Intelligent Interactive System for Collaborative Green Computing Kuo-Ming Chao, Nazaraf Shah, Adriana Matei, Tomasz Zlamaniec, Weidong Li Dept. of CDE, Coventry University Coventry, United Kingdom (csx240, aa0699, mateia, zlamanit, weidong.li)@coventry.ac.uk
Abstract— Designing an effective and efficient intelligent interactive system for energy conservation in domestic environment requires considerable deliberations on tradeoffs of meeting the criteria derived from human-machine interaction and reasoning mechanism, so the system can provide the users with right information based on right decisions at right time. In this paper, we report a mixed initiative intelligent system that employs multi-modal sensor system, context awareness model, semantics, and service-oriented architecture to provide real-time energy consumption information and recommendations for positive behaviour change on households’ energy consumption. The system adopts concept of On-Line Analytical Process (OLAP) to summarise large volume of data generated from energy consumption monitoring sensors into a set of meaningful information for the intelligent system to reason. ServiceOriented Architecture (SOA) is used to model hardware and software components, so they can be composed flexibly to meet the user and system requests. The system is currently deployed in around 250 households across Europe for evaluation. Initial households’ feedbacks are encouraging. Keywords-Recommending;Semantic Computing;Sensors;Green Computing
I.
Web;
Service-Oriented
INTRODUCTION
Green computing can refer to environmentally sustainable computing or IT, or to make the use of computers for improving energy efficiency. In this paper, we focus on utilizing ICT to make households aware of their energy consumption and to influence their behaviors towards improving the efficiency in their energy usage. Domestic electrical appliances such as washing-machine, tumble-dryer, fridge, cooker etc are essential equipments to maintain our daily life comfort, but they consume substantial portion of overall household energy. Energy efficiency can be improved or more energy can be saved, if their usage can be monitored in real-time and be optimized. However, most households are not aware of how much energy each appliance has consumed, or information about its energy consumption profile. Furthermore, they do not know the ways and means of improving their usage to save energy. So, it is important to make them aware of where and how much of the energy is being used in real-time. Consequently, recommendations on energy conservation can be presented to them, so they can act as actuator to control appliances in conjunction with automatic controller. It would be ideal that each appliance has an
Chi-Chun Lo, Yinsheng Li Institute of Information Management, National Chiao-Tung University, Hsin-Chu City, 300, Taiwan Software School, Fudan University Shanghai, China
[email protected];
[email protected] attached sensor which collects its energy consumption data for efficiency analysis, so the recommendations can be made accordingly, even though it would be expensive in terms of cost. Designing an intelligent collaborative system for energy conservation faces a number of challenges, as it needs to satisfy a number of requirements and constraints derived from the characteristics of application domain such as large amount of data, multiple users and opinions with multiple criteria, real-time responses and intelligent interactions. It requires an approach that allows the system to be flexible, dynamic and efficient. In other words, it needs a mechanism to coordinate the involved subsystems such as sensor network, data analysis tool, actuators, and contextual information system, recommending system and intelligent interactive system. However, it includes the following aspects of the system that need to be investigated. A. Intelligent System and HCI The design spaces of intelligent system and human computer interaction study are large and complex and their views can be polarised, as one attempts to maximise the automation process for decision making on behaviour of human users and the other tries to facilitate them to have better control over or improved interaction with the system. The intersection of these two spaces has been less explored by research communities due to the distinct interests of both camps [1], so to find a generic optimised solution is not a trivial task. With the advance of the multimodal sensor system, contextual modelling and related technologies, the search process in the space can be navigated and refined. B. Generalisation and Personalisation A system often is designed to serve a group of users, but each individual user has his/her preferences or arbitrary requirements need to be met. The system design has to consider both individual and group requirements, so the balance of generalisation and personalisation needs can be identified. This requirement complicates the design of an effective and intelligent interactive system. The system needs to be generic for all households, but it also needs to consider personalised information and profile to be displayed at right time.
