSYpU. 4. N. Cardozo, P. Eckersley, and J. Gillula, “Does DARPA’s Cyber Grand Challenge Need a Safety Protocol?” Electronic Frontier Foundation, 4 Aug. 2016; bit.ly/2aPxRXc. 5. A. Nordrum, “Autonomous Security Bots Seek and Destroy Software Bugs in DARPA Cyber Grand Challenge,” IEEE Spectrum, Aug. 2016; bit.ly/2arLOcR. 6. S. Jontz, “Cyber Network, Heal Thyself,” Signal, 1 Apr. 2017; bit .ly/2o0ZCVe.

7. A. Jacob, “Forget the Turing Test— There Are Better Ways of Judging AI,” New Scientist, 21 Sept. 2015; bit. ly/1MoMUnF. 8. J. Demsar, “Statistical Comparisons of Classifiers over Multiple Data Sets,” J. Machine Learning Research, vol. 7, 2006, pp. 1–30. 9. H. Reese, “Bias in Machine Learning, and How to Stop It,” TechRepublic, 18 Nov. 2016; tek.io /2gcqFrI. 10. C. Mims, “All IT Jobs Are Cybersecurity Jobs Now,” Wall Street J., 17 May 2017; on.wsj.com/2qH5VP2. 11. P. Erdi, Complexity Explained, SpringerVerlag, 2008. 12. N. Masuda and R. Lambiotte, A Guide to Temporal Networks, World Scientific Publishing, 2016.

George Hurlburt is chief scientist at STEMCorp, a nonprofit that works to further economic development via adoption of network science and to advance autonomous technologies as useful tools for human use. He is engaged in dynamic graph-based Internet of Things architecture. Hurlburt is on the editorial board of IT Professional and is a member of the board of governors of the Southern Maryland Higher Education Center. Contact him at [email protected].

Read your subscriptions through the myCS publications portal at

http://mycs.computer.org

CALL FOR STANDARDS AWARD NOMINATIONS IEEE COMPUTER SOCIETY HANS K ARLSSON STANDARDS AWARD A plaque and $2,000 honorarium is presented in recognition of outstanding skills and dedication to diplomacy, team facilitation, and joint achievement in the development or promotion of standards in the computer industry where individual aspirations, corporate competition, and organizational rivalry could otherwise be counter to the benefit of society. NOMINATE A COLLEAGUE FOR THIS AWARD!

DUE: 1 OCTOBER 2017 • Requires 3 endorsements. • Self-nominations are not accepted. • Do not need IEEE or IEEE Computer Society membership to apply.

Submit your nomination electronically: awards.computer.org | Questions: [email protected]

computer.org/ITPro

11

IT IN EMERGING MARKETS EDITOR: Gustavo Rossi, Universidad Nacional de La Plata, [email protected]

Teaching Computer Sciences in Morocco An Overview

Yousef Farhaoui, Moulay Ismail University, Morocco

I

n our current technologybased society, an individual who is ignorant of the notions of computer basics will find it increasingly difficult to find a place. As a result, digital literacy has proven to be a requirement of human development, and IT is positioning itself as a fundamental component of civic education. Indeed, the governments of developed countries explicitly recognize that neglecting the importance of computer studies in education could jeopardize the future of their countries. As a developing country, Morocco has engaged in this international trend and has oriented its educational policy in this direction. For example, with the aim of improving the quality of education and training, Morocco’s National Education and Training Charter highlights the undeniable pedagogical contribution of new information and communication as a subject: “Computer sciences promote the learning of technical knowledge

12


and contributes to the development of the learner’s rigor and autonomy. They also enable him or her to develop his or her creative potential and communication and collaboration skills.”

Computer Sciences: A School Subject Nowadays, computer sciences are recognized as a school subject in virtually all educational systems worldwide. However, the debate about the knowledge and computing skills that today’s young people are supposed to learn at school is still relevant. Although there is unanimity about the need to introduce computer sciences into the school curriculum, the approaches adopted and the experiments carried out depend on each country’s social or cultural context. These differences result in both a multiplicity of conceptions about which fields of computer studies a school must offer and which implementation approaches should be adopted. Thus, a divergence already appears at the level of the nomenclature and, more


precisely, at the level of the vocabulary arsenal used to designate the teaching of computer sciences at school.

Teaching Computer Sciences in Morocco Since 2007, computer sciences have been a compulsory school subject at middle school as well as in the first year of high school. At these levels, pupils’ ages range from 11 to 16 years old, representing the most crucial stage in the school curriculum. Indeed, during this period, pupils begin their effective taxation in society as independent individuals and start to develop critical thinking, analysis, and reasoning skills to understand and elaborate their explanations of the world. The proposed education must therefore be able to guide pupils to make coherent and global representations of their environment and the world in which they live. In pursuit of this objective, curricula, particularly those relating to IT, need to be developed.

1520-9202/17/$33.00 © 2017 IEEE

Teaching Computer Sciences in Primary Education One of the announced aims of primary education in the Moroccan educational system is to enable students to use ICT. However, after reading through Morocco’s official guidelines, we can see that the term “IT” is cited only once. These guidelines provide ambiguous guidance on the use of digital resources in teaching various subjects. In practice, ICT use remains limited, and it is often utilized only in the initiatives of a minority of teachers. In fact, this minority is made up partly of teachers involved in innovative educational projects or projects designed to produce digital content.

Teaching Computer Sciences in Middle School As already mentioned, computer sciences are incorporated into the middle school curriculum. During the three years of middle school, this school subject is meant to teach pupils to use ICT to research, process, and communicate information. However, starting from the second year, students are introduced to basic programming and problem-solving skills.

Teaching Computer Sciences in High School In the first year of qualifying secondary education, Morocco’s curriculum focuses on using ICT and introducing algorithms and programming. It is organized into four modules, each of which contains a set of coherent units. The concepts’ level of depth varies according to each sector. In analyzing the curriculum for the first year of high school, we can notice the following: First, the hourly weight of the Algorithmic and Programming module constitutes 26 percent of the program

(16 hours). Notions such as repetition and procedures (included in the middle school curriculum) are absent from Grade 1 curriculum, as are notions of tables and other data structures. Given this lack of notions, pupils cannot adopt an efficient algorithmic approach to solve a given problem, a very lowlevel result in the development of algorithmic thinking and the ability to solve problems. Second, the content of the other modules is taught at middle school, so students will not find any added value. Third, there is a lack of continuity or matching between the computer program at middle school and in the first year of qualifying high school. Knowledge and know-how do not follow a precise order in terms of difficulties, the level of notions, and the themes tackled. In addition, even if it is assumed that the student has not studied computers at middle school, the proposed program does not take into account young people’s digital culture, which can make computer courses uninteresting to them. On the other hand, the circulars for school re-enrollment insist on teaching computer sciences without resorting to group work, so that students in the core curriculum generally receive one hour per week instead of two. Accordingly, the aims are either halved or not fully achieved.

ICT for Learning The Moroccan educational system, which is part of the international trend toward deploying ICT in the learning process, has funded and promoted several initiatives in this direction. As a case in point, Morocco’s educational governmental body launched the Generalization of Information and Communication Technologies in

Education (GENIE) program in 2006. This program aims at training teachers and administrative staff to ensure the use of technological tools and develop a digital culture in schools. Thus, among the program’s focuses are the creation of multimedia rooms with an Internet connection, teacher training to integrate ICT into their practices, and the development of digital education resources. Despite the scale and ambition of this program, it has encountered various obstacles and difficulties, including inadequate computer equipment and digital educational resources adapted to the programs taught; a lack of skills and skilled human resources in ICT; resistance to change and teachers’ lack of confidence in technology; failing management of ICT projects at the local level; and a lack of long-term vision in the process of integrating ICT in schools. Despite these problems and difficulties, the initiatives for digital education committed to by official bodies are increasingly numerous: • The IT Academy and Microsoft Office Specialist Certification (MOS) program, launched in October 2013, aims to train and certify more than 300,000 educational and administrative frameworks for the MOS program over a four-year period. • The CITI project, which was launched in February 2008, has developed several applications for academic support during the college cycle in science subjects. • The ITQANE e-learning training program gave rise to a distance training scheme targeting the education profession in various regional centers for the education and training trades (CRMEF).

computer.org/ITPro

13

IT IN EMERGING MARKETS • The Massar project launched during the 2013–2014 school year aims to integrate ICT and strengthen governance in school system management. • The Collab platform was developed for in-service training for national education staff. • The E-taaloum project was established to support students from primary to secondary edu cation.

Computer Studies and ICT Integration Survey The relevance of computer education can well be assessed through student opinions. For this reason, and to have a clear view of students’ perceptions of computer sciences, I administered a questionnaire consisting of three parts: • Profile (personal information). The survey’s target audience was asked to identify their gender and the technological supports they used. • Attitude toward computer sciences. This section aimed to provide an overview of the students’ attitudes and perceptions of computer sciences as a school subject. • Attitude toward integrating ICT in science sessions. The objective of this section was to collect information on usage. Students were asked whether they had attended a session that used ICT during their academic journey, and to clarify the subject matter and answer questions about the conditions of use.

Student Attitudes toward Computer Sciences The data obtained (100 valid responses were analyzed) show that only 3 percent of the students 14


surveyed benefited from computer courses in the first year of middle school, 80 percent in the second year, and 0 percent in the third year. Among those who attended computer courses, • 85 percent believe that computer sciences as a school subject are not necessary because the student had acquired or can acquire computer skills outside the school. • 80 percent believe that the computer courses studied do not really help students develop the manipulation skills of word processing, spreadsheet, and presentation software. • With respect to the Internet and the use of ICT, the totality claimed that they had not taken any training on the use of Internet services, the production of multimedia documents, or the creation of webpages. • Finally, 80 percent claim not to have had programming courses. As for the remaining 20 percent, they say they attended programming sessions in LOGO without really knowing they were programming. This data, although from a relatively small sample, corroborates the findings of surveys in other parts of the country. Examples include the remarks made by pedagogical inspectors in high school computing, who qualify on the ground, and the studies carried out on computer science teaching in Morocco.

Student Attitudes toward ICT Use in the Classroom By asking questions about students’ attitudes toward the integration of ICT in scientific sessions at middle school, I sought to know whether ICT is indeed

integrated in the classroom. The results obtained show that ICT is used much more in biology and geology than in other school subjects. In fact, 90 students out of the 100 surveyed said they took ICT in biology and geology, whereas this percentage drops to 50 percent in physics and 2 percent in mathematics. The findings also show that the use of ICT is limited to using the video projector to display PowerPoint presentations, watch video and view photos, and sometimes present the content of the course that students must then copy to their notebooks. Only 10 percent of students said they had carried out scientific experiments using computer simulations. During such classes, students followed the presentation without using a computer, which limits the aims of the GENIE project.

T

he starting point of this article is the commonly shared reality that computing is an integral part of our current world, and digital education is a necessity for a developing society. In the school environment, computer sciences are currently omnipresent in all school subjects, providing students with the tools to master technologies that are constantly developing. It can therefore be a teaching material and a tool for teaching or learning other school subjects. To this end, educators are unanimous about conferring the status of school subject on computer sciences, and the calls to reform this teaching are multiplying. These calls point out that the IT approach must go beyond literacy, office automation, and using devices. It must be approached at a higher level—that of a science

of modeling, reasoning, analysis, problem solving, and creativity. Yousef Farhaoui is an associate professor at Moulay Ismail University, Faculty of Sciences and Techniques, Department

of Computer Science, Asia Team, Errachidia, Morocco. His research interests include learning, e-learning, computer security, big data analytics, and business intelligence. Farhaoui has three books in computer science. He is a coordinator

and member of the organizing committee and also a member of the scientific committee of several international congresses, and is a member of various international associations. Contact him at [email protected].

CONFERENCES

in the Palm of Your Hand

IEEE Computer Society’s Conference Publishing Services (CPS) is now offering conference program mobile apps! Let your attendees have their conference schedule, conference information, and paper listings in the palm of their hands. The conference program mobile app works for Android devices, iPhone, iPad, and the Kindle Fire.

For more information please contact [email protected]

computer.org/ITPro

15

GUEST EDITORS’ INTRODUCTION

Cognitive Computing

Haluk Demirkan, University of Washington Tacoma Seth Earley, Earley Information Science Robert R. Harmon, Portland State University

T

here is no doubt that computers are increasingly capable of doing things that in the past, only humans could do. Today, smart machines are becoming more like humans through their ability to recognize voices, process natural language, and learn. They learn by interacting with the physical world through devices that enable them to see, hear, smell, and touch, as well as through mobility and motor control. In some cases, they do a much faster and better job than humans at recognizing patterns, performing rule-based analysis on very large amounts of data, and solving both structured and unstructured problems.

Smart Machines Cognitive computing refers to smart systems that learn at scale, reason with purpose, and interact with humans and other smart systems naturally. Rather than being explicitly programmed, these systems learn and reason from their interactions with us and from their experiences with their 16


environment. They are made possible by advances in a number of scientific fields over the past half-century, and are different in important ways from the information systems that preceded them. Many researchers, practitioners, and educators are working diligently on this topic. Collectively, enterprises are spending billions of dollars to make these cognitive systems smarter. We already interact with services enabled by cognitive computing, such as Apple Siri, IBM Watson, Microsoft Cortana, Google Go, and Amazon Echo. But what is this hype about “smart machines becoming the most disruptive in the history of IT”? In fact, this is not a new thing. One of the enabling technologies, artificial intelligence, was defined by John McCarthy at a conference held at Dartmouth in 1956 as “the science and engineering of making intelligent machines, especially intelligent computer programs.”1 It has taken 60 years to begin to realize this vision through a combination of advances in processing power,


1520-9202/17/$33.00 © 2017 IEEE

development of new algorithms, and the explosion of data with which to train machines and enable intelligent functionality. Cognitive computing—smart machines—has great potential to reduce costs, increase efficiency, and improve outcomes by performing routine tasks and analyzing large amounts of data. The development and adoption of cognitive computing is a process and the result of a combination of machine learning algorithms and traditional knowledge engineering extended by applying breakthroughs in emerging technology. The overall goal of cognitive computing is to increase the productivity and creativity (decision making, connectivity, innovation, and augmentation) of individuals and organizations.2

In This Issue This special issue of IT Professional seeks to provide readers with an overview of the current topics and practices related to cognitive computing. In addition, it looks into the future for IT professionals as these technologies become indispensable and essential enablers of product and service development and the creation of new markets. The authors’ contributions to this collection of articles have implications for cognitive computing that go beyond the immediate application settings on which they report. They also showcase the application of an array of research methods, including surveys, experiments, and design science. We next look at each of the articles to identify the main thrust of the authors’ investigations and the relevant findings for theory and practice.

Enterprise Cognitive Computing This issue opens with “Enterprise Cognitive Computing Applications: Opportunities and Challenges,” by Monideepa Tarafdar, Cynthia M. Beath, and Jeanne W. Ross. In this article, the authors provide an overview of cognitive computing applications for the enterprise. In particular, they provide a classification of opportunities for developing enterprise cognitive computing (ECC) applications and describe challenges in implementing them. In the 33 user organizations—which represented a broad range of industries distributed across North America, Europe, and the Asia-Pacific region—the authors studied a total of 51 initiatives and use cases of

ECC applications, about 70 percent of which were either in production or in a working proof-ofconcept stage. The authors describe two unique capabilities that characterize ECC applications. The first is the processing and making sense of increasingly large and growing volumes of data. ECC applications can handle data that, although available directly to humans, can be overwhelming because of the sheer size of the corpus. Examples include medical literature on a given topic and case law of a country or state. Moreover, ECC applications can adjust and adapt their models based on new data to return increasingly accurate results. The second capability is the automation of tasks that typically require human interpretation, such as queries in a call center that can be interpreted by natural language processing (NLP) or image recognition that can identify individuals. The authors also discovered that the largest number of use cases involved ECC applications that enhanced an organization’s operational excellence. These included sophisticated search and retrieval from very large corpora of technical information, such as legal and accounting data, financial laws and regulations, and medical literature; predictive maintenance of machines; product classification; and fraud detection. ECC applications that intend to delight customers by either offering superior products and services or fostering customer loyalty and engagement formed the second largest number of use cases. The third group was ECC applications that help create a superior employee experience. Finally, the authors found that four challenges are particularly important. To take advantage of ECC’s ability to process massive amounts of data and improve efficiency and effectiveness, business leaders must choose the right tools, make sure needed data is available to those tools, consistently supervise the applications, and appropriately allocate responsibilities between humans and machines.

Cognitive Compliance in Finance This issue’s theme continues with an article by Arvind Agarwal, Balaji Ganesan, Ankush Gupta, Nitisha Jain, Hima P. Karanam, Arun Kumar, Nishtha Madaan, Vitobha Munigala, and Srikanth G. Tamilselvam, “Cognitive Compliance

computer.org/ITPro

17


for Financial Regulations.” Financial institutions are faced with rapidly changing regulatory policies and an ever-growing number of regulations. Compliance tasks are further complicated by the complex language used in regulatory documents,3 which forces banks to hire domain experts whose primary job is to identify relevant regulations and accordingly introduce or suggest changes to specific internal controls to remain compliant. This demand for skilled labor has led to growing operational costs in recent years, sometimes accounting for more than 10 percent of total operational expenses.4 To address the volume, velocity, variety, and complexity of regulations, banks are increasingly seeking technological help. The authors propose a system that aids compliance officers in understanding. They describe the architecture of Cogpliance, a cognitive compliance platform that uses machine learning, information retrieval, and NLP techniques coupled with a novel user experience design to provide an end-to-end system. The architecture consists of multiple phases—that is, data ingestion, preprocessing, data enrichment, data store, supported services, and applications. The authors also describe two applications built on this architecture: regulatory change tracking and a knowledge-graph-driven question-answering system. They show through two case studies how these applications can aid in increased compliance. Practitioners can replicate these solutions by developing necessary components such as document ingestion, data enrichment, and data storage. Organizations can apply this architecture to improve the efficiency of compliance tasks.

An Intelligent Academic Advisor In the third article, entitled “Building a Cognitive Application Using Watson DeepQA,” Christopher Asakiewicz, Edward A. Stohr, Shrey Mahajan, and Lalitkumar Pandey aim to address the gap between the technical and broad overview literature on cognitive computing by developing guidelines that can be used by nontechnical analysts and developers for the development of cognitive applications. They present a cognitive advising system that answers questions related to their university and its programs. This system uses IBM Watson’s machine learning 18


algorithms to identify the question category and provide an appropriate response. To implement the system, they developed a four-step development process as illustrated in their article. Their guidelines were developed through their experience in developing a frequently asked questions (FAQs) application for a graduate analytics program. The system answers questions from a number of perspectives: potential applicants to the program, newly admitted students, current students, faculty members, and business professionals. One of their goals was to demonstrate how easy it can be to develop a cognitive application. The application addresses a relatively simple domain with a narrow corpus of knowledge to train the system. Building cognitive business applications requires a combination of both business focus and technical skill. The business focus helps the development team to best identify those opportunities suitable for cognitive development. Technical skills (such as Python and Java programming) help in developing and deploying cognitive solutions. Having a comprehensive platform such as Bluemix on which to build cognitive applications, as well as a set of template applications (with associated code and documentation contained in GitHub repositories) that can be extended and enhanced, allowed the development team to get their prototype running quickly.

Railway Cognitive Communication Cheng Wu and Yiming Wang aim to demonstrate a railway cognitive communication model to assess the key issues of spectrum scarcity and uncertainty in motion in their article, “Cognitive Communication in Rail Transit: Awareness, Adaption, and Reasoning.” With the continuous construction of high-speed rail in recent years, wireless communication technology plays an increasingly vital role in the daily operation, safety monitoring, and efficiency improvement of highspeed rail. The authors tested their cognitive application model by using a network simulation of a wireless communication framework. They analyzed the main limitations of rail traffic wireless communication caused by the high-speed motion of trains. To solve the uncertainty of spectrum accessibility, they proposed adding a cognitive reasoning module with a Bayesian network as the

core engine. They also discuss performance features related to spectrum accessibility.

Cognitive Gaming In the final article of this issue, “Cognitive Gaming,” Wei Cai, Yuanfang Chi, and Victor C.M. Leung describe the concept of cognitive gaming with a proposed enabling architecture. Video game virtual experiences are emerging as the earliest adopters of these types of cognitive capability. After defining cognitive gaming and presenting an architectural framework, the authors discuss examples and opportunities in cognitive game content generation. Two cognitive game system optimization approaches with their inherent challenges are discussed. Building cognitive services specifically for game content is a challenging task, given that it requires complex mathematic models and large amounts of training data for learning algorithms. Collecting players’ information for game content generation is identified as an emerging trend in cognitive gaming. This is an extension of personalization approaches used in targeted advertising, which tracks consumer behavior and delivers appropriate content. The rich information available across the web makes it possible for a machine to quickly derive game players’ characteristics. With games as dynamic human-computer interaction systems, player behaviors and performances in different game scenarios pose enormous challenges in game design and optimization. These issues become more complicated in multiplayer games because a variety of factors create more uncertainties. The self-adaptation feature of cognitive computing makes it a powerful approach to address these issues.

A

s the articles in this special issue illustrate, cognitive computing encompasses the processing and making sense of large volumes of data to enhance operational excellence, improve compliance with regulations, develop academic advising systems, optimize spectrum sharing, and optimize complex game system design. Although the articles here provide an overview of a range of cognitive computing applications, the discipline and its applications are growing rapidly. Cognitive

computing is cross-disciplinary in nature and focuses on methodologies and systems that can implement autonomous computational intelligence in applications as varied as expert systems, robotics, autonomous vehicles, medical diagnostics, machine vision, translation, employee performance evaluations, planning and scheduling, marketing analytics, remote maintenance monitoring, and many others. We hope the articles presented here will motivate readers to continue the journey of discovery about what constitutes one of the most important domains for IT professionals.

Acknowledgments We know the importance of having to start somewhere to get new ideas moving, and of finding the appropriate collaborators to make initial steps and advances in new knowledge possible. We thank editor in chief San Murugesan for the vision he shared with us and for getting the discussion rolling. This issue received several submissions which underwent a two-cycle “review and revise” process before we were able to select the final articles. We would especially like to acknowledge the anonymous reviewers who so generously offered their time, effort, and helpful insights for us to make these hard choices and for helping us develop the final product. Finally, we thank the authors whose work was accepted and those whose research we were not able to publish in this edition. We hope the reviewers’ comments will strengthen their future success. We look forward to the “next generation” of IT submissions to IT Professional.