C. Large Volume of Data The introduction of multi-modal sensor system to the intelligent system can increase the accuracy of pattern analysis and predication. Large volume of data generated from sensors cannot be efficiently processed for the real-time recommending system. In addition, statistical or rule-based reasoning over a large knowledge base often is a time and resource consuming process. These constraints bring the difficulty to satisfy the efficiency required by real-time applications. D. Proactive and Responsive The system needs to be proactive, but it also needs to be reactive when is necessary. The system needs to provide the appropriate information to the user at right time and in right context. However, incomplete and uncertain information often exist in the environment. This could result from unobtainable information or missing data. It could not be effective and efficient to solely rely on the system (AI system) to make predications as they could be wrong in presence of uncertain information. So, modelling the user profile and contextual information effectively becomes an important task for increasing the accuracy. The users can contribute their inputs by filling in missing information or data in order to reduce uncertainty. E. Feedbacks The system needs to capture feedbacks from users and the reasons for accepting or disregarding a recommendation, so the system can take them into account for re-organising them accordingly. However, the system will suggest recommendations those with good ratings regardless other factors such as relevancy and context, if it takes into account single selection criteria of highest ratings. Other useful recommendations may not emerge, as the search is trapped in a confined space. A mechanism is needed to explore other potential and useful recommendations. F. Heterogeneous Systems Different technologies and programming languages have been used to realise systems and devices. This leads to a barrier to interact with them coherently. In addition, the web becomes the most accessible medium. So, it is rational to enable them to collaborate via web technologies which can also become a common platform for them to communicate. II.
BACKGROUND STUDY
With the recent improvements and fast developments in AI algorithms and technologies, more AI techniques are embedded in the interfaces to assist the users. This can be seen as a new generation of HCI. The integration between AI and HCI can be categorized into two types: Intelligent HCI and Adaptive HCI [2]. The criterion to distinguish them is based on the role of AI played in HCI. Intelligent HCI focuses on making the interface with intelligent features such as Microsoft agent providing an interface to allow the users using natural language to interact with the Microsoft Office suite. The design of the adaptive HCI has great interest in making
the system interaction with users more intelligent by using AI, but not only in interfacing. It can be used in the way to continue to interact with users according to their preferences or behaviors [1]. For example, Microsoft Word can remember words inputted by the users in the documents. When the users misspell the words, the system automatically corrects them. In the context of Dehems, the system is expected to proactively interact with the users when abnormal power usage in the appliances has been detected. It can interact with the users to get relevant information to identify the possible causes of abnormal power consumption. This leads to the study of mixed initiative interaction between computer system and human users, so they can take turn to initiate the dialogs in order to achieve their common goal or come to a joint understanding. The study of mixed initiative interaction has generated great interests since late 90 from both AI and HCI communities. Its research focuses on developing methods that enable computing systems to support an efficient, natural interleaving of contributions by people and computers, aimed at converging on solutions to problems [1]. However, difficulties of integrating automated processes with direct human input still exist. Horvitz summarized a list of principles of mixed-initiative user interface to address these challenges in 1999[3,4], but with advances in sensing and reasoning over human cognition, these challenges could be addressed partially for some specific application domains. For example, the combination of user profile, contextual information and multimodal sensor systems can remove some uncertainty in users’ perceived intentions. Multimodal human computer interaction system is simply one that responds to inputs in more than one modality or communication channel. The modality means mode of communication according to human senses or computer input and output devices. Even though they can bring advantages, filtering and coordinating large volume of data becomes an issue. The recommender systems have been widely used in ECommerce for recommending books, DVD, and holiday packages etc [5]. Its sophisticated preference-based filtering and similarity calculation approaches with ratings/voting have increased its accuracy in predication. According to Adomavicius and Tuzhilin [5], the underling modeling approaches to derive recommendations can be classified into three categories: content-based methods, collaborative-based methods, and hybrid approaches. Content-based methods focus on measuring user’s preferences on the contents of commonalities to derive a related recommendation based on a pre-defined arbitrary utility function. Since the search spaces can be very large and complex, the user profiles are often taken into account in the search. The collaborative recommendation systems are built upon a collective and relevant preferences and rating given by other users to derive the utility function while the content-based filtering approach is more concerned with individual user’s past preferences. The collaborative recommending system could be based on the similarity calculation of the entire collection of previously rated items by users or use statistical
models to reason over the past ratings to produce the predication. The hybrid approach tries to take advantage of the aforementioned approaches to address their shortcomings and exhibits their positive characteristics in order to produce more accurate recommendations. Various hybrid systems have been proposed and they proved their approaches outperforming each pure approach via empirical tests. However, there still exist a number of factors that can be included to increase performance such as contextual information and multiple criteria ratings which are not well-addressed by the research communities. Search dimensions of the system can be expanded to include environmental or user contextual information, for example, weather, location, and type of houses etc to increase predication accuracy on energy saving patterns. The introduction of multimodal interactions also can assist the accuracy of the prediction. Various database and data warehousing technologies are available to store and manage large volumes of data. They provided different structures for storing and accessing data. They normally support various functions at the system level to ensure that data consistency and system performance are maintained with multi-user accesses and in some cases with disperse locations. However, to