References 1. J. McCarthy, “What Is Artificial Intelligence?” Computer Science Dept., Stanford Univ., 12 Nov. 2007, pp. 1–15; stanford.io/2lSo373. 2. H. Demirkan, J.C. Spohrer, and J.J. Welser, “Digital Innovation and Strategic Transformation,” IT Professional, vol. 18, no. 6, 2016, pp. 14–18. 3. T.D. Breaux and A.I. Antón, “Mining Rule Semantics to Understand Legislative Compliance,” Proc. 2005 ACM Workshop on Privacy in the Electronic Society, 2005, pp. 51–54. 4. The Cost of Compliance: Global Hedge Fund Survey, AIMA, MFA, and KPMG, Oct. 2013; bit.ly/2s5oWwB.

Haluk Demirkan is a Milgard professor of service innovation and business analytics and director of the Center for Business Analytics at the Milgard School of Business, University of Washington Tacoma. His research interests

computer.org/ITPro

19


include driving service innovation and transformation with analytics and cognitive computing for value co-creation and outcomes. Demirkan is a cofounder of and member of the board of directors for the International Society of Service Innovation Professionals, and has more than 150 publications. He received a PhD in information systems and operations management from the University of Florida. Contact him at [email protected]. Seth Earley is CEO of Earley Information Science (www. earley.com). His interests include knowledge strategy, information architecture, search-based applications, and accelerating the value of analytics solutions. Earley has worked with a diverse roster of Fortune 1000 companies, helping them achieve higher levels of operating performance; is founder of the Boston Knowledge Management Forum; and is a former adjunct professor at Northeastern

University, where he taught graduate courses in knowledge management infrastructure and e-business strategy. Contact him at [email protected]. Robert R. Harmon is a professor of marketing and service innovation and Cameron Research Fellow at Portland State University. His research interests include cloud-enabled service innovation and the development and implementation of asymmetric business strategies. Harmon recently coauthored Cloud as a Service: Understanding the Service Innovation Ecosystem (Apress/Springer, 2016) with Enrique Castro-Leon of Intel. He is a member of the IEEE Computer Society and the International Society of Service Innovation Professionals, and serves as an associate editor in chief for IT Professional. He received a PhD in marketing and information systems from Arizona State University. Contact him at [email protected].

EXECUTIVE STAFF PURPOSE: The IEEE Computer Society is the world’s largest association of computing professionals and is the leading provider of technical information in the field. MEMBERSHIP: Members receive the monthly magazine Computer, discounts, and opportunities to serve (all activities are led by volunteer members). Membership is open to all IEEE members, affiliate society members, and others interested in the computer field. OMBUDSMAN: Email [email protected]. COMPUTER SOCIETY WEBSITE: www.computer.org

Next Board Meeting: 12–13 November 2017, Phoenix, AZ, USA

EXECUTIVE COMMITTEE President: Jean-Luc Gaudiot; President-Elect: Hironori Kasahara; Past President: Roger U. Fujii; Secretary: Forrest Shull; First VP, Treasurer: David Lomet; Second VP, Publications: Gregory T. Byrd; VP, Member & Geographic Activities: Cecilia Metra; VP, Professional & Educational Activities: Andy T. Chen; VP, Standards Activities: Jon Rosdahl; VP, Technical & Conference Activities: Hausi A. Müller; 2017–2018 IEEE Director & Delegate Division VIII: Dejan S. Milojičić; 2016–2017 IEEE Director & Delegate Division V: Harold Javid; 2017 IEEE Director-Elect & Delegate Division V-Elect: John W. Walz

BOARD OF GOVERNORS Term Expiring 2017: Alfredo Benso, Sy-Yen Kuo, Ming C. Lin, Fabrizio Lombardi, Hausi A. Müller, Dimitrios Serpanos, Forrest J. Shull Term Expiring 2018: Ann DeMarle, Fred Douglis, Vladimir Getov, Bruce M. McMillin, Cecilia Metra, Kunio Uchiyama, Stefano Zanero Term Expiring 2019: Saurabh Bagchi, Leila De Floriani, David S. Ebert, Jill I. Gostin, William Gropp, Sumi Helal, Avi Mendelson

Executive Director: Angela R. Burgess; Director, Governance & Associate Executive Director: Anne Marie Kelly; Director, Finance & Accounting: Sunny Hwang; Director, Information Technology & Services: Sumit Kacker; Director, Membership Development: Eric Berkowitz; Director, Products & Services: Evan M. Butterfield; Director, Sales & Marketing: Chris Jensen

COMPUTER SOCIETY OFFICES Washington, D.C.: 2001 L St., Ste. 700, Washington, D.C. 20036-4928 Phone: +1 202 371 0101 • Fax: +1 202 728 9614 • Email: [email protected] Los Alamitos: 10662 Los Vaqueros Circle, Los Alamitos, CA 90720 Phone: +1 714 821 8380 • Email: [email protected] MEMBERSHIP & PUBLICATION ORDERS Phone: +1 800 272 6657 • Fax: +1 714 821 4641 • Email: [email protected] Asia/Pacific: Watanabe Building, 1-4-2 Minami-Aoyama, Minato-ku, Tokyo 107-0062, Japan • Phone: +81 3 3408 3118 • Fax: +81 3 3408 3553 • Email: [email protected]

IEEE BOARD OF DIRECTORS President & CEO: Karen Bartleson; President-Elect: James Jefferies; Past President: Barry L. Shoop; Secretary: William Walsh; Treasurer: John W. Walz; Director & President, IEEE-USA: Karen Pedersen; Director & President, Standards Association: Forrest Don Wright; Director & VP, Educational Activities: S.K. Ramesh; Director & VP, Membership and Geographic Activities: Mary Ellen Randall; Director & VP, Publication Services and Products: Samir El-Ghazaly; Director & VP, Technical Activities: Marina Ruggieri; Director & Delegate Division V: Harold Javid; Director & Delegate Division VIII: Dejan S. Milojičić

revised 31 May 2017

20


COGNITIVE COGNITIVECOMPUTING COMPUTING

Enterprise Cognitive Computing Applications

Opportunities and Challenges Monideepa Tarafdar, Lancaster University, UK Cynthia M. Beath, University of Texas at Austin Jeanne W. Ross, Center for Information Systems Research, MIT

In this article, the authors provide an overview of cognitive computing applications for the enterprise. In particular, they classify opportunities for developing enterprise cognitive computing applications and describe implementation challenges.

C

ognitive computing applications utilize tools such as natural language processing (NLP), image recognition, intelligent search, and decision analysis to adapt their underlying computational and modeling algorithms or processing based on exposure to new data. Cognitive computing is commonly and colloquially known by many other descriptions, such as machine learning, artificial intelligence, and deep learning. We define enterprise cognitive computing (ECC) applications as those that introduce cognitive computing into software that enables an organization’s business processes. 1520-9202/17/$33.00 © 2017 IEEE

ECC aims to make business processes more efficient, accurate, relevant, and reliable. ECC applications are generating a great deal of excitement for organizations. However, largescale business impacts remain elusive. Many companies have invested in ECC applications, but most cannot point to significant benefits. An important reason for this is a lack of understanding about how ECC applications can contribute to the company’s business objectives. Our research suggests that many companies don’t fully appreciate the challenges associated with implementing ECC applications. The Center for Information


computer.org/ITPro

21

COGNITIVE COMPUTING AsiaPacific 7 (14%)

Ideation stage 8 (15%) In development 16 (33%) Discontinued 8 (15%) Europe 16 (31%)

In production 19 (37%)

(a)

North America 28 (55%)

(b)

Figure 1. Enterprise cognitive computing initiatives studied in this research grouped by (a) development stage and (b) region.

share our findings to provide an overview of cognitive computing applications for the enterprise. In particular, we classify opportunities for developing ECC applications and describe challenges in implementing those applications.

Creating superior employee experience 2 (6%)

Delighting customers 12 (33%)

Driving operational experience 21 (61%)

Figure 2. Opportunities for implementing enterprise cognitive computing applications. These applications typically aim to pursue one of three business objectives: driving operational excellence, delighting customers, or creating a superior experience for employees.

Systems Research (CISR) at the MIT Sloan School of Management conducted a study on ECC initiatives in 41 organizations, comprising 33 ECC application users and eight developers and vendors. In the 33 user organizations, which represented a broad range of industries distributed across North America, Europe, and the Asia-Pacific region, we studied a total of 51 initiatives and use cases for ECC applications, about 70 percent of which were either in production or use, or had a working proof of concept (see Figure 1). Here, we 22


Opportunities for ECC Application Development Two unique capabilities characterize ECC applications. The first is that they can process and make sense of increasingly large and growing volumes of data. ECC applications can handle data that—although available to humans—can be overwhelming because of the corpus’s sheer size. Examples include the entire medical literature on a given topic or the case law of a certain country or state. Moreover, ECC applications can adjust and adapt their models based on new data to return results that are increasingly consistent with past data. The second capability is their ability to automate tasks that formerly required human interpretation, such as queries within a call center that can now be interpreted by NLP or image recognition that can learn individuals’ identities. Thirty-five of the ECC applications we studied were in production or development. They sought to leverage these two capabilities to pursue one of three distinct business objectives: driving operational excellence, delighting customers, or creating a superior experience for employees (see Figure 2). We describe these objectives next and provide examples from among the applications in production.

Driving Operational Excellence Operational excellence is key to an organization’s ability to execute its core business processes in a reliable and repeatable way. Organizations that deliver superior operational excellence have stable core and routine processes.1 The largest number of use cases we found were ECC applications that enhanced an organization’s operational excellence. These included sophisticated search and retrieval from very large corpora of technical information, such as legal, accounting, and financial laws and regulations, as well as medical literature, predictive machine maintenance, product classification, fraud detection, and so on. In all these examples, ECC applications were used to enhance the speed, accuracy, and reliability of the organization’s various core processes, and therefore drive its operational excellence. In terms of speed, two sites (an auditing and consulting firm and a government regulatory body) used text mining, pattern matching, search, and NLP in ECC applications to parse large volumes of information and identify case law pertaining to their cases, helping to drastically speed up the grunt work of their attorneys. A global bank used Bayesian statistics and anomaly recognition for semiautomated fraud detection; the application narrowed down the list of probable cases for further consideration by the bank’s risk managers, thus speeding up the fraud-detection process. In terms of accuracy, an online retailer and a consumer analytics firm used pattern detection and data mining to quickly and reliably classify and categorize new products for correct presentation. A pharmaceutical company used genetic algorithm applications to develop models that could more accurately identify compounds that had the greatest likelihood of viability and success. In terms of reliability, we found several use cases involving predictive machine maintenance in datacenters. These applications used machine learning models that learned from past data to predict future breakdowns, helping increase the datacenters’ reliability.

Delighting Customers Those ECC applications aimed at delighting customers by either offering superior products and services or fostering customer loyalty and

engagement formed the second largest number of use cases. These included loan processing and delivery, call center support, claims processing, citizen tax services, delivery of insurance products, and customer feedback analysis. Several ECC applications allowed organizations to provide innovative, personalized, and superior products or services to customers. An online bank used Bayesian statistical modeling tools to understand customer behavior and demographics and offer loans in under seven minutes to retail customers. The bank’s leaders believed that providing loans to customers so easily and quickly gave them a significant edge over their competitors. A leading insurance company used text analytics and pattern matching to speed up its claims analysis process. Similarly, a financial services company provided insurance agents with machine learning tools that enhanced their ability to quickly answer questions about insurance policies and claims. Some ECC applications strengthen customer loyalty and trust. A pharmaceutical manufacturer used text mining and NLP to analyze customer feedback from multiple text and voice sources, which helped identify key product defects. Another bank harvested signals of customers’ intent from their interactions on the bank’s website. This enabled the bank to identify customers who were at risk of leaving and proactively offer appropriate products and solutions. This application has reduced customer churn. A government tax department used text analytics to mine laws for taxation implications so it could provide citizens accurate information about their tax liabilities.

Creating a Superior Employee Experience High-performing organizations empower their employees to seamlessly, quickly, and efficiently execute their key day-to-day tasks.2 ECC applications can help create a superior employee experience. For example, ECC applications in internal help desks help agents tackle a range of technical problems experienced by employees. An IT services organization used NLP, text mining, and pattern matching to give employees solutions to problems with their devices and applications. Machines generated answers to standard problems while handing off nonroutine and complex

computer.org/ITPro

23

COGNITIVE COMPUTING

Table 1. Cognitive computing tools for enterprise. Cognitive computing tool type

Open source tools

Vendor-provided tools

Point solutions that execute focused and relatively narrow tasks

Tensor Flow: helps systems transition from prototype to production

Luminoso: sentiment analysis

Spark: Antimalware

Naralogic: data clustering and classification

Apache OpenNLP: natural language processing

Ross Intelligence: legal discovery

Emdros: text mining

Kabbage: retail loan processing

WordNet::similarity: semantic pattern matching

IPSoft: call center management

Internally developed ECC platforms that leverage many open source tools

IBM Watson

Broad-based tools that offer a suite of solutions

problems to human experts, who were also supported by the ECC application. A life insurance company in Europe developed a similar application for its computer help desk. In another use case, employees had access to an application that used text mining and pattern matching to provide answers to questions about workrelated travel, such as passport and visa applications, currency, and airline tickets. In all these organizations, ECC applications helped create a digital environment that delivered a superior employee experience by making routine and often time-consuming processes more reliable and efficient.

Challenges in Implementing ECC Applications Although we sensed a lot of enthusiasm for the potential of ECC applications to deliver business benefits, there was a sobering recognition that achieving these benefits requires the organization to address multiple challenges. We observed that failure to acknowledge, understand, and tackle these challenges resulted in glacially paced implementations, failed implementations, and, in the worst cases, a strong disenchantment with and deep skepticism of ECC applications. Four challenges are particularly important. To take advantage of ECC’s ability to process massive amounts of data, business leaders must choose the right tools and make sure needed data is available to those tools. To take advantage of increased automation, business leaders must consistently supervise applications and appropriately allocate responsibilities between humans and machines. 24


Choosing the Right Tools Although numerous cognitive computing tools are available for ECC applications, not all tools are appropriate for all types of tasks. A tool that can predict the loan worthiness of potential customers might not be appropriate for answering call center queries because the nature of computations and data can be different for each. There are, in general, four types of tools (see Table 1), depending on whether they are point solutions that execute focused, relatively narrow tasks or broad-based applications that offer a suite of solutions; and on whether they are open source or provided by specific vendors. Point solution tools can tackle specific tasks such as NLP, text mining, and searching through unstructured data. These tools embed existing and relatively uniform ontologies—rules of vocabulary, syntax, and grammar—or a structured corpus of domain information, such as case law. In principle, organizations can develop this type of tool in-house, procure a vendor-supplied tool, or adapt an open source tool. Among the ECC applications we studied, where firms could benefit from existing ontologies, they adopted open source tools or licensed vendor-supplied tools that incorporated these ontologies. For example, an IT services company developed an ECC application for an internal technical help desk, applying open source tools that provided an ontology for NLP. In contrast, the vendor Luminoso developed an ontology for its sentiment analysis tool. Client companies build on this tool’s “base” ontology with data about their own products and customers to predict customer churn or anticipate demands for new product features.

Broad-based tools offer a set of general-purpose cognitive computing components that can be adapted to myriad specific ECC applications. For example, IBM’s Watson suite of applications has a variety of components such as NLP, pattern detection, and classification. An organization licensing this tool can develop company-specific ontologies by training the general-purpose tool with company-specific data. Broad-based cognitive tools need massive amounts of up-front data and training sets to develop algorithms. A leading bank in Europe used a generalpurpose cognitive computing tool to develop an ECC application that captured a raft of crucial information about the bank’s highest-value customers. Much of the effort in the development of this application involved gathering this data from myriad sources and then teaching the tool the relationships among different data entities. The bank succeeded in adapting the general-purpose tool to its specific ECC application. It was an exception. Because of the significant resources and costs involved in doing this, multiple companies abandoned applications of broad-based tools. For example, a heavy-machine manufacturing company tried to implement a troubleshooting ECC application using a broad-based tool. It did not achieve the desired level of accuracy, even after significant up-front effort. Eventually, the company abandoned the initiative while noting that a much simpler search engine might have been a more appropriate choice. Open source and vendor-based tools require different types of resources and skills. Open source tools call for technical expertise to program and contextualize them to the specific requirements of a particular ECC application. We observed that organizations that successfully drew on open source tools to build their ECC applications had a strong base of technical expertise (for example, a healthcare claims audit firm and two pharmaceutical companies). On the other hand, purchasers of focused tools can, in principle, benefit from the continuous improvements that the vendor makes to the base ontology. One tool vendor in our study mentioned that the high quality of its sentiment analysis tool was the direct result of continual monitoring and analysis by a product team of linguists, data scientists, and neural-network scientists that worked to improve the tool’s underlying algorithm. It thus

seems more efficient to license vendor tools and add additional company-specific ontologies on top of them to reduce requirements for validation and input from experts within the client organization. Finally, technically capable organizations can extend and enhance combinations of open source tools, such as NLP, speech recognition, and intelligent search, to develop their own broadbased tools. An organization in the IT services industry developed a broad-based ECC platform largely from different open source components that enabled its employees to provide help desk support and other customer services.

Getting the Right Data, and Getting the Data Right ECC applications critically and completely depend on massive and continually growing amounts of data to return outputs that are consistent with the data. There are three challenges concerning data. The first is identifying the right data. A particular ECC application might require external data, such as legal documents, libraries, journals, consumer reports, or real-time information feeds; or the organization’s internal data, such as product manuals, troubleshooting databases, operational databases, and subject matter expert knowledge. For example, an audit firm procured large external databases of accounting law manuals and financial statements for its accounting law analysis ECC application, while a heavy-equipment manufacturer required product manuals and troubleshooting databases to be accessed by its field service engineers in its product troubleshooting ECC application. The second challenge is ensuring that needed data is available. For instance, search and query ECC applications require significant amounts of test data to train the model. The third challenge is getting the data into the right format. ECC applications require data that is appropriately cleaned, formatted, and structured.3 Format problems might exist even with an organization’s internal data. A bank found much of its data on high-value customers in the locally stored files or memories of relationship managers. All of this had to be digitized, tagged, and structured before it could be accessible to its ECC application. Another bank found that PDF documents could not be used as-is because its ECC application could not read that format.

computer.org/ITPro

25

COGNITIVE COMPUTING

External data, which is not under the control of the organization, can present even more formatting problems. One way to acquire external data, especially data available on webpages, is to look for vendors that offer applications that can “scrape” such data. A bank deployed an ECC loan-processing application from a vendor that used another vendor’s program to scrape financial statement data from the websites of applicants’ banks. This was done in real time and saved the bank the cost and effort of acquiring the data. Another way to acquire external data is to deploy ECC applications that can interface with high-traffic websites such as Twitter and TripAdvisor to analyze data using techniques such as text mining. A number of vendors of sentiment analysis ECC tools, such as Luminoso, provide such functionality. One way to address data-related challenges is to link ECC initiatives to existing data analytics resources and capabilities. A government labor/employment agency developed a proof of concept for an ECC application that provided citizens guidance on jobs available and skills required. It leveraged curated data from a successful predictive-analytics application for job skill matching. In another example, a retail bank’s ECC application for approving loan applications was implemented by the team responsible for the organization’s data analytics applications.

Supervising on an Ongoing Basis As ECC applications handle more and more of a company’s routine tasks, it is tempting to assume they need little attention. However, a lack of supervision can cause the applications to drift as the organization’s customers and products change over time, and new policies and business rules are needed. For example, a call center ECC application would need to take into account new products, customers, services, complaints, and solutions in an ongoing way. It thus needs to be supervised by technical and subject matter experts so that the answers it returns are validated for relevance and accuracy, even in the face of these changes. Without supervision, the application’s models will lose their relevance over time, and the application will be able to handle a smaller and smaller percentage of its assigned tasks meaningfully. 26


Organizations thus face the challenge of on going supervision of the ECC application—that is, of monitoring the performance of the application and regularly recalibrating it against desired or correct outputs. An online retailer mentioned that it revised its ECC recommendation engine model on a regular basis, with some product categories requiring more frequent revisions than others. An insurance company assigned product managers as subject matter experts to oversee the outputs of its call center ECC application. Supervision is necessary for maintaining and managing the quality of the knowledge base residing in the ECC application. As noted earlier, vendors provide some supervision of their tools. A vendor of a case law analysis tool mentioned that it collected both system-generated and customer-generated feedback, which served as a basis for further improvements to its product’s algorithms. The vendor’s system recorded the number of times a client-user actually clicked on the output links of a particular search query, and the vendor also collected qualitative feedback from its clients about the tool’s performance. However, even with vendor-provided tools, it is important to monitor the output of the ECC application to ensure fit with changing organizational needs. Supervision is thus necessary to ensure that the tool retains its accuracy and relevance over time.

Clarifying the Division of Labor Clarity regarding the allocation of responsibilities between human users and the ECC application is essential to successful ECC application deployment. In general, this means that ECC applications will take on more routine tasks and leave complex or discretionary tasks to humans. In addition to retaining some prior responsibilities, ECC applications create new human tasks, such as training and sustaining the application. If executives have a good grasp on how cognitive computing works, they are more likely to make the necessary investments for data preparation, integration with other enterprise applications, and development of ECC governance structures. Understanding which aspects of a business process are routine enough to be handed over to the machine and which need to be retained by human users is often, as we found, a matter of iterative use and learning. Two banks we studied had ECC applications for

detecting fraud. They experimented with the application to understand which of the tasks the application could perform, which ones needed validation by human users, and how exceptions should be handled. Lack of clarity in articulating the respective roles of the human user and the ECC application can result in expecting too much of the application or, conversely, using human talent inefficiently. Companies will gradually learn which tasks are best assigned to people and which are best assigned to ECC applications. Initially, ECC app lications aimed at helping lawyers identify appropriate case law were limited to keyword searches that simply helped them get their hands on potentially useful reference cases sooner. As cognitive computing has been applied to such applications, the application is better able to predict which cases are most useful and how they can be applied. The shift of tasks from legal talent to machines has been gradual. Machines can now do much of the work that law clerks once did. On the other hand, lawyers still provide the reasoning and creativity that convert the case material to the needs of a specific engagement.

T

he deployment of intelligent algorithms in business processes calls for caution because, as Ray Kurzweil observed, such algorithms, if not properly understood, can lead to “runaway” and “rogue” information processing logic that is largely disconnected from underlying business realities.4 To add to Kurzweil’s concern, our research found that unrealistic expectations of mystical powers are also a potential problem. If executives regard cognitive computing as nearly magical, projects are more likely to fail. Executives leading ECC initiatives must become knowledgeable about the abilities and inabilities of cognitive computing tools. Enthusiasm is not enough—a clear-eyed view of the possible is crucial. Much has been written about high-profile applications of cognitive computing—for example, those that can win Go, Jeopardy!, or chess. These are complex but focused applications with a fixed set of objectives and logic. In contrast, business processes can have competing objectives and dynamic business rules. We are still very early in our ability to apply cognitive computing applications in the business enterprise.