choose an appropriate data storage technology depends on the nature and requirement of applications such as the size of data, available resources, data characteristics and time constraints. For example, the recordoriented (relational) storage systems places attributes (tuple) contiguously in storage as rows, so it has high performance in writing data to the physical disk. A column store architecture differs from the row store architecture, as it stores values for each single column (or attribute) contiguously. It responds better to the retrieval of a bulk of data from repository according to query specifications [6]. Even though, these technologies can increase performance, the challenges of requiring real-time responses from a large volume of data are still at large. The data needs to be preprocessed and pre-analyzed and it has to be stored appropriately as summarized and aggregated information for real-time applications. The structure of information requires flexibility and sufficient capabilities to re-organize/re-aggregate data to meet the needs of analysis and decision making tools at design time and runtime or unforeseen future requirements. OLAP is a well-known solution to this type of problems. Codd states that OLAP is made up of numerous speculative “what-if” and/or “why” data model scenarios executed within the context of some specific historical basis and perspective [7]. It, however, has difficulty to locate the required information and compose them efficiently. This requires another layer that enables to annotate these data and schemas and exhibit them, so the required data can be located and synthesized on the fly. Introduction of semantics to the data cubes that can be reasoned by inference engines could help to dynamically synthesize the data to answer high level conceptual questions. A number of middleware technologies (e.g. RPC, DCOM, RMI, CORBA and Web services etc) have been developed for
integrating heterogeneous distributed systems since 1980s. The middleware has evolved from procedural code approach to high-level service-oriented architecture, so the application developers can focus on design of reusable components to support the required functionalities regardless the barriers built by technology differences. The main characteristic of SOA is its strength in dynamic service composition at runtime, so an application can be formed by distributed composite services at run time. It adopts and defines standardized communication protocols, descriptions for service interface, facilities for service discovery, and XML language to facilitate system integration and collaboration. So, hardware and software including control mechanisms and data can be modeled as services to work together in a unifying manner to achieve their design goals. From briefly revisiting the enabling technologies and analyzing their pros and cons, to design an effective and efficient intelligent interactive system for energy conservation requires a new architectural approach that allows the required services or systems to operate in a collaborative manner. The system needs to effectively and efficiently process large amount of raw data to respond users’ queries. The system also needs to consider the users’ personal requirements/preferences and contextual information in order to provide personalised services. The recommendations about energy saving given to the users need to be useful by considering their own subjective preferences and group’s opinions. The system should be able to capture households’ opinions on usefulness of energy saving recommendations by employing multidimensional criteria. Furthermore the system should be able to bridge heterogeneous components into a coherent system. The system needs to observe individual or community level of energy consumption activities and proactively give them appropriate recommendations on optimal use of energy when necessary. The proposed approach is different from the existing ones as it needs to meet the above requirements. III.
THE PROPOSED CONCEPTUAL MODEL
We propose a new intelligent collaborative architecture which is based on AI, SOA, mixed initiative concept, and fuzzy Multiple Criteria Decision Making in combination with the concept of OLAP to address the aforementioned issues and to provide an effective solution to home energy management. The concept of mixed initiative is used to trigger various steps defined in the architecture and it can be divided into two different aspects. These are interactions between human users and machine and interactions between individual user and communities. A. Mixed Initiative Interactions Machines in this context can be computer, home appliances and devices (such as windows, and doors etc) which have an interface that allows users or computer programs to control them manually or automatically. The user uses the machine and the machine then produces the effects such as energy consumption and temperature change etc. The associated sensor collects the data and sends to a software
system in order to increase its visibility and present statistical information to the users in real time. The users are aware of the cause-effect, so behaviour change could occur or lessons can be learned if the effects are not positive. The system can initiate the interactions with the users by alerting or notifying them of predication or giving them useful recommendations based on statistical, environment contextual information and built-in knowledge. The individual user’s energy consumption behaviour could be influenced by the collected behaviours of his/her community, if such information is available. Greater impact could be generated if the users can select what groups of the community they want to get compare with. The data generated by the community with appropriate processing on them can indirectly initiate effective interaction with individual. The community consists of a collection of individuals. So, behaviour change of each individual also produces – a potential impact on the community. This would produce a positive cycle. This can be seen as data and information level interaction because the results are derived from data analysis. The other type of interactions can be regarded as social interaction which is based on their expressed preferences and opinions on the recommendations. Each individual gives the score on the energy saving recommendations. So, the recommender system based on the group ratings considering individual contextual information gives a list of recommendations. The individual can either take the advice given by the group to consume the devices or ignore it. As a consequence, the individual rate these recommendations according to his/her subjective opinions and preferences which could contribute to a new ranking of these recommendations. B. Data, Information, Knowledge, and Societal Levels In order to model different types of interactions, the system needs to have multiple levels. These are data, information, knowledge, and societal (see Fig. 1). The data are raw data sent from sensors. The information includes user profile, summarized data derived from raw data, and context information etc. The knowledge contains energy usage pattern identification, user expectation reasoning, and recommending mechanism etc. The societal is collective data, information and knowledge generated or built by the community.