Given the relative newness of ECC applications, the associated hype around them, and a lack of systematic description regarding what is possible from them, we hope this article will help researchers, practitioners, and those with a general interest in how technology benefits organizations to see past the hype.

References 1. J.W. Ross, I.M. Sebastian, and C.M. Beath, “How to Develop a Great Digital Strategy,” MIT Sloan Management Rev., Winter 2017, pp. 7–9. 2. K. Dery, I.M. Sebastian, and N. van del Meulen, “Building Digital Value from the Digital Workplace,” MIT Center for Information Systems Research briefing, vol. XVI, no. 9, 2016. 3. S. Earley, “There Is No AI without IA,” IT Professional, vol. 18, no. 3, 2016, pp. 2–8. 4. R. Kurzweil, The Singularity Is Near: When Humans Transcend Biology, Penguin Books, 2006.

Monideepa Tarafdar is a professor of information systems in the Management School at Lancaster University, UK, a research affiliate at MIT’s Center for Information Systems Research, and a visiting professor at the Indian Institute of Management Calcutta, India. Her current research focuses on human-machine partnership and its positive and negative consequences. Tarafdar received a PhD in management from the Indian Institute of Management Calcutta. She is a member of the Association for Information Systems. Contact her at [email protected] or [email protected]. Cynthia M. Beath is a professor emerita at the McCombs School, University of Texas at Austin. Her current research is on the uptake of artificial intelligence and organization design for the digital economy. Beath received a PhD in management from the University of California, Los Angeles. She is an Association for Information Systems Fellow. Contact her at [email protected]. Jeanne W. Ross is a principal research scientist at MIT’s Center for Information Systems Research. Her current research examines digital transformations and how people and machines use data to achieve strategic objectives. Ross received a PhD in management information systems from the University of Wisconsin–Milwaukee. Contact her at [email protected]. Read your subscriptions through the myCS publications portal at


computer.org/ITPro

27

COGNITIVE COMPUTING

Cognitive Compliance for Financial Regulations Arvind Agarwal, Balaji Ganesan, Ankush Gupta, Nitisha Jain, Hima P. Karanam, Arun Kumar, Nishtha Madaan, Vitobha Munigala, and Srikanth G. Tamilselvam, IBM Research–India

Compliance with regulations is getting increasingly hard because of complex documents and the sheer volume of regulatory change. The authors’ Cogpliance platform uses a cognitive approach to help users achieve regulatory compliance.

F

inancial institutions are faced with rapidly changing regulatory policies and an ever-growing number of regulations. It is estimated that by 2020, global banks will be required to comply with more than 120,000 pages of regulations.1 Failure in adhering to these regulations often leads to huge monetary fines, customer dissatisfaction, and damage to a business’s reputation; thus, CFOs see this as their topmost challenge.2 Compliance is further complicated by the complex language used in regulatory documents,3 which forces banks to hire domain experts whose primary job is to identify relevant regulations

28


and, accordingly, introduce or suggest changes to specific internal controls. This demand for skilled labor has led to growing operational costs in recent years, sometimes accounting for more than 10 percent of total operational expenses.4 To address the volume, velocity, variety, and complexity of regulations, banks are increasingly seeking technological help. To remain compliant, banks must deal with two kinds of compliance tasks. The first is dealing with internal regulatory data—that is, internal controls. Banks need to constantly monitor compliance functions at the operational level to make sure that the desired controls are being


1520-9202/17/$33.00 © 2017 IEEE

Related Work in Regulatory Compliance

G

iven the importance of compliance in today’s world, several efforts have tried to make the compliance process as easy and inexpensive as possible for organizations. Thomson Reuters Regulatory Intelligence provides a common place for compliance officers to browse through regulatory documents, search them based on keywords, and go through the updates released by regulatory authorities (tmsnrt. rs/2kkIGad). However, with this tool, documents are ingested manually, and any update released by regulatory authorities is obtained from its source and entered into the tool at the appropriate place. This limitation is overcome in the platform we propose in the main text, in which we automate data gathering, processing, and enrichment steps to reduce dependency on skilled labor. Similar tools are provided by other leading consulting firms, such as KPMG’s Regulatory Compliance Tool (bit.ly/2pKjsT2), the Deloitte Compliance Platform (bit.ly/2qUAxNP), and the Adastra Integrated Compliance Platform (www.adastracorp.com/icp). However, these are all built on years of knowledge acquisition and business-process-specific data, the details of which are propriety and often not publicly available. In addition to generic compliance products, there are also task-specific products. NICE Actimize provides a Financial Markets Compliance tool, as well as tools that help with antimoney-laundering by delving into the details of customer relationships (www.niceactimize.com). QUMAS provides a compliance platform focused on forms and document storing, and process management functionalities (www.qumas.com/ qumas-compliance-platform); PwC provides tools for risk management, audit, tax compliance, and so on (pwc.to/2qkn5As). IBM Regulatory Compliance Analytics identifies potential obligations in regulations and provides a dashboard for managing controls (ibm.co/2qkzzru).

executed as intended and that a compliance status is reported to management. This helps in determining the bank’s current risk exposure. The second task is dealing with external regulatory data. Banks need to keep track of new regulations

IBM Open Pages, on the other hand, focuses on corporate governance, risk, and compliance (ibm.co/2qHltUp). However, none of these products are targeted toward understanding complex regulatory documents or perform tasks such as question answering, change tracking, or obligation control mapping. Machine learning, or artificial intelligence in general, has just started to gain ground in regulatory compliance.1 Despite the importance of the problem, there has been limited work in compliance change tracking—in fact, to the best of our knowledge, there is no academic work particular to compliance change tracking. There is some academic work around building ontologies for regulatory change management,2 a critical component of the enterprise, governance, risk, and compliance framework. Some commercial products available in the space of regulatory change management are 360Factors (bit. ly/2pMWA4A), MetricStream (bit.ly/2pMM8KA), Thomson Reuters (tmsnrt.rs/2pMaIf7), and KPMG (bit.ly/2pLJmGh). These systems are often built on top of domain knowledge collected over years and with the help of a large workforce that constantly tracks regulatory changes and makes them available to subscribers. Furthermore, to the best of our knowledge, there is no system or commercial product targeted toward understanding regulatory documents and performing question-answering tasks, as we propose here. References 1. M. MacDonagh and W. Kluwer, “How AI Has the Potential to Transform Regulatory Compliance,” Wolters Kluwer, 6 Sept. 2016; bit.ly/2q77aFt. 2. A. Espinoza, E. Abi-Lahoud, and T. Butler, “OntologyDriven Financial Regulatory Change Management: An Iterative Development Process,” Proc. 2nd Semantic Web and Linked Open Data Workshop (SW-LOD), 2014; enc2014.cicese.mx/Memorias/ paper_95.pdf.

or changes in regulations to ensure that they remain compliant with the latest policies. Here, we focus on the second task—dealing with understanding external regulatory documents before they can be translated into internal controls.

computer.org/ITPro

29

COGNITIVE COMPUTING

Understanding regulatory documents has several applications. For instance, a new bank, or an existing bank entering a new country or business line, would like to know what regulations apply to it. Furthermore, bank employees encounter different customer situations and need to know whether a particular action is allowed by regulations. Situations such as these require a mechanism for querying and getting answers from regulatory documents. Although a bank might have understood a regulation and implemented controls to comply with it, revisions to the regulation need constant attention. Amendments, which can themselves run into tens of pages, need to be understood and compared with the original regulations to determine what has changed. This has a direct impact on the bank’s internal compliance functions, which might need to be adjusted accordingly, or new ones will need to be put in place. Hidden among these two tasks is another task for banks: correlating regulatory documents with internal controls so that coverage and gaps can be identified. Again, given that internal control policy and regulation documents can each run to several tens of pages and use very different language, it is a nontrivial task to determine association and coverage. Here, we propose a system that addresses some of these challenges by helping compliance officers to understand regulations and get their tasks done faster and with much less effort. Our cognitive compliance platform, Cogpliance, uses machine learning, information retrieval, and natural language processing (NLP) techniques coupled with a novel user experience design to provide an end-to-end system. Although we present the platform in the context of the financial domain, the techniques used are equally applicable to other domains, such as healthcare, payroll, and manufacturing.

system requires deeper understanding of regulatory documents. We must understand different entities in the document and their relationships and be able to put them together for reasoning. For instance, consider the following example containing a question, a passage containing the answer, the answer itself, and the explanation of the answer:

Our Approach

Our approach to the QA system is based on sophisticated machine learning and NLP methods. Documents are ingested into the platform, parsed, and converted into meaningful, continuous pieces of text—that is, clauses. From each clause, we extract concepts and relationships and add to a knowledge graph. To aid the extraction process, we use an ontology. This ontology can be built automatically or be provided

Our approach to regulatory compliance tasks is motivated and driven by recent advancements in both software and hardware—more specifically, advancements in artificial intelligence and in distributed computing using GPUs. Bank employees often need a questionanswering (QA) system that can make answers available at their fingertips. Building such a 30


Question (from bank): Can I open a new branch in Agra? Passage containing the answer: RRBs should allocate at least 25 percent of the total number of branches proposed to be opened during a year in unbanked rural (Tier-5 and Tier-6) centers. Answer: No. Explanation: Agra is a Tier-1 city.

In this example, we must not only understand the passage containing the answer, but also have additional information for reasoning—that is, that Agra is a Tier-1 city. The proposed system uses additional information along with the entity relationship information in the passage to reason and answer. The passage shown in this example is for illustrative purposes only and does not reflect the complexity of the language used in regulatory documents. A typical clause from a regulatory document is as follows: RRBs are permitted to open branches in Tier-2 to Tier-6 centers (with populations of up to 99,999 as per Census 2001—details of classification of centers tier-wise furnished in Annex IV) without having the need to take prior permission from Reserve Bank of India in each case, subject to reporting, provided they fulfill the conditions laid down in para 1.3. RRBs which do not satisfy the said conditions may approach the Regional Office of RBI for prior approval. Their applications will be considered on the basis of conditions laid down in para 1.2.

Enrichment Blekko User document upload Custom crawler

Kafka Document conversion HDFS

Ontology creation Document ingestion

Preprocessing

Rule-based information extraction

Solr index

Search Alerts

Machine-learningbased annotations Custom annotators

Regulatory repository

Service APIs

Knowledge graph

Querying

Q&A system

Reasoning

Actionables

Ontology service

Enrichment

Knowledge graph construction

Services

Applications

Figure 1. The Cogpliance cognitive compliance platform. The system architecture consists of multiple phases, from data ingestion to the application layer.

by domain experts based on the questions that the system is expected to handle. The knowledge graph further contains the information about the external world obtained from sources such as Wikipedia, Yago, and Freebase. Although there are sophisticated methods available for building an ontology automatically, they still require experts’ intervention, one limitation of the current technology. Another application is regulatory change tracking, in which organizations need to know the difference between regulations and their amendments. Our approach to this task is to use machine learning modules such as semantic comparison, document alignment, and semantic difference finders, which can find differences between regulatory documents and their amendments. These modules’ outputs are also stored in the knowledge graph, so they are available to organizations whenever needed. These machine learning modules can also be used for compliance mapping—that is, generating a map between regulations and organizations’ policy and controls. Organizations often need this map to identify potential gaps in their controls and analyze the impact of any changes to those controls. Our approach is to use semantic comparison and document alignment to generate this mapping.

Cognitive Compliance Platform Our Cogpliance platform can support various services and applications. Figure 1 shows a diagram of the platform architecture. The architecture consists of multiple phases—that is, data ingestion, preprocessing, data enrichment, data store, supported services, and applications.

Data Ingestion and Preprocessing In the data ingestion and preprocessing phases, regulatory documents are obtained from multiple sources. We use the term “regulatory document” to denote any document that might contain information or updates about regulations. These documents cover a broad range; they might refer to documents published by regulatory authorities such as the US Securities and Exchange Commission or Federal Reserve System, news articles that contain updates about regulations, or software patches released by commercial products such as SAP or PeopleSoft. These regulatory documents can be crawled from websites, retrieved from repositories, or uploaded by users. Due to the varied nature of sources, these regulatory documents come in many formats, including PDF, HTML, and text files, and thus undergo several preprocessing steps, such as document format conversion, data cleaning tasks (for instance, content extraction from tables),

computer.org/ITPro

31

COGNITIVE COMPUTING

annexures, external reports, and removal of unwanted tags. After the data is preprocessed, it is stored in a distributed file system—namely, the Hadoop Distributed File System (HDFS).

relational database (SQL store), a graph database (Titan), or a search index, such as Apache Solr. Typically, applications store data in more than one type of store.

Data Enrichment

Applications and Services

In the data enrichment phase, documents are read from HDFS and enriched with tags. These tags include concepts, relationships, and attributes of concepts. For example, a Bank and a Regulator could be identified as concepts, and a Regulator can have one-to-many relationships with the Bank. An attribute of the Bank concept could be bank type—that is, an Investment Bank or a Commercial Deposit Bank. A formal schema of these identified concepts, relationships, and attributes is captured as an ontology. We call the process of adding additional information (tags) to the documents annotation and the programs that enable this annotators. The platform is enabled with various annotators to help with document enrichment:

The final phase in this architecture is the application layer. Several cognitive applications are supported on top of the proposed architecture. These applications communicate with the underlying platform using REST APIs.

• regular-expression-based annotators, • machine learning annotators, and • rule-based annotators. Regular-expression-based annotators are used to extract document structure and entities such as date, time, numbers (transaction amounts, monetary figures, deadlines), and so on. Machine learning annotators help in concept detection, document similarity, and so on. Domain-specific rules help in these tasks whenever domain information must be incorporated. The final step of the data enrichment phase is knowledge graph construction. This typically involves storing concepts as vertices and relationships as edges. The attributes of concepts are stored in the vertices, while the relationship type is stored as an attribute of the edges. The ontology serves as the schema for the knowledge graph. This knowledge graph can then be queried for search, comparison, and other applications.

Data Store The third phase of the architecture is the data store. By data, we mean the original ingested documents and annotations and the knowledge graph created from them. Based on the application’s requirements, data can be stored in a 32


Supported Services We next describe some of the applications and services supported by the final layer of the Cogpliance architecture.

Regulatory Change Tracking As mentioned, regulatory change tracking helps organizations keep track of changes in regulations and answer questions such as, “What are the changes in the newly published regulation that we care about?” A conventional approach to solving this problem is to deploy a large workforce that constantly keeps track of sources in which regulatory bodies publish changes. On a regular basis, employees browse through these sources to first identify relevant regulations and then compare them with the previous version to identify the changes. This manual process is not only time-consuming but also prone to errors. In this section, we describe an approach built on the architecture presented in Figure 1 that can be used to automate this process. Our regulatory change tracking application monitors all documents that can possibly contain regulatory changes. These documents can be sourced from various places, such as government websites, news websites, or software patch documents released by commercial product vendors. In the application, these documents first go through the ingestion and preprocessing phases. Then, the application identifies the relevance of the documents to the organization’s area of operation. A machine-learning-based annotator—that is, a classifier—is used for relevance identification. These relevant documents are then further processed to identify the change in the regulation; for this, we extract the exact piece of text that mentions this change. Again, a machine-learning- and NLP-based module that

can understand the semantics of text is used for this extraction. These changes are extracted and made available to users through the knowledge graph. They are semantically compared with the software patch descriptions to raise appropriate alerts notifying the compliance officer of whether the software patches cover the regulatory changes. In cases where they do not, organizations need to implement the patches in their systems on their own.

Regulatory Search and QA Searching regulatory documents is enabled by a regulatory repository built on top of the data store, as Figure 1 shows. In addition to the keyword search, the regulatory repository enables faceted search by indexing the additional information extracted during the enrichment phase. This faceted search helps users filter results based on predefined categories, such as industry, regulator, topic, and date. For the QA system, the regulatory documents are converted into text from various formats, and the structure of the documents (sections, subsections, lists, paragraphs, tables, and so on) is extracted using annotators. An obligations extractor is used to identify obligations in the documents. This obligations extractor can be a rule-based system (manually defined linguistic patterns), a statistical machine learning classifier (supervised and requiring sufficient training data), or a hybrid (rule-based plus statistical) model. Given that regulatory documents undergo considerable revision and redrafting, compliance officers might want to refer to previous document drafts. This is enabled by versioning the documents. Different versions of the documents, extracted obligations, and other annotated data are made available through the knowledge graph. As discussed, a QA system can help compliance officers respond to questions asked by different departments in their organization. Instead of looking up the regulators’ website or documents, the knowledge graph can be queried directly in natural language. This reduces both the manual effort involved and human error. This system can also help relatively inexperienced professionals to leverage the institutional knowledge gained over time by more experienced officers. This institutional knowledge is captured by the QA system through a feedback mechanism.

The user interface for such a QA system is as important as the content itself (see our prior work5). In the proposed system, our user interface supports search, natural language querying, and the hierarchical representation of the document, showing the clause containing the answer, its section, and possibly a related annex for further document exploration.

Case Studies Let’s now examine the case studies we conducted for both the regulatory change tracking and regulatory search and QA services.

Regulatory Change Tracking Case Study One organization we worked with employs four regional compliance managers and around 227 payroll practitioners concentrated in North America, Latin America, Europe, and Asia for regulation change tracking. On average, these officers spend roughly five hours per day in compliance activities. The bulk of this time is spent on reading through regulatory updates obtained through paid subscription services and determining if any action needs to be taken. This effort exponentially increases without access to regulatory update services. Given what such services cost per user license, organizations tend to limit their access to a select few senior officers. A potential miss could lead to penalties in thousands to millions of dollars. With Cogpliance deployed, the system has received positive feedback from the organization and is estimated to bring down practitioners’ time spent on articles to very minimal by providing relevant and actionable information filtered by geographical or other factors.

Regulatory Search and Q&A Case Study In another case, a bank had a dedicated compliance team to understand existing regulations, formulate its policies around those, respond to queries from ground staff from various business divisions, and help those staff with compliance-related queries in day-to-day business operations. The limited availability of this small compliance team has created a dependency in other units, and queries are often delayed, leading to customer dissatisfaction. Furthermore, the operational staff simply delegates all responsibility to the compliance team.

computer.org/ITPro

33

COGNITIVE COMPUTING

With Cogpliance deployed, the operations team not only has direct access to regulatory content to get their queries answered, but the bank’s interpretation of it is available to them 24-7. In the second scenario, an example from the customer query “What does the regulation say about check book issuance to visually impaired people?” will be answered by a combination of search results, the document exploration view, and knowledge graph answers. Search results point the user directly to the relevant clauses. The user can better understand the context by exploring the section and other clauses using the document exploration view. The knowledge graph answers compliment the search results by highlighting the potentially important aspects of the clauses. This direct access to answers considerably reduces the turnaround time for the operations team, while also relieving the compliance team from repeated requests.

P

ractitioners can replicate the solutions from our case studies by developing necessary components such as document ingestion, data enrichment, and data store, the specific details of which would depend on the application and data being processed. Users or practitioners can use solutions built on this architecture to help them with their compliance tasks. As shown in the case studies, our regulatory change tracking solution can be used by practitioners to receive alerts about regulatory changes and actionable items based on these changes. The QA solution can be used to obtain real-time and up-to-date information about regulations.

The primary input to the architecture is unstructured regulatory text, which opens up multiple challenges for researchers. These include processing complex PDF documents to extract tables, figures, multiple columns, and headers and footers correctly; improving NLP techniques for handling complex sentences in this domain; and noise filtering for data coming in from unofficial sources such as the news. A future version of this system would entail not only improved individual components that can better handle data ingestion, processing, and enrichment, but also more sophisticated methods to handle regulatory text. Better and deeper semantic understanding 34


of regulatory text would enable us to take this system to a level at which it can start to perform advanced cognitive tasks, such as reasoning and inference.

References 1. A. Kumar, Regulatory Compliance Management in Banks: Challenges and Complexities, Infosys, 2016; infy.com/ 2q4s0r9. 2. N. Amato, “Compliance Costs, Complexity Rising in Finance,” J. Accountancy, June 2016; bit.ly/28OkszA. 3. T.D. Breaux and A.I. Antón, “Mining Rule Semantics to Understand Legislative Compliance,” Proc. 2005 ACM Workshop Privacy in the Electronic Society, 2005, pp. 51–54. 4. The Cost of Compliance: 2013 KPMG/AIMA/MFA Global Hedge Fund Survey, Oct. 2013; bit.ly/2qtsmbv. 5. N. Madaan et al., “Visual Exploration of Unstructured Regulatory Documents,” Proc. 22nd Int’l Conf. Intelligent User Interfaces Companion, 2017, pp. 129–132.

Arvind Agarwal is a researcher at IBM Research–India. His research interests are in machine learning, in particular, problems related to social computing, recommendation systems, data mining and visualization, with an emphasis on large-scale algorithms, and matrix and graph algorithms. Agarwal received a PhD in computer science from the University of Maryland, College Park. He received the best student paper award at the 2010 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases. Contact him at [email protected]. Balaji Ganesan is a research software engineer at IBM Research–India. His research interests include web search, computational advertising, and knowledge graphs. Ganesan received an MS in computer science from the University of Arizona. Contact him at [email protected]. Ankush Gupta is a software engineer at IBM Research– India. His research interests include natural language processing, information retrieval, and extraction and data mining. Gupta received an MS by Research in computer science and engineering from the International Institute of Information Technology, India. He received the 2013 IBM Research Division Award for his work on developing a live system to deliver real-time, personalized, and locationbased offers and recommendations to customers on various delivery channels such as SMS, email, and NetBanking. Contact him at [email protected].

IEEE INTERNET COMPUTING

march • april 2016

nov ember • december 2015


jaNUaRy • fEbRUaRy 2016 INTERNET ECONOMICs

ClOUd sTORaGE

MEasURING ThE INTERNET

ExPlORING TOMORROw’s INTERNET

ThE INTERNET Of YOU

VOl. 20, NO. 1

www.COMPUTER.ORG/INTERNET/



VOl. 20, NO. 3

VOl. 20, NO. 4



VOl. 20, NO. 2

vOl. 19, NO. 6

October 9, 2015 3:26 PM


May • jUNE 2016

jUly • aUGUsT 2016

MaRCh • aPRIl 2016

NOvEMbER • DECEMbER 2015

Cover-1

Srikanth G. Tamilselvam works in the Knowledge Engineering Department at IBM Research–India. His research interests include software engineering, interactive user interfaces, and machine learning. Tamilselvam received an MS in computer science and technology from Mysore University, India. He is recognized as an IBM Master Inventor for his contribution to the IBM-focused patent portfolio. Contact him at [email protected].





IC-19-06-c1

july • augus t 2016

Arun Kumar is a senior researcher and manages the Cognitive Solutions research group at IBM Research–India. His current work is focused on applying cognitive computing technology in the finance and IT support domains. Kumar has filed several patents and has coauthored numerous refereed papers in reputed journals and conferences. He received a PhD from the Indian Institute of Technology, Madras. Contact him at [email protected].

Vitobha Munigala is a research software engineer at IBM Research–India. His research interests include natural language processing and smart energy analytics. Munigala received an MS in computer science from the Indian Institute of Technology, Bombay. Contact him at [email protected].

IC-20-01-c1

IC-20-03-c1

IC-20-02-c1

Cover-1

january • february 2016

Hima P. Karanam is a senior technical staff member at IBM Research–India. His research interests include largescale data curation, information integration, and data quality. Karanam received an MS in computer science from the Indian Institute of Technology, Madras. Contact him at [email protected].

Nishtha Madaan is a research software engineer at IBM Research–India. Her research interests include business process mining and natural language processing. Madaan received an MS by Research in computer science from the International Institute of Information Technology, India. Contact her at [email protected].

may • june 2016

Nitisha Jain is a member of the Cognitive Solutions team in the Knowledge Engineering and Data Platforms department at IBM Research Lab–India. Her research interests include distributed systems, data analytics, information retrieval, and knowledge discovery. Jain received an MSc in computer science from the Indian Institute of Science, Bangalore. She has authored several research papers for international conferences as well as a journal paper for IEEE Transactions on Services Computing. Contact her at [email protected].

Cover-1

Cover-1

December 7, 2015 1:45 PM

April 13, 2016 8:45 PM

February 11, 2016 10:30 PM

Want to know more about the Internet?

This magazine covers all aspects of Internet computing, from programming and standards to security and networking.

www.computer.org/internet

computer.org/ITPro

35

COGNITIVE COMPUTING

Building a Cognitive Application Using Watson DeepQA Christopher Asakiewicz, Edward A. Stohr, Shrey Mahajan, and Lalitkumar Pandey, Stevens Institute of Technology

Academic advisors assist students in academic, professional, and personal matters. The authors’ cognitive advising system uses IBM Watson’s cognitive intelligence to identify question categories and then answer questions accordingly.

A

ccording to IBM, we are entering a third great era of computing—the cognitive computing era—characterized by computers’ deep intelligence and ability to both understand problems that were hitherto tractable only for humans and, importantly, learn from experience. Cognitive computing offers powerful technologies with the potential to augment human capacity and understanding and thus has significant commercial and societal importance.1 The first two computing eras, the tabulating era (1900 to 1940) and the programming era (1940 to the present) are distinguished by the

36


development of our capacity to instruct computers to perform complex operations at high speed and with unparalleled accuracy and reliability. In the second era, the exponential growth in computational power and storage capacity changed big data from a challenge to an unprecedented source of opportunity and innovation. Enabled by the explosion of computing power, advances in machine learning, and the newfound ability to handle massive structured and unstructured datasets, cognitive computing is at last helping to realize the potential of some 50 years of artificial intelligence research. The parallel developments associated with big data and cognitive computing


1520-9202/17/$33.00 © 2017 IEEE

provide unprecedented opportunities and challenges for business and society. With regard to challenges, we need only consider the human displacement potential of robotics and deep intelligence applications.2 Here, we address the gap between the technical and broad overview literature on cognitive computing by developing guidelines that can be used by nontechnical analysts and developers to analyze and develop a common class of cognitive applications. Our focus is on the DeepQA question-and-answer (Q&A) systems that first gained notoriety when IBM’s Watson defeated the best human Jeopardy! players.3 We developed our guidelines based on our experience developing a frequently asked questions (FAQs) application, ASK@, for the MS in Business Intelligence and Analytics (BI&A) program at the Stevens Institute of Technology. The system answers questions from several types of clients: potential applicants to the program, newly admitted students, current students, faculty members, and business professionals. When fully developed, our Intelligent Academic Advisor (IAA) will relieve faculty advisors and university administrators from the burden of answering hundreds of questions daily, while helping us develop an accessible, natural language interface that will provide accurate answers about many process, curricula, and policy questions. Moreover, this “ground truth” will not be subject to the necessarily limited and fragmented knowledge of university personnel. As it learns, IAA will become more authoritative and will grow in terms of coverage and ease of use over time. Moreover, ASK@ can be adapted to applications other than academic advising simply by changing the corpus. Our application is not unlike other DeepQA systems. In particular, a Virtual Teaching Assistant for an online class developed at the Georgia Institute of Technology used Watson technologies and could answer questions with 97 percent accuracy, “with the remainder being answered by human assistants.”4 The advantage of the IAA application for our research is its simplicity. It addresses a relatively simple domain, and, consequently, it is relatively easy to develop the corpus of knowledge, train the system, and gain accurate assessments of its performance in terms of accuracy, ease of use, and user satisfaction.

DeepQA systems differ from the transactional and web-based inquiry systems of the programming era because they can learn and become more intelligent over time. A DeepQA system deals with unstructured data in the form of a large knowledge corpus that could be derived from specially developed text, publications, websites, or blogs; supports natural language inquiries; and can usually converse with users through some form of intelligent dialog.

Design Approach Traditional systems developed and deployed as “expert systems” have relied on either preconfigured answers and rules-based recommendations or on content and document management based on search features. Cognitive systems developed or deployed as “assistants” leverage automated interactive connections to knowledge assets while expanding the conversation by introducing new content and insights. These systems usually learn from their interactions with users and adapt to language and situational exigencies.5 Figure 1 provides an overview of the Q&A process flow associated with any Q&A system. The development of the ASK@ intelligent advisor application we describe illustrates an instance of this process flow. Initially, the system was more than acceptable for answering most FAQs; however, it didn’t perform as well with the infrequently asked questions. As more answers were added to the core categories, response accuracy improved significantly. As depicted in Figure 1, there were two very different corpora of information. The first was our database of questions and answers; this initial set of 293 questions and answers was developed manually. The second was our corpus of content coming from FAQs, brochures, syllabi, and the web. The second corpus was important in enhancing system confidence around the types of answers that might be suitable for answering a specific question. Our proposed, four-step system development approach is keyed to the Q&A process flow illustrated in Figure 1.

Business Objective Statement

Developing a cognitive business application6 begins with framing the business problem or question to be addressed and outlining the scope of

computer.org/ITPro

37

COGNITIVE COMPUTING

Text to speech

Speech to text

Questions

Answers

Update Q&A database

Q&A database

Yes

Is answer correct?

…

+

Dialog

Web

Natural language classifier

Retrieve and rank

Syllabi

Document conversion

Brochure

Question recognition and response: text

Corpus

Watson services FAQs

voice

Q&A process flow

Response and confidence

Update Q&A database

No

Human response to question

No Yes

If confidence >90%

Measure Watson/human accuracy over time

Figure 1. Question-and-answer (Q&A) process flow. This flow can be associated with any Q&A system.

data needed, the architectural approach to be used, the anticipated results, and what business impact the application will have. Table 1 displays the business objective statement for the IAA. With an appropriate development focus, development teams can then go about the process of leveraging IBM’s Bluemix platform (www.ibm .com/cloud-computing/bluemix) to quickly assemble the necessary services and deploy their application quickly.

Corpus Development The second step involves developing the corpus outlined in the data scope section of the business objective statement. The initial corpus of IAA consisted of official answers to anticipated FAQs that had already been compiled by various university departments. This is shown as the Q&A database in Figure 1. This database was sufficient for the initial development step and initial application deployment. As Figure 1 illustrates, the ASK@ corpus will later be expanded to include information contained in other written materials, such as 38


the BI&A program’s brochure, syllabi, and website content.

Initial Development The third step is to train the application using the initial Q&A corpus. In the IAA project, because the Q&A database required no specialized knowledge and was authoritative in nature, the initial training was performed by the developers. In less-restricted applications, domain experts might be necessary for this step. In any case, the initial development step resulted in the development of synonyms and antonyms, and application-specific vocabulary and abbreviations. As we describe next, the training step utilizes cognitive computing’s learning capability together with the evaluation metrics.

Continuous Improvement The fourth step, continuous improvement, involves learning during actual system use. For example, as we describe later, the IAA application will be made available to potential and current

Table 1. Intelligent Academic Advisor business objectives. Component

Description

Business problem

The program receives hundreds of questions from potential applicants, including many frequently and some infrequently asked questions. Engaging the interest of potential applicants can be important in both differentiating our program and ensuring that interest eventually turns into applications.

Business question

Can the program answer both frequently as well as infrequently asked questions from potential applicants?

Data and scope

The training data includes emails with frequently asked questions (FAQs) from prospects. The document corpus encompasses Business Intelligence and Analytics program, School of Business, and Stevens Institute of Technology websites.

Architectural approach

The architecture uses the Watson Retrieve and Rank and Natural Language Classifier APIs.

Anticipated results

The application (if available via the web or as an app) will help potential applicants find answers to their questions; it will help differentiate our program and will potentially turn interest into applications.

Business impact

Turning interest into applications has a significant benefit to the school as well as to the potential applicant (by saving time, reducing frustration, and helping match the interests of our audience with our offerings). If successful, the application can be expanded from the program to the School of Business and ultimately to Stevens.

student users, and the corpus will be expanded to include material other than the authoritative Q&A material in the initial application. The feedback from actual use will be used to improve both system accuracy and the range of questions that it can faithfully answer.

Watson System Components A critical component of developing and ultimately enhancing a system such as the IAA is for the developers (in our case, business school students) to have access to both the tools and a platform on which to run them. For our developer audience, we leverage Bluemix, which facilitates students working in small, agile development teams. As Figure 1 shows, the initial ASK@ application facilitated users finding the answers to specific questions related to a specific corpus. As the system is enhanced, it will begin to advise and recommend based on the context of the question being asked and the individual doing the asking (applicant, student, graduate, and so on). Table 2 describes the Bluemix services that are relevant to our system. The initial instantiation of the IAA used a hybrid combination of the Natural Language Classifier and Retrieve and Rank (R&R) services.7 Over time, enhancements will leverage the remainder of the services.

Classification and Confidence A critical objective in developing a cognitive assistant such as ASK@ is to have some means of measuring accuracy in question classification, answer search, and retrieval accuracy, as well as Watson versus human answer accuracy over time. Table 3 shows four measures of specific interest. By collecting these measures as the system is used, we will obtain a better feel for how well it learns, first by better understanding the question being asked and then by judging the accuracy of the answers obtained. As students and potential applicants begin to ask questions that are not easily found with a keyword search, the system will more closely approach a human’s answering accuracy.

Developing the Initial Q&A Corpus An initial set of 293 questions and corresponding answers was developed manually. Most of the questions and answers were taken directly from the FAQs links on the BI&A program, School of Business, and Stevens Institute websites, as well as the Registrar, Admissions, and Career Development websites. As stated earlier, we envisaged three main groups of users (clients): potential students and BI&A program applicants (collectively, applicants), current students, and potential employers.

computer.org/ITPro

39

COGNITIVE COMPUTING

Table 2. IBM Bluemix services. Watson service

Description

Natural Language Classifier

Applies cognitive computing techniques to return the best-matching classes for a sentence or phrase

Retrieve and Rank

Helps users find the most relevant information for their query using a combination of search and machine learning algorithms to detect “signals” in the data

Document conversion

Converts a single HTML, PDF, or Microsoft Word document into a normalized HTML, plain text, or set of JSON-formatted answer units that can be used with other services

Dialog

Allows a developer to design the way an application interacts with an end user through a conversational interface

Concept insights

Links documents that you provide with a preexisting graph of concepts based on Wikipedia (for example, “machine learning”)

Relationship extraction

Parses sentences into their various components and detects relationships between the components, which include parts of speech (noun, verb, adjective, conjunction, and so on) and functions (subjects, objects, predicates, and so on)

Text to speech

Provides a REST API (RESTful web service) to synthesize speech audio from an input of plain text; multiple voices, both male and female, are available across Brazilian Portuguese, English, French, German, Italian, Japanese, and Spanish

Speech to text

Bridges the gap between the spoken word and its written form, using machine intelligence to combine information about grammar and language structure with knowledge of the composition of an audio signal to generate an accurate transcription

Table 3. System accuracy measures. Measure

Description

Question classification accuracy

Determines the number of questions that are accurately classified over time by tracking classification confidence. The classification training database will be updated to improve classification accuracy as the corpus grows in the specific domain.

Answer confidence level

Determines the number of answers that are accurately obtained from the corpus over time by tracking response confidence. The question-and-answer (Q&A) database will be updated with answers that receive a high confidence rank (greater than 90%). For answers that receive a low confidence rank (less than 90%), either the Q&A database or the corpus will be updated.

Search and retrieval accuracy

Determines the number of search queries resulting in the most accurate answer over time by tracking the response rank. It also tracks whether answers retrieved from the Q&A dataset actually match the search criteria (“Did you find the answer to your question?”).

Watson/human accuracy

Determines the number of questions that are accurately answered by Watson vs. humans over time by tracking classification confidence. As the Q&A dataset and domain corpus grow, the number of correct questions and answers provided by Watson should eventually surpass that provided by humans.

A simple ontology was developed to guide question generation. The ontology consisted of each of these client entities and the organizational elements of the university, school, and program with which the clients interact. The interaction between these entities occurs through processes and events. Questions regarding the entities themselves and the associated processes 40


and events then followed naturally. The objective of proceeding in this way was to develop a complete set of categories for the questions. The number of questions and answers in each category is expected to grow as the system learns, but the categories helped facilitate the generation and organization of the initial set of questions. Figure 2 shows the top level of the ontology and

50 45 40 35 30 25 20 15 10 5

s on si is

dm 9A

ro

3S

gr

am

re

qu

7I

ire

nd

m

us

en

ts

try

ty ul ac 6F

um ric

ul

ra

ur

pr A I&

4B

fB lo oo

ch

5C

og

in us

te 2S

m

s es

ns ve

er oy

Em

pl

to 15

en em

14

Pl

ac

s

ce ffi

en ud st

nt rre Cu

13

Re

Fi

na

8P

nc

12

ts

e of

ra gi

st

up /s id

la ia

fic

e rt

po

nt de tu

ls

r's

of

fic of

ok ob 1H na io te

rn

at 11

In 10

fic

e

en

0

Total

Figure 2. Top question categories. Most questions and answers were generated in the “BI&A program” and “curriculum” categories.

the number of questions generated initially in each category. As shown in the figure, most questions and answers were generated in the “BI&A program” and “curriculum” categories. It remains to be seen whether the numbers of questions and answers in the other areas are sufficient to satisfy the demand for information in these areas—“financial aid/support” is a case in point. The major challenge regarding the academic domain versus other domains is that there are questions outside of the specific academic area (such as lifestyle or housing) with more than one particular answer. This can be seen in Figure 2, where we have a significant number of questions with specific answers in the “program,” “curriculum,” “requirements,” and “admissions” categories but far fewer in other question categories (for instance, “international student office,” “Stevens,” or “faculty”). Narrowing the cognitive application to just one

or two categories would have made the effort easier but would have resulted in a less versatile solution.

Initial Development Apache Solr performs a full-text search of the corpus and returns possible answers based on matching keywords. R&R is built on top of Apache Solr. It contains several proprietary machine learning techniques known as learning-to-rank algorithms and chooses the best combination of algorithms for the training data to improve error handling and resiliency. In the real world, words are often used interchangeably. A natural language application interface must be able to interpret such words. While training the R&R model, a supplementary text file containing a list of synonyms and keywords was provided for this feature. During the training process, the model builds the corpus along with synonyms lists and keywords.

computer.org/ITPro

41

COGNITIVE COMPUTING

Figure 3. Question interface for the Intelligent Academic Advisor application. The cognitive system returns a response to the user and ranks the answers.

As a result, whenever a new word or keyword that is not a part of training data is found, the Solr search agent fetches all the possible synonyms of it from the synonyms list, and that list is searched against the corpus. This is often known as paraphrasing. Using this paraphrasing technique, we increase the possibility of fetching more results. Synonyms, keywords, or peculiar words must be mapped to a more generic word to get the best performance benefit from this Solr capability. For example, in the IAA application, we incorporated synonyms for “course fee” as “expense,” “cost,” “term fee,” and so on. For keywords such as “Hoboken,” we included the frequently used term “HBKN.” This will provide the same search results for “Hoboken” and “HBKN” whenever it is used in the user query. Application logs are useful in determining application performance and improving performance over time. Application logs are different from the service logs generated and collected by the Bluemix service container, which logs every request made to the service handler. The service logs contain information about service events, such as the service start and stop time stamp, the application identifier of each HTTP request, response time, and service requests made from all of the applications. This log is often used to debug the service instance performance and instance failures using 42


the global transaction ID associated with every service instance. Application performance is recorded in the application logs, which are implemented at the application level. It is used to track the user inputs and browse activity events along with the system response. In the IAA application, we used user inputs and system responses to improve the corpus and the overall application. For every user query, the system responds with a set of answers and the confidence scores associated with these answers. If an answer set contains answers with confidence scores lower than a fixed threshold value, we conclude that we need to add more relevant answers to the corpus document. When the model is retrained with the new corpus, the confidence levels will improve. Similarly, we can keep track of the most frequently asked questions and the users’ evaluations of those answers. If certain answers are not widely accepted, we might need to add more relevant answer documents. In this way, we keep track of poorly performing user queries and add relevant answers to increase the application’s overall performance. In a multi-user environment, the size of the application logs can be huge. In such cases, we consider embedding a third-party log-management service, such as Splunk, which provides better collection and analytics services on application logs. Figure 3 shows the question interface for the IAA application and the response from the cognitive system. The arrow and number to the left show how many places the ranker moved the answer from Solr and in which direction these results moved based on the model that was trained. So, in the example in the figure, the first result was originally the fifth result in Solr, so it moved up four spots; the second result did not move at all; and the third result was originally the first result in Solr but moved down two spots. “Domain expert review” on the right shows the relevancy assessment that was part of the ground truth for this result. The more bars, the higher the relevancy. So, results with more bars on the top set of results mean the ranker is doing a good job of reranking the results. In Figure 3, the R&R system has combined several discrete answers to provide a satisfactory answer to the English Language Competency

query. In the IAA’s current development state, the response to other user queries might not be as complete and accurate. In these cases, performance will improve with use during the continuous improvement phase we describe next.

Continuous Improvement As mentioned previously, our application leverages a hybrid combination of the Natural Language Classifier service (for better understanding the question’s context) and the R&R service (for increasing confidence surrounding the answer to the question) (see http://rar-stevens-faq-demo. mybluemix.net). The classifier helps to narrow the context for the question based on training data. As more variations in how questions are asked find their way into the training data, a higher confidence level will indicate improved classification accuracy. R&R focuses on identifying the answer or set of answers most likely to be associated with the question. The service is built on top of Apache Solr and gives results based on a combination of search and machine learning algorithms. Only the highest-ranking results, based on response confidence, are returned as answers. In addition to improvement surrounding the corpus, we also see improvement in how the corpus is used. The first enhancement will be in the question response interface (such as voice input and output). The second enhancement will be in enabling the application to address multiple questions (for instance, dialog).

A

s mentioned, the advantage of the IAA application for our research is its simplicity. It addresses a relatively simple domain, and, consequently, it is relatively easy to develop the corpus of knowledge, train the system, and gain good assessments of its performance in terms of accuracy, ease of use, and user satisfaction. The initial IAA application demonstrates the power of a cognitive assistant in helping to find answers to frequently and not so frequently asked questions. To enhance its capability, we are looking to both expand its corpus of knowledge (by including additional documents and webpages) and enhance its ability to handle multiple questions and individual user context.

The IAA application offers significant value to those who use it, becoming more authoritative over time and reducing the burden of faculty advisors and administrators having to answer hundreds of questions daily. In addition, because this particular cognitive system is part of a semesterlong course in Cognitive Computing in Business, it exposes student teams with diverse skills and backgrounds to both developing and deploying cognitive systems. Building cognitive business applications requires a combination of both business focus and technical skill. The business focus helps the development team to best identify those opportunities suitable for cognitive development. Technical skills (for example, Python and Java programming) help in actually developing and deploying cognitive solutions. Having a comprehensive platform such as Bluemix on which to build cognitive applications, as well as a set of template applications (with associated code and documentation contained in GitHub repositories) that can be extended and enhanced, allows the development team to get its prototype up and running as quickly as possible.

References 1. J.E. Kelly and S. Hamm, Smart Machines: IBM’s Watson and the Era of Cognitive Computing, Columbia Univ. Press, 2014. 2. E. Brynjolfsson and A. McAfee, The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies, W.W. Norton & Company, 2016. 3. J. Markoff, “Computer Wins on ‘Jeopardy!’: Trivial, It’s Not,” New York Times, 16 Feb. 2011; nyti. ms/1AkYdWx. 4. J. Maderer, “Artificial Intelligence Course Creates AI Teaching Assistant,” News Center, Georgia Tech., 9 May 2016; b.gatech.edu/1TaImCG. 5. J.E. Kelly III, Computing, Cognition, and the Future of Knowing: How Humans and Machines Are Forging a New Age of Understanding, IBM white paper, 2015; ibm. co/1LJ6K9Z. 6. J. Hurwitz et.al., Cognitive Computing and Big Data Analytics, John Wiley, 2015. 7. C. Ackerson, “Developing with IBM Watson Retrieve and Rank,” blog, 6 Apr. 2016; bit.ly/2pXYhfK.

Christopher Asakiewicz is an industry professor, program director of the Business Intelligence and Analytics program, and director of the Stevens Alliance for Innovation

computer.org/ITPro

43

COGNITIVE COMPUTING

and Leadership at the School of Business, Stevens Institute of Technology; and president of IT Strategy and Management Consulting. His research interests include collaborative research and discovery, cognitive systems, and analytics. Asakiewicz has a strategic background in business technology management, information technology, consulting, and education, and a 21-year history as a vice president of global business technology at Pfizer. Contact him at christopher. [email protected]. Edward A. Stohr is a professor of information systems, codirector of the Center for Technology Management Research, and program coordinator for the MS in Business Intelligence and Analytics program at the School of Business, Stevens Institute of Technology. His published research focuses on the problems of developing computer systems to support work and decision making in organizations. Stohr was a faculty member at New York University’s Stern School of Business for more than 20 years. Contact him at [email protected].

Shrey Mahajan is a business continuity analyst at Continuity Logic. His research interests include emerging technologies, business, and investment. Mahajan received an MS in information systems from the Stevens Institute of Technology. Contact him at [email protected]. Lalitkumar Pandey is a business intelligence consultant professional at Capgemini. His research interests include efficient information retrieval methods and cognitive machine learning. Pandey has four years of experience at Teradata. He received an MS in computer science from the Stevens Institute of Technology. Contact him at [email protected].