Fig. 1. Levels in the Conceptual Model
The data mainly associated with sensors can be divided into static data and real-time data. The static data such as historic sensor data, household ID and sensor ID etc will be converted into information for various uses (e.g. average monthly energy consumption for certain type of houses). The real-time data are stored into cache memory with high
performance accessibility which can be processed and displayed to the users in nearly real-time. Information includes static and dynamic information. The information such as household profile, expectation on energy usage, appliance information is relatively stable. Not much change needs to carry out after the information is established. The dynamic information, regarded as data cube, is mainly derived from sensor data which is changing and growing over time. The data cubes can be combined in different ways to support the queries posed by the users. The knowledge base includes ontology for domestic appliance classification, semantics to describe OLAP, and rules and tips for recommending energy saving as well as a reasoning mechanism for ranking recommendations according to group’s opinions. Individual household data, information and knowledge can be pre-processed and organized in a form that can be consumed at societal level. The individual household energy consumption data within the community, for example, can be aggregated to form a neighbourhood energy consumption profile. Data generated by different types of households can be stored into separated data cubes (sets) for further comparison and analysis. Individual’s preference on energy usage recommendations can be collected and reasoned to produce the groups’ preference. Context is taken into account to tag all these contents of levels in order to increase their usability. For example, each recommendation is stored in the ontology with unique ID and interface, so it can be accessible directly. The contents in each level can be directly accessed by the third party software or applications with appropriate authentications or internal processes by following the specifications defined in the interfaces. Each cube has a schema, so its contents can be populated offline. A number of threads or processes are running constantly to monitor new inputs of data and update the contents of the cube. The new cube could be created according to the requirements when the existing cubes cannot satisfy them. C. Service Layers The involvements of various hardware and software technologies create the challenge of providing a collaborative platform to allow them to communicate and interact. So, the SOA was adopted to model the heterogeneous technologies by introducing coherent interfaces and communication protocol with service demand, broker and supply layers. The contents of data, information, knowledge and societal are modelled as services and they can consume other services or provide services depending on the rules of mixed initiatives. In other words, the human users can be regarded as service provider and demander, since they could act as actuators in the domestic appliance control and could issue queries to the system for their required information and recommendations. Service broker plays an important role in service discovery and composition. Service demanders issue the request and the service broker selects the appropriate services and composes them. The selection criteria are based on functional specifications and non-functional requirements with group consensus.
IV.
THE DESIGN OF ARCHITECTURE
The design of the architecture can be viewed as a system construction which consists of a number of subsystems and these are Recommendation Browser, Intelligent Interacting System, Context-awareness Recommendation System, Appliance Ontology, and Group Consensus Feedback System. Each subsystem is enabled by service-oriented technologies partially or entirely and each has explicit interfaces for interactions with others (see Fig. 2). A. Intelligent Interacting System The intelligent interacting system is the key interface facility to link the human users and machines. The intelligent interacting system includes a knowledge base that has capability to interpret the requests from the human users to provide appropriate functions by identifying and composing appropriate services. It interacts with other subsystems (see Fig. 2) to supply correct information, giving timely and useful advices on energy savings to the users. It also links to context aware recommendation system to proactively provide appropriate information. The system has access to the user profile, so any unnecessary information will be filtered out and only relevant interactions will take place. For example, if the user does not have a tumble dryer, its related recommendations will not be shown.
has consumed more energy than the average of its similar type of houses and family, the system starts look up for possible causes and give them related advices. The other dimension to support proactive recommendation is through the analysis on appliance usage. If the usage of a particular appliance in a household is higher than the previous period or similar type of households, the system will be activated to interact with users to identify any possible inappropriate use or malfunction. Household type will be taken in account before altering the users, if energy consumption is higher than other households. It also contains environment variables such as weather and location which can be combined with appliance usages to provide useful recommendations. For example, the system can provide useful tips in relation to drying washings by retrieving information from the Google weather forecast centre when it detects that the washing machine starts. It has a constant monitoring mechanism to observe and analyze energy usages generated from the sensors which are associated with particular appliances. It collaborates with the intelligent interacting system to identify the causes and to generate appropriate advices for users. C. Ontology Repository The Ontology Repository includes three different ontologies (appliance, society and energy usage recommendation) and they are sharable knowledge and are modelled using semantic web technologies. They are built upon SUMO ontology [8] as it standardises a number of universal terms which can be extended to define domain specific terms (see Fig. 3).
Fig. 2. The Design Architecture
It mainly works on summarized information and composes them in order to produce effective information for knowledge base to reason. It also integrates with the recommendation knowledge base and ontologies to provide intelligent recommendations based on their energy usages. It also provides an interface that allows the users to rate the proposed recommendations according to given sub-criteria. B. Context-awareness Recommendation System The system provides a facility that can proactively provide recommendations on energy saving to the users according to the expectation of energy usage from individual household and societal, the condition of their appliance usages, or environmental variables. It includes a context model to record the norms in energy consumptions based on different groups of households and societal. Number of family members living in a house and type of houses, for example, are used as criteria for comparison. If a household
Fig. 3, Domestic Electrical Appliance and Societal Ontology
The appliance ontology classifies appliances according to their functions and defines their energy usage profile. Currently we classify 10 concepts for the top abstract level of electrical domestic appliances which inherit the properties from machine concept in the SUMO ontology. The repository also has societal ontology that defines the household attributes and their members. With combining the information
in the user profile and these standardized definitions, the information at different societal levels such as an individual household, local community, or region, can be created to form a base line for comparison and analysis. The related information can be retrieved and aggregated accordingly. We also encode energy saving advices and their semantics in the recommendation ontology. We classify energy consumption activities related to various electrical appliances in home environment [9] into a hierarchy. The ontological representation of hierarchy provides the semantics to these activities and provides a rich structure for reasoning rules and knowledge sharing. These energy consumption activities are distinct in a way that they are uniquely associated with various appliances. Abnormal or undesired energy consumption may be associated with any energy consumption activity taking place in a home environment. D. Recommendation Knowledge Base The diagnosis process is trigged by an abnormal energy consumption event which in fact deviation from energy consumption profiles of an appliance. The system checks the profiles of the concerned appliance. Based on abnormal values of the appliance the system will gather data acquired by other sensors and user profiles in order to use this data to diagnosis the underlying cause of abnormal event. The data from other sensors such as occupancy and temperature sensors provide contextual information which surround the period during which abdominal event was detected. If all necessary data required by reasoning process is not available within the system, then system will interact with user get information related to abnormal event [8]. Once necessary information are acquired the reasoning process concludes the underlying cause of the abnormal event and generate advice for household which is based on recommendation and energy/money wastage of not dealing with the cause of the abnormal event. For example on detection of abnormal peak in washing machine activity the component triggers the rule base which in turn invokes rule related to abnormalities of washing machine. The rules interact with ontology to get information regarding washing machine profiles and pieces of advice encoded in the ontology. In case underlying cause of abnormality is high washing temperature, the rule will generate the advice concerning appropriate washing temperature and explains the consequences of high temperature setting. E. Group Rating on Recommendations Recommendations about energy saving are made available to the users based on their ratings and a number of other measures (such as contextual information, personal profiles, energy saving knowledge bases etc) to ensure that the appropriate recommendations will be presented. The users give their ratings to the recommendations based on their usefulness, but it is a vague, subjective, and not easy criterion to measure. The sub-criteria (importance, timely, and relevance) are introduced to give the users to express their intentions in details. Fuzzy rating is adopted for the users to express their opinions, so fuzzy reasoning is used to calculate the rating score. The system collects individual ratings over a