Harlan D. Mills Award

Call for Software Engineering Award Nominations Established in memory of Harlan D. Mills to recognize researchers and practitioners who have demonstrated long-standing, sustained, and impactful contributions to software engineering practice and research through the development and application of sound theory. The award consists of a $3,000 honorarium, plaque, and a possible invited talk during the week of the annual International Conference on Software Engineering (ICSE), co-sponsored by the Deadline for 2018 Nominations: IEEE Computer Society Technical Council on Software Engineering.

1 October 2017 Nomination site: awards.computer.org

44


The award nomination requires at least 3 endorsements. Self-nominations are not accepted. Nominees/nominators do not need to be IEEE or IEEE Computer Society members.

COGNITIVE COGNITIVECOMPUTING COMPUTING

Cognitive Communication in Rail Transit Awareness, Adaption, and Reasoning Cheng Wu and Yiming Wang,

Soochow University, China

Applying cognitive radio in railway wireless communication systems is a cutting-edge research field. The authors’ rail cognitive radio model uses a Bayesian network with channel contextual features to probabilistically infer the likelihood of spectrum accessibility.

H

igh-speed rail is an important part of modern railway transportation, which is closely related to humans’ daily activities. Countries worldwide have built or are building large-scale high-speed rail. For example, a plan for high-speed rail has been included in the US’s technology development outline.1 It is estimated that the length of

1520-9202/17/$33.00 © 2017 IEEE

the high-speed rail network in China will reach 50,000 km by 2020.2 With the continuous construction of highspeed rail in recent years, wireless communication technology plays an increasingly vital role in its daily operation, safety monitoring, and efficiency improvement.3 There are three requirements for the use of wireless communication technology


computer.org/ITPro

45

COGNITIVE COMPUTING Core network Internet

Control center

SGW AGW

AGW

(iii)

(i)

Aggregation network Satellite

LTE-R Wayside devices

GSM-R WLAN CDMA

RoF

Access network TAT

(ii) High-speed train

(a)

Train network

(b)

Figure 1. Network model of high-speed railway wireless communication. (a) There are three typical concerns when using railway wireless communications: information exchange, data collection, and Internet connectivity. (b) The network architecture has four layers. (AGW: access gateway; SGW: service gateway; TAT: train access terminal.)

in modern railway transportation. The first is information exchange between the train and the control center (for example, download the track database, upload the train speed and location), as Figure 1a(i) shows. Second, more and more sensor devices have been deployed along rail lines and are used to collect environmental data, surveillance images, and device status,4 shown in Figure 1a(ii). Third, and most importantly, with the development of mobile Internet and social networks, people need to maintain Internet connectivity moment by moment. Passengers’ data expectations have evolved toward being able to receive and send information (email and messages) continuously or stream high-quality, realtime video anytime and anywhere on a journey,5 as Figure 1a(iii) shows. The evolution of human behaviors in using wireless communications places higher demands on the future wireless communication networks of high-speed rail. On one hand, railway wireless communication must be able to provide broad46


band data services. This requires that the base stations (BSs) along the railway provide enough spectrum to transmit communication signals.6 In fact, spectrum scarcity is a common, inherent problem of wireless communication.7 Spectrum itself is an expensive and limited resource. It is mandated by government agencies for long-term operator use that often spans several decades. So far, the spectrum of 6 GHz and below has been almost entirely allocated.8 In addition, the insatiable appetite of current wireless users has led to the rapid growth in 4G LTE networks and is driving the industry to 5G at an unprecedented pace. Although the growth of bandwidth and equipment has fueled the industry’s boom, the amount of available spectrum has not kept pace with it. On the other hand, railway wireless communication networks must meet the special demand of providing stable dataflow throughout highspeed movement.9 During high-speed movement, existing vehicle wireless communication

technology often demonstrates considerable performance instability and spectrum uncertainty. When speeds are up to 350 km per hour, problems inevitably occur, such as fast handoff, Doppler shift, and penetration loss.10 Fast handoff is a key factor in reducing communication performance. For overlapping regions of a certain size, the faster the moving speed, the shorter the time spent crossing the overlap region. When the time that the train traverses the overlap region is less than the minimum delay required to handle a handoff, the handoff process cannot be completed.11 Therefore, the efficient management of wireless spectrum for scarcity and uncertainty in motion is becoming key to future railway wireless communication. This article aims to establish an achievable railway cognitive communication model to address the key issues of spectrum scarcity and uncertainty in motion. By introducing the physical structure of the railway wireless communication network, we determine that uncertain spectrum accessibility in motion is the key special issue for wireless communication in rail transit, beyond the common problem of spectrum scarcity. To solve this uncertainty of spectrum accessibility, we propose adding a cognitive reasoning module with a Bayesian network as the core of the cognitive engine. At the same time, we discuss the channel contextual features and performance symptom features that have a strong causal relationship to spectrum accessibility. This formulation is expected to open a new chapter in the study of cognitive communication in rail transit.

Wireless Communication in High-Speed Rail Figure 1b illustrates a typical network architecture for broadband wireless communication in highspeed rail. This multilayer architecture supports data transmission among fast-moving trains, wayside sensors, the railway control center (RCC), and the Internet. It consists of four layers.12 The core network provides data services for wireless communication in high-speed rail, which includes two major actors: the Internet and the RCC.13 The Internet provides multimedia services for train passengers’ entertainment, and the RCC enables railway signaling transmissions and remote access to devices on trains and at waysides for data acquisition and system maintenance.

The aggregation network is responsible for data aggregation and forwards data from the access network to the core network. It has two kinds of gateways: the access gateway (AGW) and the service gateway (SGW). The AGW serves as an interface connecting the access network and the aggregation network, and the SGW serves as an interface connecting the aggregation network and the core network. The AGW gets data from the access network and forwards data to the SGW.14 The access network is close to the railway and delivers data services directly to the train network. It provides the last hop in the data links. Different kinds of wireless access technologies are used to provide data access for high-speed trains. They work together to satisfy the requirements of data services, such as a high data rate, low latency, and fast handover. Finally, the train network provides broadband Internet access for train passengers using a train access terminal (TAT). The TAT connects to devices on the access network through an antenna mounted outside the train body. The signals received from the TAT are then fed to gateways and access points on every train car for data access or to the control terminal for train control. This wireless communication network architecture supports the entire road network of highspeed railway. Normally, this network covers a large geographical area. The train in motion often encounters different geological conditions and electromagnetic environments, which inevitably results in noise from continuous change and difficulty detecting sources of interference. Ultimately, fast motion affects channel quality and results in a change in spectrum availability. Consequently, the spectrum holes in the BS group are unstable. Furthermore, the high-speed train passes quickly through the coverage areas of multiple BSs. Each BS has different licensed devices with unique spectrum occupancy rules, leading to discontinuity in the spectrum holes among the BS group. A novel architecture with cognitive capability is thus necessary for train communication to solve the problem of spectrum accessibility uncertainty when trains are in high-speed motion.

Cognitive Radio in Rail In fact, although spectrum is licensed to operators for long-term use, a large portion of this assigned spectrum is used sporadically, leading to

computer.org/ITPro

47

COGNITIVE COMPUTING Radio environment

Radio environment

Radio stimuli

Radio stimuli Radio parameters

Spectrum sensing Demand request

Spectrum mobility

ACT Spectrum mobility

Spectrum sensing

Demand request Spectrum sharing

PUs detected?

Channel capacity

Yes

Cognitive devices

Yes Enter overlapped?

Channel capacity

Decide Spectrum decision Yes

QoS bad? Yes Orient

GPS info policy

Signal transmit

Spectrum sharing

Sense

Physical situation

Signal transmit

Radio parameters

Train network

Physcial sensing

Policy changed?

Spectrum decision Access network

(a)

(b)

Figure 2. Cognitive model. (a) The cognitive radio cycle helps to solve the spectrum scarcity problem. It is improved in (b) the rail cognitive radio cycle.

low utilization and the waste of valuable frequency resources. To address this critical problem, the US Federal Communications Commission (FCC) has recently approved the use of unlicensed devices in licensed bands.15 Consequently, dynamic spectrum access techniques have been proposed to solve current spectrum inefficiency problems.16–18 This new area of research foresees the development of cognitive radio (CR) technology to further improve spectrum efficiency. The basic idea of CR is that unlicensed devices (also called CR users or CUs) share wireless channels with licensed devices (also known as primary users or PUs) that already have an assigned spectrum, and must thus vacate the band once the PUs’ activities are detected.19 To achieve this, CUs must continuously monitor spectrum for the presence of PUs and reconfigure the radio front end according to the demands and requirements of the higher layers. This capability can be realized, as shown in Figure 2a, by the cognitive cycle, which comprises the following spectrum functionalities: determining the portions of spectrum currently available (spectrum sensing), selecting the best available channel (spectrum decision), coordinating access to this channel with other users (spectrum sharing), and effectively 48


vacating the channel when a PU is detected (spectrum mobility). The cognitive cycle lets CR effectively solve the spectrum scarcity problem. Applying CR technology in high-speed rail has attracted much interest in the research community and government agencies. Rail spectrum scarcity is the research focus at this stage. The US Federal Railroad Administration (FRA) proposed a genetic-algorithm-based CR method to improve the performance of train control systems to achieve train safety and efficiency.20 The French Urbanisme des Radio Communications project was one of the first projects in France and Europe to raise the problem of optimizing spectral resources in the Paris region while considering the transport field and, particularly, urban guided systems.10 These innovations have introduced cognitive technology into railway wireless communications to a certain extent to alleviate the scarcity problem. However, the spectrum accessibility uncertainty caused by high-speed motion still exists. The next step should be enhancing the binding with the specificity of railway wireless communication. For general wireless networks, the cognitive cycle gives too much consideration to

Table 1. Improved rail cognitive radio model.

PUs’ spectrum occupancy rules while ignoring other factors. This is because the emergence of a CU is often sudden, and there is no fixed spectrum occupancy rule to follow. However, the railway wireless network is different. Trains usually run according to specific schedules. Under ideal conditions, the train always passes regularly and periodically through BS coverage areas. This phenomenon enables the cognitive cycle in rail transit to consider the impact of other factors on communication performance, such as potential physical interference, other user activities, and spatial-temporal changes in the spectrum.

Improved Cognitive Radio Model The principle of rail-CR is to combine AI-based decision-making theory and learning algorithms with a soft-defined radio (SDR) platform to address railways’ needs. Figure 2b shows an improved rail-CR model. This model significantly evolves an SDR platform into four modules (see Table 1). Combined with the functionalities of the traditional cognitive cycle, the model enables the use of spectrum sensing to observe the rail radio environment and physical sensors (such as GPS, acceleration sensors, and temperature sensors) to determine current physical situations; orient to the particularity of spectrum policies and goals; apply spectrum decisions to seek an optimal way to change the configurations of the SDR for adaptation; and use spectrum sharing and mobility to apply those changes to the SDR. In this model, the sensing module provides four trigger conditions for the decision module. The triggering conditions from radio environment sensing are whether any PU is detected using the occupied spectrum and whether the occupied spectrum’s quality of service (QoS) supports poor data transmission (blue diamonds in Figure 2b). The triggering conditions from physical situation sensing are whether the train is running into overlapping BS areas and whether the spectrum policy of the area where the train is located changes (green diamonds in Figure 2b). The improved rail-CR model provides a new SDR platform for high-speed rail wireless communication. The nature of this model is to solve the problem of spectrum scarcity—that is, cognitive users can dynamically and opportunistically access the spectrum holes for transmission with-

Module

Description

Sense

Radio environment sensing and physical situation sensing

Orient

How to reach particular goals and fit within local policies

Decide

The changes needed to adapt to a situation

Act

How to apply the new configuration to the radio environment

out affecting PUs. More importantly, it provides a mechanism for resolving the uncertain spectrum accessibility caused by high-speed movement by considering more contextual factors. This work is usually done by the cognitive engine, which is responsible for learning and predicting channel accessibility in radio operations and decision making about how the CU uses the spectrum holes.

Cognitive Reasoning for Spectrum Accessibility The uncertainty in rail-CR’s spectrum accessibility comes from the instability, discontinuity, and unpredictability of spectrum holes owing to the fast movement of the high-speed train itself. When the high-speed train enters the coverage region of a particular BS, the cognitive engine’s primary task is to select the channel with the best QoS for data transmission. On one hand, channel selection is triggered by various radio-dependent cues or physical-saliency cues; on the other hand, the cognitive engine arouses a series of communication performance symptoms. These cues and symptoms (also called clues) can be seen as a relationship between causes and effects that directly indicates spectrum accessibility status. To reduce the uncertainty in accessibility reasoning, more clues should be introduced in the reasoning process. Considering the cause-and-effect relationship among spectrum accessibility clues, the features from these clues can be divided into two types: channel contextual features and performance symptom features (Figure 3). Channel contextual features. The following three types of contextual features influence spectrum accessibility: • Licensed user activities. When a CU and a PU have a conflict over using the same channel in a CR network, the spectrum shared by the

computer.org/ITPro

49

COGNITIVE COMPUTING

Licensed user activity

Potential physical interference

Spatialtemporal changes xtual

Conte Performance symptom

s

s

feature

Channel accessibility

Successful transmission

Packet delay

Channel switching

Network throughput

Power consumption

Figure 3. A directed acyclic graph. The graph represents the causal relationship in spectrum accessibility.

CU would seriously affect the data transmission of the PU. This kind of concurrent spectrum occupancy goes against the principle of protecting licensed devices. Hence, for channel accessibility of CUs, PUs’ activities must be considered, which can be seen as the primary cause of lowering link throughput. • Potential physical interference. Some physical factors, such as multipath reflections, channel gain, thermal noise, propagation loss, transmission power, and reception gain, cause an unpredictable transmission failure or unacceptable QoS. We cannot ignore causes from varying physical factors. • Spatial-temporal changes. The movement of a highspeed train causes rapid changes in the train’s location and time zone. Different locations can mean changes in the surrounding geographical environment, whereas different times can mean that spectrum usage will change. These changes will affect the choice of spectrum channels.

Performance symptom features. Channel contextual features are determined through a priori knowledge that can influence spectrum accessibility under a single BS. Correspondingly, after choosing a channel, various performance symptoms can be used to validate whether the chosen channel is optimal. The following network metrics can be employed as performance symptom features: • Successful transmission. If link disconnection, 50


channel-induced errors, and transmission collision are not observed in a given transmission slot, then the packet is successfully transmitted from the sender to the receiver. • Channel switching. Channel switching is the action taken by the CU operating on a specific channel to switch to another channel. Each switch would result in a transmission delay, which could waste several milliseconds. In fact, such a delay could easily go beyond 10 ms, which would have a significant negative impact on performance, especially if consecutive switch operations occur. Another problem associated with consecutive channel switches is packet dropping. • Network throughput. After assigning the channel, either the train or the access point in urban scenarios will start to transmit data over the channel. By calculating the throughput of CUs, we can easily judge whether this assignment is optimal. • Power consumption. The transmitter and receiver of the train or BS will consume more power if the assignment is not good enough.21,22 For example, if some packets are dropped during receiving or the receiver finds out that there are occasional bit flips in the datastream, the transmitter and receiver might consume extra power to retransmit. • Packet delay. End-to-end delay or one-way delay refers to the time it takes for a packet to be transmitted across a network from the source to the destination. Waiting for a PU to leave or, in peer devices, for the transmission queue could cause packet delay. It is also another significant indicator for measuring network performance.23

Bayesian Network Model Clearly, having an abundance of features with a causal relationship to spectrum accessibility can help us explicitly evaluate whether the choice of spectrum is optimal. The difficulty is then in how to construct a model that can consolidate evidence from these features to reduce possible uncertainties during reasoning about spectrum accessibility. Quite sensibly, a Bayesian network might be a feasible technique for coping with such uncertainty; such networks provide consistent and reliable semantics for inferring causal relationships (causes and effects) using an

Channel contextual features

Licensed user activity

Spatial-temporal changes

Performance symptom features

Channel accessibility

Rail radio Bayesian network (multiple features fusion)

Potential physical interference

Successful transmission Channel switching Network throughput Power consumption Packet end-to-end delay Adjust

Figure 4. Bayesian network model of spectrum accessibility in rail cognitive radio. Channel contextual features and performance symptom features are fused to deduce channel accessibility, and accessibility is further fed back as an adjustment of performance symptoms.

intuitive, directed acyclic graphical representation. Thus, a probabilistic model based on Bayesian networks is used for cognitive reasoning about spectrum accessibility. The goal of a Bayesian network model is to infer the hidden spectrum accessibility hypothesis with some uncertainty from partially available contextual features or observed symptom features; thus, three inference steps are proposed. The first is to set spectrum accessibility as the target hypothesis for which spectrum selection is obviously the target variable that we want to infer. The step then identifies the corresponding feature variables with a strong causal relationship to spectrum accessibility. Contextual features that can influence spectrum accessibility include PU activity, potential physical interference, and spatialtemporal changes caused by high-speed movement. Symptom features, which are the effects of spectrum selection, include successful transmission, channel switching, throughput, power consumption, and packet delay.

The second step in CR is to construct a directed acyclic graph by grouping channel contextual features into parent nodes and performance symptom features into child nodes according to their causal hierarchical relations with spectrum accessibility. The last step is to calculate a conditional probabilities table and reason the likelihood of spectrum accessibility. Combining all three steps together, the Bayesian network model of spectrum accessibility is constructed as shown in Figure 4. In rail-CR, spectrum selection sends an adjust back to the radio environment, which would further influence performance symptom features until the dynamic network converges to a stable status.

Validation We use network simulation tool NS-2 to develop our wireless communication framework (Figure 5). The construction of the railway environment, CR functionalities, and the reasoning mechanism are realized by an independent C11 module in which the protocols of the physical

computer.org/ITPro

51

COGNITIVE COMPUTING

Bayesian reasoning module Set target hypothesis and feature variables

Construct directed acyclic graph with conditional probability

Calculate the likehood of spectrum accessibility

Real-time cognitive radio module Railway physical situation repository

Railway wireless environment repository

Spectrum sensing Spectrum decision

• GPS info • Radio policy • Base station location

Physical sensing

• TX power • Spectrum • PU activity

Spectrum sharing

Spectrum mobility

…

…

Real-world simulation platform

Figure 5. Simulation platform of our rail cognitive radio framework. The Bayesian reasoning module works as the core of spectrum decision making, directing the real components of the simulation platform to achieve spectrum management.

layer, link layer, and network layer are modified accordingly. The real-time cognitive radio module supports the CR architecture, which consists of the sense, orient, decide, and act modules. The reasoning module describes the Bayesian network framework and some common learning functions. The railway wireless environment repository contains information on spectrum settings and transmission power, and a two-state Markov model to define PU activities. The railway physical situation repository contains information on train speed and location, radio policy, and BS locations and coverage, as well as different network protocols for railway environments. The real-world simulation platform provides a simulation railway communication environment along with a fixed rail line. There are 10 BSs and a total of 10 available spectrums. Each BS has 10 PUs. Each PU is authorized to use a dedicated spectrum for data transmission. The spectrum occupancy of each PU is different and is described by a two-state Markov model with the randomly generated rate parameters 52


,lbusy, lidle.. On our simulation platform, each tower is used to simulate a BS. We use “on” and “off” LED lights with different colors to represent whether the PU is occupying its spectrum. The PU spectrum occupancy rules are generated by stand-alone workstations. The communication function of each BS is done through the Zigbee node. The CU detects the spectrum hole and allocates spectrum every 0.1 seconds. The trains run along the rail line at an average speed of 10 meters per second. Each run of the train can be considered an episode. The train runs for 60,000 episodes and is considered an experiment. To evaluate the proposed BN-based method, we compare the performance of the BN-based, genetic-algorithm-based,20 and Round Robin (RR)-based schemes. The RR-based scheme adopts the following principle: once the PU occupies the spectrum, the CU switches to the next channel in turn with the same probability. No channel has priority. This method is simple, easy to implement, and will not “starve to death.” We

R

ail-CR is an effort that recently launched and is still in its initial stages. It is a typical application of cognitive computing in the field of rail transit wireless communication. It is used to cope with the increasingly complex wireless communication behaviors of rail transit, meet communication requirements, and enhance communication efficiency. Next steps for our research will focus on realizing the CR module in the cognitive engine of rail transit. In future experiments, we will use a large amount of training to obtain a priori conditional probabilities of channel contextual features under spectrum accessibility and spectrum accessibility under performance symptom features. In addition, the CR module has many other research needs, including rail-CR-specific case-based reasoning on network topology, more channel contextual features and performance symptom features associated with spectrum accessibility, and cognitive algorithms for multi-BS collaboration.

References 1. D.R. Peterman, J. Frittelli, and W.J. Mallett, The Development of High Speed Rail in the United States: Issues and Recent Events, Congressional Research Service, 2012. 2. K. Smith, “China Steps up Railway Diplomacy in East Asia,” Int’l Railway J., 3 Feb. 2016; bit.ly/2ryGq4b. 3. R. He et al., “High-Speed Railway Communications,” IEEE Vehicular Technology Magazine, Sept. 2016, pp. 49–58. 4. V.J. Hodge et al., “Wireless Sensor Networks for Condition Monitoring in the Railway Industry: A Survey,” IEEE Trans. Intelligent Transportation Systems, vol. 16, no. 3, 2015, pp. 1088–1106. 5. A. Amanna et al., “Railway Cognitive Radio,” IEEE Vehicular Technology Magazine, Sept. 2010, pp. 82–89. 6. B. Ai et al., “Challenges toward Wireless Commu-

0.8 0.6 0.4 RR-based GA-based BN-based

0.2 0

0

10,000

20,000

(a)

30,000

40,000

50,000

60,000

50,000

60,000

Number of epoch 70 60

Average number of channel switching

consider two performance indicators, such as the probability of successful transmission and the number of spectrum switches, for results validation. Figure 6 shows the simulation results on network performance. The results confirm that the Bayesian network-based scheme can significantly improve the network’s data transmission rate through a period of reasoning, avoid mistakes in channel switching, and optimize the efficiency of spectrum management.