period of time and evaluates them to identify group opinions. The evaluation method is based on the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) to help the system to reason over fuzzy ratings given by the users. The linguistic variables, parameterized by triangular fuzzy numbers, are used for evaluating the weights of criteria and the ratings of each recommendation [9]. F. Multiple Modal Sensors and Summarised Information Multiple modal sensors generate multiple streams of data which draws the picture of the monitored environment or object with appropriate aggregation and analysis. These data, however, can grow into unmanageable size which cannot be processed efficiently enough to make real-time responses even with distributed high performance computers. The concept of OLAP is introduced to pre-process and aggregate these data into multiple dimensional data cubes. This could increase the system performance, as the data cubes can be reused. It can reduce the processing time, even though a number of redundant data sets (cubes) will be created. Its advantages exceed the disadvantages. However, as the number of cubes increases, the difficulty of finding the required ones also rises. The criteria for searching the right ones do not only require schema compatibility, but also semantic matching. We proposed an ontological approach to describe the abstract contents of each cube which can automatically be reasoned by computer programs [6]. Each cube is represented as a service, so the content can be accessed without difficulty. These cubes can be composed as another services to be consumed by other applications. G. Recommendation Browser Recommendation browser is a facility allowing the users to traverse the knowledge base to improve their knowledge in energy savings. This facility could compensate the shortfalls in automatic recommendation mechanism which could be trapped in local search area. V.
IMPLEMENTATION
The system adopts web technologies to provide the interfaces for the end users to interact, as it provides good facilities for presentation and accessibility. We used Protégé [10] as a tool to create and model the ontology which eventfully is stored in OWL in order to increase its reusability. Currently we have around 60 concepts relating to domestic electricity appliances having been defined and around 70 tips regarding energy saving having been encoded (see Fig. 4). The inference engine which supports reasoning over knowledge bases is Jess [11]. JessTab [12] was used to bridge to the gap between Jess and ontology. So, the relations between the concepts defined in ontology can be exploited and reasoned by Jess. FuzzyJ was used to model fuzzy concepts and to reason over them in conjunction with Jess. We use an open source system, Synapse, which is an open source based on Apache web server to be our Enterprise Service Bus (ESB). ESB is not only a web server container but also can forward the message to other web services accordingly which are deployed on other
servers. Service Component Architecture is adopted to support service realization. More details can be found in [13]. The database system used for managing the sensor data is Informix which provides efficient database access. Currently we did not use any data warehousing commercial tools to support OLAP, because they do not support semantics. Instead, we designed the cubes as tables stored in relational database which is integrated with Jess and OWL to provide semantics and reasoning for cube selection and composition. The paper in [6] has further descriptions on the ontological OLAP. The sensors we used to monitor and collect the energy usage are Plugwise sensor and electricity clamp meter. The Plugwise sensor monitors individual electricity appliance or a collection of appliances which share the same plug. The clamp meter is attached to the wire of main electricity switch of a household that gathers the total electricity consumption. Both make use of radio transmissions to transfer the data to an Internet gateway which processes and send data to a centralized server for other applications to consume. A similar approach is adopted to monitor household gas consumptions. So, the users can view their energy consumptions in real-time (nearly). We also used these data to feed into the OLAP system. The data will be pushed into permanent storage after it reaches to a threshold in terms of time and size. Since large amount of data generated from different sensors need to be stored in permanent storage, a mechanism is designed to ensure that they are consistent and can be retrieved by the users in real time. So, the information shown in the dashboard with real-time information comes from cache memory. Each household has a unique ID and each sensor in the household also has ID, so the data collected from each sensor (modal) will be sent to a centralized database via a gateway and data collector. Each modal normally is associated with a device or a group of devices.
the ratings, they are stored in the database and will be used for generating/revising group consensus ratings.
Fig. 5. Recommendations and their ratings
Fig. 6 shows alert containing of Google weather forecast in the region where the user is located, when the user starts the washing machine. The system is able to detect that and proactively and automatically invokes the rules to retrieve the recommendations from the ontology about what actions can be taken in order to conserve the energy.
Fig. 6. Contextual Information
Fig. 4. Mapping Concepts in SUMO ontology
Fig. 5 displays the energy saving recommendations to the user after the user logs in to the system. The system currently only selects top three recommendations with the highest scores. The user can rate them according to their usefulness. After rating on the main criterion has been given, the user can give their ratings on the sub-criteria. Once the users finished
Fig. 7 indicates the system interacting with the user in order to identify the possible cause of the energy inefficiency. In this case, the fridge has consumed more than the weekly average, so the system alters the user and asks the questions in order to identify the underlying cause and to give appropriate recommendations. The system invokes the knowledge base when the user gives the answer. This process will continue until all the uncertainties in the knowledge base have been clarified and the right recommendations can be determined.