Average probability of successful transmission

1

50 40 30 20

RR-based GA-based BN-based

10 0

(b)

0

10,000

20,000

30,000

40,000

Number of epoch

Figure 6. Simulation results on network performance. The results confirm that our Bayesian network-based scheme can significantly improve the network’s data transmission rate. nications for High-Speed Railway,” IEEE Trans. Intelligent Transportation Systems, vol. 15, no. 5, 2014, pp. 2143–2158. 7. W. Zame et al., “Cooperative Learning and Coordination for Cognitive Radio Networks,” IEEE J. Selected Areas in Communications, vol. 32, no. 3, 2014, pp. 1–30; bit.ly/2pQCcjb. 8. J. Yang and H. Zhao, “Enhanced Throughput of Cognitive Radio Networks by Imperfect Spectrum Prediction,” IEEE Communications Letters, vol. 19, no. 10, 2015, pp. 1738–1741. 9. L. Tian et al., “Seamless Dual-Link Handover Scheme in Broadband Wireless Communication Systems for High-Speed Rail,” IEEE J. Selected Areas in Communications, vol. 30, no. 4, 2012, pp. 708–717. 10. M. Berbineau et al., “Cognitive Radio for High Speed Railway through Dynamic and Opportunistic Spectrum Reuse,” Proc. 5th Conf. Transport Research Arena (TRA): Transport Solutions from Research to Deployment, 2014; trid.trb.org/view.aspx?id51327710. 11. K. Li et al., “Cooperative and Cognitive Wireless Networks for Train Control Systems,” Wireless Networks, vol. 21, no. 8, 2015, pp. 2545–2559. 12. S. Xu et al., “A Survey on High-Speed Railway Communications: A Radio Resource Management Perspective,” Computer Comm., July 2016, pp. 12–28; bit.ly/2rNbDwK.

computer.org/ITPro

53

COGNITIVE COMPUTING

13. D.T. Fokum et al., “A Survey on Methods for Broadband Internet Access on Trains,” IEEE Communications Surveys & Tutorials, vol. 12, no. 2, 2010, pp. 171–185. 14. K.D. Singh et al., “Optimizing TCP over Cognitive Radio Networks for Trains,” Proc. 12th Int’l Conf. ITS Telecommunications Optimising, 2012, pp. 678–682. 15. Spectrum Policy Task Force Report, US Federal Communications Commission, ET docket no. 02-155, Nov. 2002. 16. I.F. Akyildiz et al., “NeXt Generation/Dynamic Spectrum Access/Cognitive Radio Wireless Networks: A Survey,” Computer Networks J., Sept. 2006, pp. 2127–2159. 17. S. Haykin, “Cognitive Radio: Brain-Empowered Wireless Communications,” IEEE J. Selected Areas in Communications, vol. 23, no. 2, 2005, pp. 201–220. 18. J. Mitola III, “Cognitive Radio for Flexible Mobile Multimedia Communication,” Proc. IEEE Int’l Workshop Mobile Multimedia Communications (MoMuC), 1999, pp. 3–10. 19. J. Mitola and G.Q. Maguire, “Cognitive Radio: Making Software Radios More Personal,” IEEE Personal Communications, vol. 6, no. 4, 1999, pp. 13–18.

Cutting Edge

stay on the

of Artificial Intelligence IEEE

J a n ua ry/ f E b r ua ry 2 016

Also in this issue: aI’s 10 to Watch January/FEBruary 2016

IEEE

real-Time Taxi Dispatching

56 68

Cheng Wu is an associate professor at the School of Urban Rail Transportation, Soochow University, China. His research interests include neural and fuzzy systems, dynamic multi-agent systems, intelligent human-computer interaction, and big data technology and applications. Wu has authored or coauthored more than 20 publications in various international journals and conference proceedings. He received a PhD in computer engineering from Northeastern University. Contact him at [email protected]. Yiming Wang is a full professor at the School of Urban Rail Transportation, Soochow University, China. Her research interests include wireless communications, cognitive wireless sensor networks, and intelligent transportation technology and applications. Wang has authored or coauthored more than 60 publications in various international journals and conferences. She received a PhD in communications from Nanjing University of Posts and Communications. Contact her at [email protected].

IEEE Intelligent Systems provides peer-

from flu Trends to Cybersecurity 84

P

U

T

T

I

N

G

A

I

I

N

T

O

P

R

A

C

T

I

C

reviewed, cutting-edge articles on the

E

Online BehAviOrAl AnAlysis

theory and applications of systems that perceive, reason, learn, and

VOLuME 31

act intelligently.

Cover-1

January 11, 2016 6:06 PM

The #1 AI Magazine

www.computer.org/intelligent


IEEE

nuMBEr 1

www.computer.org/intelligent

IS-31-01-C1

54

20. Railway Cognitive Radio to Enhance Safety, Security, and Performance of Positive Train Control, tech. report, US Dept. Transportation, Federal Railroad Administration, Feb. 2013. 21. A. He et al., “Minimizing Energy Consumption Using Cognitive Radio,” Proc. 2008 IEEE Int’l Performance, Computing, and Communications Conf., 2008, pp. 372–377. 22. A. He et al., “System Power Consumption Minimization for Multichannel Communications Using Cognitive Radio,” Proc. IEEE Int’l Conf. Microwaves, Communications, Antennas, and Electronics Systems, 2009, pp. 1–5. 23. S.C. Lin, “End-to-End Delay Reduction via In Network Computation in Cognitive Radio Sensor Networks,” Proc. IEEE Global Communications Conf. (GLOBECOM), 2013, pp. 408–413.



COGNITIVE COMPUTING

Cognitive Gaming

Wei Cai, Yuanfang Chi, and Victor C.M. Leung, The University of British Columbia

As intelligent networked computing becomes pervasive, an emerging trend is to apply a cognitive computing paradigm to video game design and development. This article describes the concept of cognitive gaming and discusses its enabling architecture.

R

ecent hot topics in research, including the Internet of Things, cloud computing, and virtual reality (VR), can all benefit from powerful cognitive systems able to perceive human and environmental variables, make intelligent decisions, and give feedback.1 Video games are a virtualized and enhanced reflection of reality. Focused on providing virtual experiences to human players, gaming systems are emerging as the earliest adopters of cognitive capability. Furthermore, games are the simplest platforms for demonstrating intelligent technologies. In fact, public interest in artificial intelligence (AI) was raised by a Go game between Google’s 1520-9202/17/$33.00 © 2017 IEEE

AlphaGo program and Lee Sedol, a South Korean professional Go player ranked as a 9 dan. AlphaGo’s unexpected win made it famous worldwide. The program builds neural networks that guide the optimal solution search procedure, which is a learning approach through dataset training rather than predetermined strategies specified by the designer. Recently, Master, the upgraded version of AlphaGo, won 60 straight games by defeating a litany of world champions, which further proved AI’s advantage. Here, we investigate the design and development of cognitive gaming, survey the current application of cognitive computing in the game industry, and discuss its future.


computer.org/ITPro

55

COGNITIVE COMPUTING Cognitive engine Decision making Content generation

Data mining Contextual perception

Environmental data Input Player inputs

Cognitive Game Content Generation

Providing intelligent responses to players’ commands is the fundaVideo System mental requirement of a gaming optimization … Vision mixture system. Under this circumstance, computer-based AI is a frequently Vibration Output Resource integration mentioned principle in tradiAudio tional game design, even though most implementations are simple scripts that let a nonplayer characFigure 1. The architectural framework of cognitive gaming. This ter (NPC) provide predefined feedarchitecture has three fundamental elements: input, a cognitive engine, back to players’ input. Recently, and output. advanced AI algorithms have started to provide sophisticated responses. For instance, AlphaGo represents Cognitive Gaming Definition the most cutting-edge decision making in chess The term “cognitive gaming” carries different placement in a Go game. meanings in diverse contexts. For example, it can In this article, we extend cognitive computing be used to refer to games that test players’ brains to the broader aspect of procedural generation,3 or challenge their cognitive abilities. It is even adopted as the name of a professional e-sports in which the gaming system can intelligently genteam based in the US. In this article, we define erate dynamic game content.4 In this approach, cognitive gaming as the integration of the cognithe game engine predefines only the rules and tive computing paradigm1 in video games. Figure elements that generate future game content, which will be subject to players’ behaviors and 1 depicts the architectural framework of cognienvironments. For example, the simplest and eartive gaming, which consists of three fundamental liest approach is based on pseudorandom numelements: ber generation, which is used by Elite, a space exploration game, to create a very large universe. • input—the interface that perceives detailed enIt falls in the category of game-space generation. vironmental data and players’ information in Another example is rhythm-based level generareal time and delivers this data to a cognitive tion, which generates different levels for 2D platanalytical engine as quickly as possible; formers (such as Super Mario Bros.) based on the • cognitive engine—the analytical component that pattern of a player’s hand movements when playconsists of data mining, decision making, and ing the game.3 Furthermore, variations in scenes, resource integration components; it cooperatively leverages rich resources in the cloud weapons, characters, obstacles, and other design and applies reasoning, inducting, and learning features could be generated using a genetic algomethodologies to support cognitive capacity; rithm and verified by a game-playing agent.4 The and purpose is to give gamers a more immersive expe• output—the interface that provides various reprience when they are engaged in a gaming session. resentations for players, including the video On top of traditional approaches, cognitive display, vision mixture, vibration, and audio. gaming leverages richer sources of content and explores environmental data from a deeper graAs a software system that provides real-time dation in real time. These approaches will be interaction with players, the system design applied in all perspectives, including to augshould be based on low-latency principles.2 The mented reality (AR), sensors, and players’ behavior. For instance, a cognitive version of Super application of cognitive gaming can be classified Mario Bros. might mix the Mushroom Kingdom into two categories: cognitive game content genwith players’ vision, so that Mario might jump eration and cognitive game system optimization. from one edge to another surface in reality. We discuss these two directions in the following Instead of preset bricks, coins, mushrooms, sections. 56

Player behavior identification


and tortoises, these items might show up according to players’ surrounding environmental information, which can be deducted from multiple sources. An example could be the traffic information returned from the Bing Maps API. A cognitive game system could create more enemies when players are in crowded traffic. Also, Mario’s speed might also be subjected to the movement of a player’s terminal. Given the playing scenario of a fast-moving train, the accelerometer in the gaming device might trigger a high-speed mode for Mario that enables some hidden warp zones. Furthermore, players’ social network status, clicks, and swipes during gaming sessions and the rhythm of Mario’s jumps could also be potential data for content generation. The game scene might be brighter if a player is identified as being in a negative mood, whereas his or her fast pace might result in a burst of enemy attacks.

Augmented Reality An AR game can be considered an advanced version of VR.5 The design principle for VR games is to track players’ eyes and render corresponding game images, thus making the depth of field more realistic. In contrast, AR games capture videos from players’ field of vision (or simply use mobile devices’ back cameras to record videos), analyze the videos’ content, and then dynamically generate virtual content to be displayed to players’ eyes. A well-designed AR system can mix virtual items with reality, leading players to see an indistinguishable vision of their world. An AR game is a typical application of cognitive computing. Industry giants have been investigating the possibility of AR games. For example, the woorld Android game on Google Project Tango lets players decorate their space with fantastical objects and then see how they interact in silly, playful ways (www.funomena.com/woorld). Similarly, the RoboRaid game on Microsoft HoloLense gives players a mixed reality, first-person shooting experience, in which target enemies are coming at the player from every possible direction in his or her physical environment (bit.ly/2q8gqre). Certainly, real-time video pattern recognition is the most challenging issue to be addressed in AR games. State-of-the-art computer vision techniques still have immature image analysis accu-

racy, not to mention high costs and long delays in data processing. Research projects, such as Google Cloud Vision and Microsoft Cognitive Services, are looking for breakthroughs in this area with help from machine learning technologies.

Sensors Another cognitive gaming approach leverages all kinds of available sensors to harvest environmental data from around game players and use it as the input for content generation.6 A classic case is to locate players with GPS and adopt the

A well-designed AR system can mix virtual items with reality, leading players to see an indistinguishable vision of their world. location information as an important parameter for gaming content. For example, Pokemon Go (www.pokemongo.com), a recently popular mobile game, produces different pokemons (pocket monsters) such as magikarp, a special carp fish, according to players’ locations. A magikarp will be available only when the gamer approaches a waterfront, such as a lakeside or harbor side. A similar concept was implemented in an earlier Android game, Ingress (www.ingress.com). Besides GPS, the trend in cognitive gaming is to extend the application of information usage to all kinds of sensors in mobile devices, including gyroscopes and accelerometers. In fact, these sensors have been adopted as input equipment for a long time— for example, in the Wii game console produced by Nintendo. However, in cognitive gaming, these sensory data should target content generation. For example, when a player is traveling in a vehicle, the acceleration his or her mobile device senses can be used as a parameter for game scenario creation.

Player Behavior Collecting players’ information for game content generation will certainly become the trend in cognitive gaming. This is an extension of the

computer.org/ITPro

57

COGNITIVE COMPUTING

personalization approach for targeted advertising, which tracks consumers’ behavior and delivers appropriate and accurate advertisements to them. The Internet’s rich information allows a machine to quickly understand game players’

If players connect their accounts to their social networks, the gaming system can collect and analyze their personal profiles and social feeds.

characteristics. Game systems can analyze gamers’ behavior models, such as selection of game genres, length of gaming sessions, strategy during the sessions,7 and even the key strike frequency and mouse movement traces, to deduce a particular gamer’s characteristics and preferences, then generate content to meet his or her needs. If players connect their accounts to their social networks, the gaming system can even collect and analyze their personal profiles and social feeds to give them a unique, personalized gaming experience. The game rating system can also be implemented in this paradigm: teen players might receive different content than adults, such as that with restrictions on violence, blood and gore, sexual content, or strong language. Furthermore, cognitive computing is a potential solution for difficulty settings, the most critical challenge in game design. An excellent game design should post appropriately difficult tasks, setting solvable but challenging targets to engage players step by step. Existing practice relies on experienced game designers, which isn’t suitable for different levels of players. In the future, cognitive computing will take over this task and provide fine-tuning services for game difficulties, thus enabling customized gaming experiences for different players with different skill levels.

Opportunities and Issues Building cognitive services specifically for game content is never easy because it involves complicated mathematic models and requires large 58


amounts of data as training sets for machine learning algorithms. Leading enterprises running large machine learning clusters are way ahead in this new playing field. For instance, IBM Japan has announced Sword Art Online (ibm. co/2rEu6Ly), which will utilize the IBM Watson cloud to enable cognitive game design. Does this mean that small game companies, especially startups, have no chance in this evolution? Not really. IT franchises would like developers to be attached to their platforms and create rich content for their marketplace so that they can continuously receive high revenue from infrastructure services. Hence, they share their core services through customized APIs. This opens up opportunities for startups to leverage these rich resources with a pay-as-you-go model. From the developers’ perspective, integrating cognitive computing services from public cloud providers is a shortcut. Figure 2 illustrates how a game development leverages several services from the Google cloud platform and the IBM Bluemix cloud. For instance, the player’s terminal can capture video and voices, along with traditional input from players’ keyboard and mouse actions. In this case, the environmental videos can be analyzed by pattern recognition algorithms, such as the Google Cloud Vision service, to extract context information. For voice, the cloud might predict players’ emotions through a tone analyzer, while the verbal content can be converted to text-based messages for further semantic analysis, which will be performed with natural language processing algorithms. Considering that more and more games are operated across multiple countries, cloud translation services for text information are also important elements in future game development. After the cognitive content generator gathers such information, dynamic content generation can be conducted with the help of AI algorithms and external resources, including historic gaming data, related multimedia content from the Internet, players’ social network updates, and so on. As Figure 2 illustrates, cognitive services from the public cloud, such as IBM Watson Conversation, can also be imported as external resources. An illustrative scenario will be creating an NPC to imitate a real person and chat with players in game scenes. As the output of the

Server

Response • text • graphic • audio • video

Conversation response

Conversation request

The IBM Watson conversation

Command

IBM language translator Tone information

IBM text-to-speech services Voice response

Translated text messages

Google natural language API

Context information Google cloud vision

Text

International service

Inter-avatar information

Cognitive content generator

Mouse input IBM tone analyzer

Captured video

Text command

Text message

IBM speech to text Text message Voice message

Client

Figure 2. A cognitive gaming system integrated with public cloud services. Raw inputs (video and voice) are processed by cloud services before reaching the cognitive content generator.

cognitive content generator, the response would be in the form of scene renderings, game element behaviors, and varied dynamic design elements. As depicted in Figure 2, text responses can be transformed into voice responses with help from the IBM text-to-speech service. However, adopting public cloud services also introduces uncertainty to the gaming system because the quality of service (QoS) from these clouds is out of developers’ control; in particular, today’s service level agreements (SLAs) are still coarse-grained and hard to audit. Hence, how to guarantee QoS in accessing these cloud services is still a critical challenge. Moreover, how to customize existing cloud services for specific gaming needs is an open issue. For instance, regarding the chatting NPC example described in Figure 2, it is still difficult to create different NPCs with different story backgrounds in different narratives, because it is challenging to create separate artificial neural networks for different NPCs, which must be fed with distinct knowledge data and profiles.

Cognitive Game System Optimization Optimization is a never-ending task for quality-ofexperience (QoE)-critical systems such as video games. By optimizing the scenario rendering and

computational workload distribution, the cognitive gaming system is expected to deliver a better QoE with the same hardware foundations. As a dynamic human-computer interaction system, different players’ distinct behaviors and performances in different game scenarios pose a stringent requirement in game system design and optimization. These issues become even more complicated in multiplayer games because the variety of environmental (for example, network or terminal) parameters brings in more uncertainty. The self-adaption feature of cognitive computing makes it a powerful methodology to address these issues.

Scenario Rendering Modern video games pursue realistic 3D game scenarios and avatar rendering, imposing a high demand on computational resources. Different from other applications, scenario rendering for video games is highly latency-sensitive. The experience of most game players will suffer from any response delay higher than 200 ms,8 while VR games have an even tighter restriction on the delay. Although the computation performance of PCs and mobile devices has improved greatly in recent years, it is no match for highly

computer.org/ITPro

59

COGNITIVE COMPUTING

sophisticated graphic algorithms that demand abundant computational power, especially in mobile devices with limited computational resources. This becomes a critical issue in game development and deployment. Consequently, many advanced games provide video options for gamers, enabling them to customize the quality of scenario rendering according to their hardware capacity. However, these manual settings with a large number of video options confuse some players, whereas others do not even realize that such functions exist. Furthermore, online and cross-platform games also limit the usage of these settings due to the variety in network and terminal hardware. To this end, cognitive computing can be adopted to optimize scenario rendering. A leading research work on this topic is a comprehensive study of cloud gaming that hosts game engines in cloud servers, renders game scenarios remotely in the cloud, and streams the gaming video to players through the Internet.9 Under this circumstance, network quality becomes the key element that determines system performance, given that the bottleneck is the real-time video transmission. Therefore, the early-stage work focuses on adapting video to network parameters—for instance, one scheme monitors network QoS parameters and adapts scenario rendering parameters such as scenario depth, scenario details, and video quality to improve the perceived gaming experience.10 Another idea perceives avatars’ interaction with game scenarios and determines different importance levels of objects to be rendered in the scene.11 Under weak network or hardware conditions, the proposed system would render only critical object sets in game scenes to minimize the workload for scenario rendering. Both of these works can be easily extended to conventional video games. More interestingly, players’ behavior can also be included as a source of parameters in determining optimal rendering solutions. For example, foveated rendering tracks players’ eyes and selectively renders the important areas around their gaze.12 By reducing the rendering quality of vision edges, this approach significantly reduces the target frame rate and increases resolution compared to traditional real-time renderers. 60


Workload Distribution The cognitive computing paradigm can also be applied to dynamic workload distribution in gaming systems. Here, workload refers to all kinds of computational needs in game programs. Along with the development of game content and forms, especially those massive multiplayer online role-playing games (MMORPGs), the downloading and rendering of game elements are no longer one-time operations on players’ terminals. Instead, progressive downloading and computational offloading become the trend. In progressive downloading implemented by Utomik (utomik.com), the terminal could start a gaming session with a small proportion of the game resources downloaded from the cloud server; the rest of the game then downloads while the game progresses. Progressive downloading requires a cognitive approach to predict a player’s progress and determine the data to be delivered in advance. It potentially reduces downstream transmissions, given that there are many branches in narratives. On the other hand, computational offloading leverages the rich computational resources in the cloud to execute game logic on cloud servers, and then transmits the results to the terminal via the network. In this context, cognitive computing means intelligent decisions in dynamic offloading. CloneCloud, a pioneering study on this topic, enables the elastic execution of general applications between mobile devices and the cloud.13 It proposes a framework that combines static program analysis with dynamic program profiling to balance workload and optimize execution time or energy. The future game system should extend the modality of CloneCloud to perceive and predict all environmental parameters involved, and determine players’ gaming progress and avatar behavior to perform cognitive offloading for system performance optimization.

Opportunities and Issues Cognitive scenario rendering and workload distribution should be built on the concept of component-based gaming architecture, given that scenario rendering needs to separate different objects for rendering, and workload distribution requires the workload itself to be migrated from a terminal to the server or other terminals. An obvious challenge for such an architecture is the decomposition complexity or, to be more specific,

the decomposition level (for example, data level, task level, or function level). The decomposition level defines the frequency at which components interact with each other and thus the rate of data exchange between them. Because components could be remotely executed, a high data exchange rate (high decomposition level) between remote components could be detrimental to both system performance and communication cost. As the decomposition level varies with game genre, how to find the appropriate level of decomposition remains the biggest challenge. As Figure 3 depicts, recent work at the University of British Columbia has designed and implemented a testbed for decomposed gaming, following the cognitive computing paradigm.14 In such a system, video games are decomposed into software components that are cognitive to players’ behavior and execution environments in both the terminal and the cloud. The authors developed several game prototypes over the proposed testbed for cognitive workload distribution. Preliminary experimental results demonstrated that well-designed cognitive algorithms can help improve the performance of a gaming system— for instance, by increasing the frame-per-second value in game execution.

T

he game program is a perfect playground for cognitive computing. Besides the conventional application of using cognitive science as the basis of NPC intelligence, we envision that game content will be partly or entirely generated by a cognitive engine in the near future, involving functions such as AR integration, sensory data usage, and identification of players’ behaviors. Also, the adaptive nature of cognitive computing makes it an attractive solution for video games. Cognitive scenario rendering and workload distribution will provide significant optimization for those systems suffering from a varying execution environment.

References 1. D.S. Modha et al., “Cognitive Computing,” Comm. ACM, vol. 54, Aug. 2011, pp. 62–71. 2. M. Claypool and K. Claypool, “Latency and Player Actions in Online Games,” Comm. ACM, vol. 49, no. 11, 2006, pp. 40–45.

Figure 3. A testbed for decomposed cognitive gaming developed at the University of British Columbia. The left screen displays the workload distribution of different components for the tank game displayed on the right screen. 3. M. Hendrikx et al., “Procedural Content Generation for Games: A Survey,” ACM Trans. Multimedia Computing, Communications, and Applications, vol. 9, no. 1, 2013, article no. 1. 4. L.T. Pereira et al., “Learning to Speed Up Evolutionary Content Generation in Physics-Based Puzzle Games,” Proc. 28th IEEE Int’l Conf. Tools with Artificial Intelligence, 2016, pp. 901–907. 5. A. Phillips, “Games in AR: Types and Technologies,” Proc. IEEE Int’l Symp. Mixed and Augmented Reality— Arts, Media, and Humanities, 2009, pp. 9–10. 6. R. Shea et al., “Location-Based Augmented Reality with Pervasive Smartphone Sensors: Inside and beyond Pokemon Go!,” IEEE Access, to appear, 2017; doi:10.1109/ACCESS.2017.2696953. 7. W. Cai et al., “Quality of Experience Optimization for Cloud Gaming System with Ad-Hoc Cloudlet Assistance,” IEEE Trans. Circuits and Systems for Video Technology, vol. 25, no. 12, 2015, pp. 1–14. 8. M. Jarschel et al., “Gaming in the Clouds: QoE and the Users Perspective,” Mathematical and Computer Modelling, vol. 57, no. 11, 2013, pp. 2883–2894. 9. W. Cai et al., “The Future of Cloud Gaming,” Proc. IEEE, vol. 104, no. 4, 2016, pp. 687–691. 10. S. Wang and S. Dey, “Rendering Adaptation to Address Communication and Computation Constraints in Cloud Mobile Gaming,” Proc. IEEE Global Telecommunications Conf., 2010; doi:10.1109/GLOCOM.2010.5684144. 11. I.S. Mohammadi, M.R. Hashemi, and M. Ghan bari, “An Object-Based Framework for Cloud Gaming Using Player’s Visual Attention,” Proc. IEEE Int’l Conf. Multimedia & Expo Workshops, 2015; doi:10.1109/ ICMEW.2015.7169781. 12. A. Patney et al., “Towards Foveated Rendering for Gaze-Tracked Virtual Reality,” J. ACM Trans. Graphics, vol. 35, no. 6, 2016, article no. 179.

computer.org/ITPro

61

COGNITIVE COMPUTING

13. B.G. Chun et al., “CloneCloud: Elastic Execution between Mobile Device and Cloud,” Proc. 6th Conf. Computer Systems (EuroSys), 2011, pp. 301–314. 14. W. Cai et al., “MCG Test-Bed: An Experimental TestBed for Mobile Cloud Gaming,” Proc. 2nd Workshop Mobile Gaming (MobiGames), 2015, pp. 25–30.

Wei Cai is a postdoctoral research fellow in the Department of Electrical and Computer Engineering at the University of British Columbia (UBC), Canada. His research areas include cloud computing, cognitive systems, Internet of Things, finance technology, and software engineering. Cai received a PhD in electrical and computer engineering from UBC. He is a member of IEEE. Contact him at [email protected]. Yuanfang Chi is an application engineer at Oracle. Her research areas include cognitive computing, cloud computing, and real-time software architecture. Chi received a master’s of applied science in electrical and computer

Read all your IEEE magazines and journals your WAY on

engineering from the University of British Columbia, Canada. She is a student member of IEEE. Contact her at [email protected]. Victor C.M. Leung is a professor of electrical and computer engineering and holder of the TELUS Mobility Research Chair at the University of British Columbia (UBC), Canada. His research interests are in the broad areas of wireless networks and mobile systems. Leung received a PhD in electrical engineering from UBC. He is a Fellow of IEEE, the Royal Society of Canada, the Canadian Academy of Engineering, and the Engineering Institute of Canada. Contact him at [email protected].



Now there’s even more to love about your membership...

Introducing myCS, the digital magazine portal from IEEE Computer Society. Go beyond static, hard-to-read PDFs with an easily accessible, customizable, and adaptive experience.

There’s No Additional Cost!

►LEARN MORE AT: mycs.computer.org IEEE myCS half Page Space Ad 2016_4-26-16.indd 1

62


4/28/16 3:03 PM

DATA ANALYTICS EDITOR: Seth Earley, Earley Information Science, [email protected]

The Problem With AI Seth Earley, Earley Information Science

S

cience is actually pretty messy. When I was a chemistry undergraduate, I loved the theory behind biochemistry— the endless complexity allowed by simple rules; how massive, complex cellular machines could arise from a few building blocks. In lab, however, I struggled to make the simplest reactions work. Starting with pure crystalline compounds and expensive laboratory equipment, when the result was also expected to be crystalline, I ended up with piles of brown goo—with my instructor concluding, “Well, it could be in there” in reference to the experiment’s objective. Data science is also very messy. Frequently the starting point is the data equivalent of brown goo—messy, poor quality, inconsistent data—with the expectation that pure crystalline results will be the output of the next best action, personalized marketing campaigns, highly effective custom email campaigns, or a cross-department, cross-functional, 1520-9202/17/$33.00 © 2017 IEEE

360-degree understanding of customers and their needs. Artificial intelligence (AI), though broadly applied these days to mean almost any algorithm, is primarily dependent on some form of machine learning. Machine learning in turn is frequently fueled by what is called big data (high-velocity, highvolume, highly variable data sources) but can also be fueled by traditional data sources.

Variable Does Not Mean Poor Quality There is a common misconception that “variable” data can mean “messy” data and that “messy” data can mean “poor-quality” data. Simply put, variable does not mean messy, and messy does not mean poor quality. Variable data is data that has different formats and structures. To use it, we need to understand how the different types of data can be used as signals to achieve a result. Twitter data is very different than transactional data. The two together can provide insights about how social


trends impact sales. Messy data can be missing values or can be in formats that are difficult to ingest and process. The data can be very good, but requires work to get it into a format for processing. A recent article in Sloan Management Review stated that Organizations can now load all of the data and let the data itself point the direction and tell the story. Unnecessary or redundant data can be culled … [This process is] often referred to … as ‘load and go.’1

While conceptually accurate, there is much left open to misinterpretation. “All the data” needs to be defined. Does it mean all product data, social media data, accounting data, transactional data, knowledge base data? Clearly “all” is an overgeneralization. And this approach has its drawbacks. Sandy Pentland, MIT professor, remarked at the recent MIT CIO Symposium that “Putting all of your data in a data lake makes it convenient for hackers to go to one place to steal all of your data.”

computer.org/ITPro

63

DATA ANALYTICS No matter what the scope is, we have to select data that is appropriate to the domain of the problem space. The data needs to be in a consistent format. It cannot contain incorrect values. If the data is incorrect or missing, then the algorithm cannot function correctly unless we are making accommodations for those issues. “It’s an absolute myth that you can send an algorithm over raw data and have insights pop up,” according to Jeffrey Heer, a professor of computer science at the University of Washington, as quoted in the New York Times.2 Technology writer Rick Delgado notes that “many data scientists jokingly refer to themselves as data janitors, with a lot of time spent getting rid of the bad data so that they can finally get around to utilizing the good data. After all, bad data can alter results, leading to incorrect and inaccurate insights.”3 In a recent conversation I had with Laks Srinivasan, chief operating officer (COO) of Opera Solutions, he asserted that “80 percent of the work the data scientists are doing is data cleaning, linking, and organizing, which is an information architecture (IA) task, not a data scientist function.” Opera, founded in 2004, was one of the firms that tied for first prize in a Netflix contest that was offering US$1 million to the company that could beat its recommendation engine by 10 percent or more. (The three-year contest, which ended in August 2009, awarded the prize to a team from AT&T Labs, which submitted its response just minutes before Opera.) Opera is an example of a company that developed a platform to help data scientists in many aspects of analysis, feature engineering, modeling, data preparation, and algorithm operationalization. 64


A Range of AI Applications AI applications exist along a spectrum. At one end lies embedded AI, which is transparent to the user, but makes applications work better and easier for them. Spelling correction is an example that people take for granted. Machine translation is another. Search engines use machine learning, and AI, and, of course, speech recognition, which has made enormous progress in recent years. At the other end of the spectrum are the applications that require deep data science and algorithm development expertise. The people who develop these applications are technical experts with deep mathematical and data science knowledge. They devise and tune the algorithms that provide advanced functionality. Along the continuum are the platforms and development environments that make use of the tools (many of which are open source). These applications require various levels of configuration and integration to provide capabilities.

Types of Cognitive Computing For example, consider a type of “cognitive computing” application. Cognitive computing is a class of application that helps humans interface with computers in a more streamlined, natural way. Such applications are also capable of processing information in a less traditionally structured manner to provide a range of answers, with probabilities based on the user’s context and details about the data sources. One type of cognitive computing application is the processing of large amounts of patient observational data and providing a “second opinion” about a diagnosis. Physicians are using this approach

to augment their knowledge and experience when developing treatment regimens. Another type is creation of an intelligent virtual assistant (IVA) that retrieves answers to procedural questions rather than lists of documents. IVA functionality requires various mechanisms that are powered by machine learning. The first is speech recognition, which translates spoken language into text. The next is a mechanism for deriving intent from the user query or utterance. Intent can be based on training sets and examples of phrase variations, or it can be from parsing language to derive meaning.

The Role of Machine Learning Each of these approaches leverages machine learning. Some dialog management approaches can use mechanisms akin to language translation. Given enough questions and enough answers, a machine learning algorithm can “translate” questions into the correct responses. When the intent is derived via natural language understanding or training set classification, a response can be retrieved from a corpus of content via a ranking algorithm that uses signals generated through determining the intent of the user as well as additional metadata that can inform the user’s context—anything from purchased products, to configured applications, to demographic or social media data. Inference can use relationships mapped in an ontology— for example, products associated with a particular solution or steps to troubleshoot a specific device configuration. Some of this knowledge is inferred from the data and some is intentionally structured—the knowledge engineering approach to AI.

Contextualizing Endless Knowledge Sources Organizations have enormous repositories of knowledge in the form of processes, procedures, manufacturing techniques, research methodologies, embedded designs, programming code, configured applications, technical documentation, knowledge bases of various kinds, engineering libraries, expert systems, traditional libraries, technical publications, scientific, engineering, and trade journals—the list of explicit knowledge sources is endless. Historically, humans have always limited the scope of the information that they consume—for example, by picking up a book on a topic, searching for a specific area in a library, pursuing a specialized library, or seeking out a particular journal. Even in our digital age, engineers will go to engineering sites for nuanced, specialized information. Scientists will go to scientific sites, and so on. Information from highly diverse sources cannot be processed as raw data inputs for any purpose without restriction. It needs to be parsed, curated, packaged, contextualized, and componentized for consumption by users or ingested by systems for application to a limited number of scenarios. As powerful as it was, the Jeopardy-playing Watson program required specific information sources to function correctly.

Can Curation Be Automated? Machines can help when given the correct scaffolding and representative training sets. Data and content sources can be processed by machine algorithms, overlaying the structure and identifying patterns in the information to assist in componentization and contextualization. The process is iterative and

requires human judgment and inputs to fine-tune results. Those results might be the componentized information containing specific answers to questions rather than large amounts of text. When the content is fine-tuned and componentized, the specific answers can be more readily retrieved. A user looking for an answer does not want a list of documents, but the answer to the question. Bots and intelligent virtual assistants are designed to respond with an answer or a short list of suggestions presented in the correct context (the user’s query or intent). Autotagging and autoclassification machine learning algorithms can apply the correct metadata to content to allow for those contextualized results.

The Role of Ontologies Ontologies are the containers of metadata—the knowledge scaffolds or structures that can be abstracted from systems of knowledge and applied to other bodies of information for organization and contextualization. The ontology can capture the relationships between knowledge elements and ways of organizing those elements—for example, the list of user intents with corresponding actions. A taxonomy of products can be related to a taxonomy of solutions composed of those products. Or a list of problem types can be associated with corresponding troubleshooting approaches. Tools such as virtual assistants become channels for knowledge structured with an ontology, along with rules and contexts that apply to specific problem sets. Take, for example, the task of servicing a customer who is trying to set up and operate a new fitness tracker. Instead of searching on the website or calling the help desk, the customer might try

typing a question into the company’s support chat bot. The bot interprets the natural language question as an intent, and the ontology allows retrieval of the correct responses from a knowledge repository. The ontology manages intents and responses as well as terminology and phrase variations for algorithm training. The advantage of a natural language question over a search is that it becomes easier to derive the user’s intent when they ask a fully formed question rather than typing a few ambiguous keywords. A bot can also be programmed to further disambiguate intent by requesting more detail from the user. This type of natural interface can also be used to access corporate information sources—running a financial analysis or retrieving underwriting procedures, for example.

Maturing Algorithms Still Necessitate Data Clean-Up While machine algorithms play an important role in both the preparation of data and interpretation of user intent, these types of applications require a significant amount of knowledge engineering to be successful. As machine learning algorithms mature, the heavy lifting will become more invisible and behind the scenes, and data or content preparation as well as application tuning and configuration will constitute the bulk of the work and require the greatest effort. With data scientists increasingly in short supply, business users will need to perform more analysis so that a backlog does not develop behind scarce data science resources. Data preparation is a major challenge, and operationalizing capabilities is an even bigger one. This is because knowledge of deep analysis approaches is becoming lost in translation from the laboratory

computer.org/ITPro

65

DATA ANALYTICS environment to the operational environment. Given that detailed machine learning approaches are less accessible to business people, there is increasingly a gulf between the business world and the IT world. However, two trends are in play. Sophisticated tools are becoming more commoditized, while more advanced capabilities are being made available to business people through platform approaches. The key component of data preparation, data operationalization, and translation between business challenges and analytical tools is the semantic layer—the glossaries, thesaurus structures, metadata standards, data architectures, and quality mechanisms. As the tools get more mature, organizations will get value from them only if they take control of the things that will not be commoditized by the marketplace— their data, content, processes, and semantic translation layers. For example, organizations will not get a competitive advantage by building speech recognition. That problem has been solved (for the most part— it is still improving, but building the algorithms from scratch would not have business value). They will, however, gain a competitive advantage from servicing their customers uniquely with a speech recognition agent that accesses the knowledge they have about their customers and serves up the products and content they need.

Rethinking High-Power Analytics As demand is exploding for big data analytics, data scientists are increasingly in short supply. When a company is building predictive models or machine learning models, a few factors stand out. Every journey starts out with raw data, so if a company is doing multiple projects for the same 66


client and the same department, multiple teams start with the same raw data, which can be inefficient. The second factor is that so much of the work data scientists are doing is data cleaning, linking, and organizing, which, as Srinivasan mentioned, is an IA task, not a data scientist function. The third factor is that even after the data is cleaned up and models are developed that accurately predict (for example) who is likely to buy a certain product, it takes a lot of time to go from the data science sandbox to actually operationalizing the analytics that create a business impact. This disconnect occurs because the development environment and the production environment are very different. As Srinivasan explains, “The data scientists might build a model using SAS in the sandbox and using certain data sets, but the IT department needs to re-code the variables and models in Java or optimize R code to scale in Hadoop when the application goes into production. At this point, the data is also very different because it goes beyond the test datasets, so the data scientists have to retest it against the model. Finally, even when the projects are in production, all these insights and know-how [are] fragmented into documents or code or people’s heads. As staff turns over, knowledge is lost.”

Re-Imagining the Analytics Lifecycle When Opera began considering how to address these issues, it came up with the approach of fundamentally re-imagining the analytic development lifecycle by developing a “semantic layer” between the data layer (raw data) and the use case, application, and UI layer. The thought was that the company could preprocess

the data to a point, independent of its future use, and then apply AI and machine learning in converting big data to small data. By putting a semantic layer around analytic models and tools, all the users can find them once the semantic layer is operationalized. According to Srinivasan, “By making data independent of use cases and operationalizing it, and then making it machine-learningdriven and AI-driven, the signals learn about the data. The system becomes a learning system, not a static, one-time data modeling system. It becomes a continuous feedback, loop-based, living, breathing kind of a central nervous system, in the enterprise.” In other words, the semantic layer acts as a way to translate business problems into the inputs needed to query a big dataset. The technical predictive algorithms operate under the covers, and this complexity is hidden from the user. The algorithms simply have to point to the big data sources (that are correctly cleansed and prepared, of course) and then provide their parameters as inputs to predict their outcomes, run simulations, segment audiences, customize campaigns, and so on.

Developing an Orchestration Layer In the case of Opera, the company went on to build a platform from the ground up to create and manage the signal layer, and ran mission-critical applications on it. The platform, called Signal Hub, processes data from about 500 million consumers for global blue-chip clients across industries. This approach allowed Opera to essentially outsource the data science work, operate on its platform, and sell solutions to business buyers. When Opera developed and then productized

the platform as an orchestration layer in 2013, many organizations did not have the IT or data science resources to fully exploit the power of advanced tools. The market has matured since then, and that strategy—to productize as an endto-end AI and machine learning enterprise platform by hardening with security, scalability, and governance capabilities—provides valuable lessons for organizations building data-driven solutions. Thinking about data as a service and the platform as an orchestration layer between business problems and technology solutions can help organizations achieve dramatic improvement in data scientists’ productivity, and in the productivity of business analysts and business intelligence workers. “The maturing of technologies and emergence of platforms is democratizing insights derived through machine learning and capabilities provided by AI in a way that we say makes ordinary people extraordinary,” says Srinivasan. “If all the insights and expertise are buried in a small team within a company, it doesn’t really leverage the value of AI tools to be used by an average call center rep.” The concepts of data as a service and platforms as an orchestration layer have far-reaching implications for the future of AIdriven enterprises. Not only can data be more fully exploited by this paradigm, but so can knowledge and content—the raw material on which cognitive applications are being developed. According to Henry Truong, CTO of TeleTech, a $1.4 billion call center services firm, “Organizations can normalize knowledge in the same way that they normalize data—through componentizing knowledge into the building blocks that provide solutions to problems. The knowledge ontology becomes the data

source to orchestrate more and more process actions, that, in our case, prevents service disruptions.” This approach is beginning to be exploited in ways that allow for interoperability between platforms that are exposing functionality through a services layer. Those “normalized knowledge bases” are powering chat bots that are driving the next-generation digital worker.

Leveraging Platforms and Orchestration Layers Many organizations are attempting to build their own platforms and believe this is required to create a competitive advantage from machine learning and AI capabilities. The key decision point is whether the platform is the differentiator or whether it is the data and orchestration layer that will be the differentiator. “I frequently hear CIOs say they have a platform or that they are building machine learning. The problem is that it is easy to go through $100 million or more, and a lot of pain and suffering. I say, ‘Do not try this at home’ in my presentations and hope they take it to heart,” cautions Srinivasan. A core premise for success with advanced analytics is that organizations need to build metadata structures and ontologies to define relationships among data elements relevant to their companies. Srinivasan continues: “That is the investment that organizations should be making rather than building their own platforms. They should be building their own representation of the core of the business, the soul of the business, which is the ontology that can embody all that knowledge of processes and customers. Insights can then be fed back into the ontology, so it becomes that living, breathing thing. It is a semantic layer that evolves around that.”

M

ost of the work that data scientists do is “data janitorial” work, as opposed to science work, and there is a gulf between prototype and sandbox, and innovation and production. In addition, having pockets of knowledge and expertise throughout the enterprise, which may be gone when an employee leaves, poses a problem when the knowledge is not institutionalized or captured in a system. Organizations are best off if they focus on understanding their own data, focus on the business problems they are trying to solve, and build the semantic layers that can allow for data portability across various platforms. This lets them take advantage of bestof-breed solutions and not become locked into a particular vendor that does not abstract the business problem, analytic, data, and platform layers required to operationalize the fast-evolving advanced machine learning analytic and AI technologies.

References 1. R. Bean, “How Big Data Is Empowering AI and Machine Learning at Scale,” MIT Sloan Management Rev., 8 May 2017; bit.ly/2psZyMm. 2. S. Lohr, “For Big-Data Scientists, ‘Janitor Work’ Is Key Hurdle to Insights,” New York Times, 17 Aug. 2014; nyti.ms/2kl1V3Y. 3. R. Delgado, “Why Your Data Scientist Isn’t Being More Inventive,” Dataconomy, 15 Mar. 2016; bit.ly/ 2rxoy73.

Seth Earley is CEO of Earley Information Science (www.earley.com). He’s an expert in knowledge processes, enterprise data architecture, and customer experience management strategies. His interests include customer experience analytics, knowledge management, structured and unstructured data systems and strategy, and machine learning. Contact him at [email protected].

computer.org/ITPro

67

SECURING IT EDITORS: Rick Kuhn, NIST, [email protected] Tim Weil, Scram Systems, [email protected]

Can Blockchain Strengthen the Internet of Things? Nir Kshetri, University of North Carolina at Greensboro

B

lockchain—a kind of distributed ledger technology—has been described in the popular press as the next big thing. Put simply, a blockchain is a data structure that makes it possible to create a tamper-proof digital ledger of transactions and share them. This technology uses public-key cryptography to sign transactions among parties. The transactions are then stored on a distributed ledger. The ledger consists of cryptographically linked blocks of transactions, which form a blockchain (bit.ly/2sgabnq). It is impossible or extremely difficult to change or remove blocks of data that are recorded on the blockchain ledger. Regarding the question of whe ther blockchain can strengthen the Internet of Things (IoT), the answer—based on this research—is “maybe.” Observers have noted that the blockchain– IoT combination is powerful and is set to transform many industries.1 For instance, IoT devices can carry out autonomous 68


transactions through smart contracts.2 Combined with artificial intelligence (AI) and big data sol utions, more significant impacts can be produced. A natural question is thus what roles can blockchain play in strengthening IoT security? To demonstrate this problem’s significance, consider the following example. In October 2016, the US-based DNS provider Dyn faced cyberattacks. Dyn said the attacks originated from “tens of millions of IP addresses,”3 and at least some of the traffic came from IoT devices, including webcams, baby monitors, home routers, and digital video recorders.4 These IoT devices had been infected with malware called Mirai, which controls online devices and uses them to launch distributed denial-ofservice (DDoS) attacks. The process involves phishing emails to infect a computer or home network. Then the malware spreads to other devices, such as DVRs, printers, routers, and Internet-connected cameras employed by stores and businesses for surveillance.5


From a security standpoint, a main drawback of IoT applications and platforms is their reliance on a centralized cloud. A decentralized, blockchain-based approach would overcome many of the problems associated with the centralized cloud approach. Some point out that blockchain could provide military-grade sec urity for IoT devices.6 There is no single point of failure or vulnerability in blockchain, except with the clock needed for time stamping. Considering these observations, this column provides insights into ways in which blockchain might strengthen IoT security.

Incorporating Blockchain into IoT Security Blockchain’s incorporation into IoT is being supported through a wide variety of measures intended to strengthen security. Several companies are leading initiatives to integrate blockchain into their production and supply chains. For instance, IBM is using its large cloud infrastructure to provide

1520-9202/17/$33.00 © 2017 IEEE

blockchain services for tracking high-value items as they move across supply chains. The IBM Watson IoT Platform’s built-in capability also allows users to add selected IoT data to private blockchain ledgers that can be included in shared transactions. The platform translates the data from connected devices into the format that blockchain contract APIs need. It is not necessary for the blockchain contract to know the specifics of the device data. The platform filters device events and sends only the data that is required to satisfy the contract (ibm.co/2rJWCPC). All business partners can access and supply IoT data in a decentralized fashion and can verify each transaction.7 Data is not collected, stored, or managed centrally. Rather, it is protected and shared among only the parties involved in the transaction. Startups such as Provenance use blockchain to promote trust in the supply chain by providing transparency and visibility when the product moves from the source to the customer.8 Others are creating new business models that eliminate the need for centralized cloud servers. For example, Filament, a blockchain-based solutions provider for IoT, has launched wireless sensors, called Taps, that allow communication with computers, phones, or tablets within 10 miles (bit.ly/2rsxZYf). Taps create low-power, autonomous mesh networks that enable companies to manage physical mining operations or water flows over agricultural fields. Taps don’t rely on cloud services. Device identification and intercommunication is secured by a blockchain that holds the unique identity of each participating node.9 One key application is likely to be in the next generation of the industrial

network (the Industrial Internet). Filament’s blockchain-based applications involve sensors connected in a decentralized system and use autonomous smart contracts. This means that devices communicate securely with each other, exchange values, and execute actions automatically. For instance, Filament’s Tap can be attached to drilling rigs in remote locations. Based on predefined conditions, a rig might know that it requires a piece of machinery and thus might send a request to an autonomous drone.10 Measures are also taken at interorganizational levels. A group

can be achieved.13 In this regard, a key challenge that arises in some applications is that it is difficult to ensure that the properties of physical assets, individuals (credentials), resource use (energy and bandwidth through IoT devices), and other relevant events are stored securely and reliably. This aspect can be handled relatively easily for most IoT devices. For instance, a private blockchain can be used to store cryptographic hashes of individual device firmware. Such a system creates a permanent record of device configuration and state. This record can be used to verify that a given

Blockchain-based identity and access management systems can be leveraged to strengthen IoT security. of technology and financial companies have announced that they have formed a group to set a new standard for securing IoT applications using blockchain. Companies joining the group include Cisco, Bosch, Bank of New York Mellon, Foxconn Technology, Gemalto, and blockchain startups Consensus Systems, BitSE, and Chronicled.11 This group hopes to establish a blockchain protocol to build IoT devices, applications, and networks.12

Identity and Access Management Systems Blockchain-based identity and access management systems can be leveraged to strengthen IoT security. Such systems have already been used to securely store information about goods’ provenance, identity, credentials, and digital rights. As long as the original information entered is accurate, blockchain’s immutability

device is genuine and that its software and settings have not been tampered with or breached. Only then is the device allowed to connect to other devices or services. Returning to the Dyn example, IP spoofing attacks were launched for the later versions of the Mirai botnet. Blockchain-based identity and access management systems can provide stronger defense against attacks involving IP spoofing or IP address forgery. Because it is not possible to alter approved blockchains, it is not possible for devices to connect to a network by disguising themselves by injecting fake signatures into the record.14 The earlier example involving Filament’s Taps illustrates this point.

Cloud vs. Blockchain Models In the cloud model, IoT devices are identified, authenticated, and connected through cloud servers,

computer.org/ITPro

69

SECURING IT Table 1. How blockchain can address Internet of Things (IoT) challenges. Challenge

Explanation

Potential blockchain solution

Costs and capacity constraints

It is a challenge to handle exponential growth in IoT devices: by 2020, a network capacity at least 1,000 times the level of 2016 will be needed.

No need for a centralized entity: devices can communicate securely, exchange value with each other, and execute actions automatically through smart contracts.

Deficient architecture

Each block of IoT architecture acts as a bottleneck or point of failure and disrupts the entire network; vulnerability to distributed denial-of-service attacks, hacking, data theft, and remote hijacking also exists.

Secure messaging between devices: the validity of a device’s identity is verified, and transactions are signed and verified cryptographically to ensure that only a message’s originator could have sent it.

Cloud server downtime and unavailability of services

Cloud servers are sometimes down due to cyberattacks, software bugs, power, cooling, or other problems.

No single point of failure: records are on many computers and devices that hold identical information.

Susceptibility to manipulation

Information is likely to be manipulated and put to inappropriate uses.

Decentralized access and immutability: malicious actions can be detected and prevented. Devices are interlocked: if one device’s blockchain updates are breached, the system rejects it.

where processing and storage are often carried out. Even if devices are a few feet apart, connections between them go through the Internet.15 First, IoT networks that have high costs are a concern in the centralized cloud model. Gartner estimated that in 2016, 5.5 million new IoT devices were connected every day.16 It is estimated that by 2020, a network capacity that is at least 1,000 times the level of 2016 will be needed.17 The amount of communication that needs to be handled will increase costs exponentially. Second, even if economic and manufacturing challenges are addressed, each block of the IoT architecture could act as a bottleneck or point of failure that can disrupt the entire network.18 For instance, IoT devices are vulnerable to DDoS attacks, hacking, data theft, and remote hijacking. Criminals might also hack the system and misuse data. If an IoT device connected to a server is breached, everyone connected to the server could be affected. 70


Consider smart water meters and associated risks. Twenty percent of California’s residents have smart water meters, which collect data and send alerts on water leakage and usage to consumers’ phones. Likewise, the Washington Suburban Sanitary Commission (WSSC) in Washington, DC, is planning to integrate IoT into its system. Water-usage data can tell criminals when residents are not home. Perpetrators can then burglarize homes when their residents are away.19 Third, the centralized cloud model is susceptible to manipulation. Collecting real-time data does not ensure that the information is put to good and appropriate use. Consider the water supply system example just discussed. If state officials or water service companies believe that the evidence might result in high costs or lawsuits, they can censor, edit, or delete data and analysis. They can also manipulate findings. For instance, consider the water crisis in the city of Flint, Michigan, which began in 2014. Flint authorities insisted for months that city water

was safe to drink.19 Citing official documents and findings of researchers who conducted extensive tests, a CNN article asserted that Michigan officials might have altered sample data to lower the city’s water lead level.20 It was reported that the Michigan Department of Environmental Quality and the city of Flint discarded two of the collected samples. A researcher said that the discarded samples had high lead levels. Including them in the ana lysis would have increased the level above 15 parts per billion (PPB). According to the US Environmental Protection Agency, water supply companies are required to alert the public and take action if lead concentrations exceed the “action level” of 15 PPB in drinking water (bit.ly/1qKMLVE). Blockchain can eliminate many of the drawbacks described in Table 1. In blockchain, message exchanges between devices can be treated in a similar way as financial transactions in a bitcoin network. To exchange messages, devices rely on smart contracts. Blockchain cryptographically signs

Downstream supply chain partners/ device owners

Upstream supply chain partners • Tracing back products to the origin of the raw material • Pinpointing the source of problematic parts/ items

Device manufacturers/network providers

• Identifying users of vulnerable devices • Guaranteeing return of products in case of recalls • Registering updates, patches, and part replacements throughout the lifetime of a product

Figure 1. Blockchain’s role in improving overall security in supply chain networks. With blockchain, it is possible to access immutable records for various aspects of transactions involving a product to understand key vulnerabilities in the upstream supply chain. This technology can also help strengthen downstream supply chain partners’ and device owners’ precautionary and defensive cybersecurity measures.

transactions and verifies those cryptographic signatures to ensure that only the message’s originator could have sent it. This can eliminate the possibility of man-in-themiddle, replay, and other attacks.6 Blockchain’s proponents have forcefully argued that this new technology can save us from “another Flint-like contamination crisis.”19 Projects such as the WSSC’s integration of the IoT in supply systems can be upgraded with sensors such as near-infrared reflectance spectroscopy (NIRS) to include data on chemical levels. If such a system had been installed in Michigan, Flint’s water service company could have found the lead contamination when it exceeded healthy levels. Blockchain can provide the “second layer of crisis prevention” in such cases.20

Ensuring Supply Chain Security Blockchain can ensure supply chain security (see Figure 1). It also makes it possible to contain an IoT security breach in a targeted way after discovery of the breach. Blockchain can facilitate

handling and dealing with crisis situations such as product recalls due to security vulnerabilities. Blockchain’s public availability means that it is possible to trace back every product to the origin of the raw materials, and transactions can be linked to identify users of vulnerable IoT devices. IoT-linked security crises, such as the cyberattacks on Dyn, could have been handled better if the supply chains had adopted blockchain. For instance, China-based Hangzhou Xiongmai Technologies, which makes Internet-connected cameras and accessories, recalled its products in the US that were vulnerable to the Mirai malware. However, it is difficult to determine the devices’ owners. Blockchain is suitable for complex workflows. It can be used to register time, location, price, parties involved, and other relevant information when an item changes ownership. The technology can also track raw materials as they move through the supply chain, are transformed into circuit boards and electronic components, are integrated into

products, and are sold to customers. Blockchain can also be used to register updates, patches, and part replacements applied to any product or device throughout its lifetime. It is easier to track progress in addressing vulnerabilities and send warnings and notifications to owners.8

B

ased on the evolving mechanisms and forces described here, a promising future seems likely for the use of blockchain in addressing IoT security. For instance, some of the key security challenges associated with the cloud can be addressed by using the decentralized, autonomous, and trusted capabilities of blockchain. Blockchain’s decentralized and consensus-driven struc tures are likely to provide more secure approaches as the network size increases exponentially. Blockchain enables the verification of the attributes it carries. Blockchain-based transactions are easily auditable. Due primarily to this and other features, blockchain

computer.org/ITPro

71

SECURING IT can play a key role in tracking the sources of insecurity in supply chains as well as in handling and dealing with crisis situations such as product recalls that occur after safety and security vulnerabilities are found. And as mentioned, blockchain-based identity and access management systems can address key IoT security challenges such as those associated with IP spoofing.

Acknowledgments I thank Jeff Voas for numerous edits and suggestions on previous versions of this article. Certain commercial entities, equipment, or materials may be identified in this document in order to describe an experimental procedure or concept adequately. Such identification is not intended to imply recommendation or endorsement, nor is it intended to imply that the entities, materials, or equipment are necessarily the best available for the purpose.

References 1. K. Christidis and M. Devetsikiotis, “Blockchains and Smart Contracts for the Internet of Things,” IEEE Access, May 2016, pp. 2292–2303. 2. Blockchain in Banking: A Measured Approach, Cognizant Reports, 2016. 3. “3rd Cyberattack ‘Has Been Resolved’ After Hours of Major Outages: Company,” NBC New York, 21 Oct. 2016; bit.ly/2eYZO46. 4. N. Perlroth, “Hackers Used New Weapons to Disrupt Major Websites Across US,” New York Times, 21 Oct. 2016; nyti.ms/2eqxHtG. 5. E. Blumenthal and E. Weise, “Hacked Home Devices Caused Massive Internet Outage,” USA Today, 21 Oct. 2016; usat.ly/2eB5RZA. 6. J. Coward, “Meet the Visionary Who Brought Blockchain to the Industrial IoT,” IOT World News, 14 Dec. 2016; bit.ly/2s8la1w. 7. A. Kaul, “IBM Watson IoT and Its Integration with Blockchain,” Tractica, 1 Aug. 2016; bit.ly/2rsOp2M.

72


8. B. Dickson, “Blockchain Could Help Fix IoT Security after DDoS Attack,” VentureBeat, 29 Oct. 2016; bit.ly/2dXNaNO. 9. B. Dickson, “How Blockchain Can Change the Future of IoT,” VentureBeat, 20 Nov. 2016; bit.ly/2qXZWXw. 10. S. Pajot-Phipps, “Energizing the Blockchain—A Canadian Perspective,” Bitcoin Magazine, 26 Jan. 2017; bit.ly/2r7IIEc. 11. J. Brown, “Companies Forge Cooperative to Explore Blockchain-Based IoT Security,” CioDive, 30 Jan. 2017; bit.ly/2quIMfv. 12. E. Young, “Tech Giants and Blockchain Startups Unite to Make IoT Apps More Secure,” The Cointelegraph, 30 Jan. 2017; bit.ly/2kNtm7w. 13. C. Catallini, “How Blockchain Applications Will Move Beyond Finance,” Harvard Business Rev., 2 Mar. 2017; bit.ly/2m2ZIZQ. 14. S. Kumar, “Not Just for Crypto cash: How Blockchain Tech Could Help Secure IoT,” IoT Agenda, 13 Feb. 2017; bit.ly/2m8H9Gr. 15. A. Banafa, “IoT and Blockchain Convergence: Benefits and

Challenges,” IEEE Internet of Things newsletter, Jan. 2017; bit.ly/2n1y8jq. 16. R. Van der Meulen, “Gartner Says 6.4 Billion Connected ‘Things’ Will Be in Use in 2016, Up 30 Percent From 2015,” Gartner press release, 10 Nov. 2015; www.gartner.com /newsroom/id/3165317. 17. S. Waterman, “Industry to Government: Hands Off IoT Security,” Fedscoop, 17 Nov. 2016; bit .ly/2g4oXYX. 18. A. Banafa, “A Secure Model of IoT with Blockchain,” OpenMind, 21 Dec. 2016; bit.ly/2j2QUkH. 19. R. Hackett, “How Blockchains Could Save Us from Another Flint-Like Contamination Crisis,” Venturebeat, 25 Feb. 2017; bit.ly/2mx11zp. 20. D. Debucquoy-Dodley, “Did Michigan Officials Hide the Truth about Lead in Flint?” CNN, 14 Jan. 2016; cnn.it/2r0aiF9.

Nir Kshetri is a professor of management in the Bryan School of Business and Economics at the University of North Carolina at Greensboro. Contact him at [email protected].

IT Professional (ISSN 1520-9202) is published bimonthly by the IEEE Computer Society. IEEE Headquarters, Three Park Avenue, 17th Floor, New York, NY 10016-5997; IEEE Computer Society Publications Office, 10662 Los Vaqueros Circle, PO Box 3014, Los Alamitos, CA 90720-1314; voice +714 821 8380; fax +714 821 4010; IEEE Computer Society Headquarters, 1828 L St. NW, Suite 1202, Washington, DC 20036. Visit www .computer.org/subscribe for subscription information. Postmaster: Send undelivered copies and address changes to IT Professional, Membership Processing Dept., IEEE Service Center, 445 Hoes Lane, Piscataway, NJ 08854-4141. Periodicals Postage Paid at New York, NY, and at additional mailing offices. Canadian GST #125634188. Canada Post Publications Mail Agreement Number 40013885. Return undeliverable Canadian addresses to PO Box 122, Niagara Falls, ON L2E 6S8, Canada. Printed in the USA. Editorial: Unless otherwise stated, bylined articles, as well as product and service descriptions, reflect the author’s or firm’s opinion. Inclusion in IT Professional does not necessarily constitute endorsement by the IEEE or the Computer Society. All submissions are subject to editing for style, clarity, and space.

Technology Solutions for the Enterprise Editor in Chief

Editorial Staff

San Murugesan, BRITE Professional Services, [email protected]

Editorial Management: Rebecca Deuel-Gallegos

Associate Editors in Chief

Irena Bojanova, NIST, [email protected] J. Morris Chang, University of South Florida, [email protected] Robert R. Harmon, Portland State University Linda Wilbanks, US Department of Education, [email protected]

Editorial Product Lead: Bonnie Wylie, [email protected]

Department Editors

Publications Coordinator: [email protected]

Data Analytics: Seth Earley, Earley Information Science, [email protected] IT in Emerging Markets: Gustavo Rossi, Universidad Nacional de La Plata, [email protected] IT and Future Employment: George Strawn, US Nat’l Academies of Sciences, Engineering, and Medicine, [email protected] IT Trends: Irena Bojanova, NIST, [email protected] Life in the C-Suite: Joseph Williams, CloudEconomist.com, [email protected] Mastermind: George Strawn, [email protected] Securing IT: Rick Kuhn, NIST, [email protected] Tim Weil, Scram Systems, [email protected]

Director, Products & Services: Evan Butterfield

Editorial Board

CS Magazine Operations Committee

Cover Designer: Jennie Zhu-Mai

Senior Manager, Editorial Services: Robin Baldwin Manager, Editorial Services: Brian Brannon Director of Membership: Eric Berkowitz Senior Advertising Coordinator: Debbie Sims

2017 IEEE Computer Society President Jean-Luc Gaudiot

Saeid Abolfazli, YTL Communication, Xchanging Malaysia Stephen J. Andriole, Villanova University Wesley Chou, US Department of Defense Fulvio Corno, Politecnico di Torino Claudio Giovanni Demartini, Politecnico di Torino Haluk Demirkan, University of Washington–Tacoma Vladimir Dimitrov, University of Sofia Reza Djavanshir, Johns Hopkins University Jinan Fiaidhi, Lakehead University, Canada Simon Fong, University of Macau G.R. Gangadharan, Institute for Development & Research in Banking Technology Vladimir Getov, University of Westminster Bin Guo, Northwestern Polytechnical University George F. Hurlburt, STEMCorp Samee U. Khan, North Dakota State University Maria R. Lee, Shih Chien University Sunil Mithas, University of Maryland Arijit Mukherjee, Innovation Labs Kolkata, Tata Consultancy Services Arpan Pal, Tata Consultancy Services Charalampos Z. Patrikakis, Piraeus University of Applied Sciences Rajiv Ranjan, Newcastle University Hiroyuki Sato, University of Tokyo Jilei Tian, BMW Technology Chicago

George K. Thiruvathukal (chair), Gul Agha, M. Brian Blake, Jim X. Chen, Maria Ebling, Lieven Eeckhout, Miguel Encarnação, Nathan Ensmenger, Sumi Helal, San Murugesan, Yong Rui, Ahmad-Reza Sadeghi, Diomidis Spinellis, VS Subrahmanian, Mazin Yousif

CS Publications Board Greg Byrd (VP for Publications), Alfredo Benso, Irena Bojanova, Robert Dupuis, David S. Ebert, Davide Falessi, Vladimir Getov, José Martínez, Forrest Shull, George K. Thiruvathukal Writers: Access www.computer.org/itpro/author.htm. Letters to the Editors: Send letters to [email protected]. Subscribe: Visit www.computer.org/subscribe. Subscription Change of Address: Contact [email protected]. Missing or Damaged Copies: Contact [email protected] Reprints of Articles: Contact [email protected]. IEEE prohibits discrimination, harassment and bullying: For more information, visit www.ieee.org/web/aboutus/whatis/policies/p9-26.html.

Advisory Board

Jin-Fu Chang, President, Yuan-Ze University and Board Chair, Institute for Information Industry, Taiwan Wushow Chou (EIC Emeritus), North Carolina State University Stephen L. Diamond, EMC Corporation Karen Evans, KE&T Partners Thomas Jepsen, IT consultant Simon Liu, US Nat’l Agricultural Library Sorel Reisman (chair), California State University, Fullerton Henry Schaffer, North Carolina State University George Strawn Published by the

R euse R ights

R eprint Permissions: Educational

1) is not made for profit; 2) includes this IEEE endorsement of any third - party products or services. Authors and their companies are permitted to post the accepted version of IEEE- copyrighted material on their own Web servers without permission, provided that the IEEE copyright notice and a full citation to the original work appear on the first screen of the posted copy. A n accepted manuscript is a version which has been revised by the author to incorporate review suggestions, but not the published version with copy- editing, proofreading, and formatting added by IEEE. F or more information, please go to: http :// www. ieee .org/publications_standards/publications/rights/paperversionpolicy. html . Permission to reprint/republish this material for commercial , advertising, or promotional purposes or for creating new collective works for resale or redistribution must be obtained from IEEE by writing to the IEEE Intellectual Property R ights O ffice, 445 H oes L ane, Piscataway, NJ 08854-4141 or pubs - permissions@ ieee.org. Copyright © 2017 IEEE. A ll rights reserved. and

or personal use of this material is permitted without fee , provided such use:

notice and a full citation to the original work on the first page of the copy; and

3)

does not imply

A bstracting and Library Use: A bstracting is permitted with credit to the source. L ibraries are permitted to photocopy for private use of patrons, provided the per- copy fee indicated in the code at the bottom of the first page is paid through the Copyright Clearance Center , 222 Rosewood D rive , Danvers, MA 01923.

computer.org/ITPro

3

SkillChoice™ Complete Now with expanded libraries and an upgraded platform!

3,000+

MENTORSHIP

online courses

u Valued at

$3,300!

28,000+ books

6,000+

Practice Exams

videos

15,000+ Books24x7 titles

OVER 20x as many resources as before

One membership. Unlimited knowledge. Did you know IEEE Computer Society membership comes with access to a high-quality, interactive suite of professional development resources, available 24/7? Powered by Skillsoft, the SkillChoice™ Complete library contains more than $3,000 worth of industry-leading online courses, books, videos, mentoring tools and exam prep. Best of all, you get it for the one low price of your Preferred Plus, Training & Development, or Student membership package. There’s something for everyone, from beginners to advanced IT professionals to business leaders and managers. The IT industry is constantly evolving. Don’t be left behind. Join the IEEE Computer Society today, and gain access to the tools you need to stay on top of the latest trends and standards.

Learn more at www.computer.org/join.

ACCESS TO SKILLSOFT IS AVAILABLE WITH

cognitive computing

cognitive computing

Suggest Documents

Cognitive Informatics & Cognitive Computing

Cognitive Informatics & Cognitive Computing

cognitive computing

Cognitive Informatics & Cognitive Computing

Cognitive Computing

Cognitive Informatics, Cognitive Computing, and Their ...

Cognitive Computing | August 2011 ...

Symbiotic Cognitive Computing - Tom Erickson

Cognitive Informatics and Cognitive Computing in Year 10 and Beyond

[PDF] Towards Cognitive Cities: Advances in Cognitive Computing ...

[PDF] Cloud and Cognitive Computing - Google Sites

Cognitive Computing Architectures for Machine (Deep ... - MDPI

Blockchain Technology, Cognitive Computing, and Healthcare ... - JAIT

Computing Narratives of Cognitive User ... - Semantic Scholar

Extended Sparse Distributed Memory - Cognitive Computing ...

Cognitive Computing Programming Paradigm: A Corelet Language ...

On Cognitive Computing - University of Calgary

Soft Computing Technique of Fuzzy Cognitive

Applying Cognitive Computing for Meaningful Customer ... - Wipro

Semantic, Cognitive, and Perceptual Computing

Cognitive Computing Resource Management for a Ubiquitous ...

On Cognitive Computing - University of Calgary

Guided Cancer Care Through Cognitive Computing

Blockchain Technology, Cognitive Computing, and ...