In terms of evaluation, we have evaluated the performance of our service server against the increase in the number of services. It shows that the service server raises its processing time linearly, as the number of services increases linearly. As the number of feedbacks given by the users increases in two folds, the required processing time to identify their group consensus on the ratings of recommendations also increase linearly. Currently, we are collecting feedbacks from the users about its usage in order to understand its usability. ACKNOWLEDGMENT
Fig. 7. Interacting Intelligent System
So far, we have implemented most required functions for the subsystems. The system is live on www.dehems.org [14]. Even though, the system has built-in functions for the users to remotely control the appliances, they have been disabled due to safety reason. VI.
DISCUSSION AND CONCLUSION
Mixed-initiative concept has been researched and applied to different technologies. Mixed-initiative solutions in combination with sensors and contextual have improved the system performance and reduced the system uncertainty. It has been used in intelligent agents [15] to guide the agent how to interact with human users and other researchers [16] used the concept of mixed imitative interactions and sensors to improve search in dangerous environment. However, they did not address performance in relation to data size. We introduced contextual information and summary data derived from sensors with group ratings to make recommendations, so it can offer more relevant recommendations. In [5], the authors encouraged the researchers to include these aspects for future research. We proposed a multiple dimensional interactions that allow machine, individual or communities to take initiates for interactions. The researchers [1] in this field have less concern with this aspect of the research. In this work, we have integrated the mixed initiative approach with background data processing technique for OLAP to improve efficiency and effectiveness in producing information. In this research, we also analyzed and enabled different types of interactions. The interactions do not only take place between machine and human users, but also between machines, human and communities. The form of their interactions could be direct and indirect. The popularity and affordability of multimodal sensors increase their deployment in monitoring the surrounding environments and human activities. The data generated from them helps to understand more about their interaction patterns. This can contribute to the recommendation mechanism in order to produce more accurate advices.
This research is carried out as a part of DEHEMS which is funded from the European Community’s Seventh Framework Program FP7/2007-2013 under grant agreement No.224609. We acknowledge all DEHEMS members for their useful input to this paper. REFERENCES [1] H. Lieberman, User Interface Goals, AI Opportunities, AIIMagazine, Vol 30, No 4, 2009 [2] F. Karray, M. Alemzadeh, J. A. Saleh and M. N. Arab, HumanComputer Interaction: Overview on State of the Art, International Journal on Smart Intelligent Systems, Vol. 1, No. 1, 2008 [3] E. Horvitz, Principles of Mixed-Initiative User Interfaces, CHI '99 Proceedings of the SIGCHI conference on Human factors in computing systems, pp159-166, 1999 [4] E. Horvitz, Reflections on Challenges and Promises of Mixed-Initiative Interaction, AI Magazine, Vol. 28, No 2, pp 19-22, 2007 [5] G. Adomavicius, A. Tuzhilin, Toward the Next Generation of Recommender Systems: A Survey of the State-of-the-Art and Possible Extensions, IEEE Transactions on Knowledge and Data Engineering Volume 17 Issue 6, pp. 734-749, 2005 [6] N Shah, C Tsai, M Marinov, J Cooper, P Vitliemov, K-M Chao. Ontological On-line Analytical Processing for Integrating Energy Sensor Data. IETE Tech Rev, Vol. 26 pp.375-87, 2009 [7] E.F. Codd, S.B. Codd and C.T. Salley, “Providing OLAP to UserAnalysts: An IT Mandate”, 1993. [8] IEEE SUO working group, SUMO (Suggested Upper Merged Ontology), http://protege.stanford.edu/ontologies/sumoOntology/sumo_ontology.ht ml, 2011 [9] K-M Chao, N. Shah, R. Farmer, A. Matei, D-Y Chen, H. Schuster-James, R. Tedd, "A Profile Based Energy Management System for Domestic Electrical Appliances," IEEE 7th International Conference on EBusiness Engineering (ICEBE), pp.415-420, 2010 [10] Protégé, http://protege.stanford.edu/, 2010 [11] Jess, http://www.jessrules.com/, 2009 [12] JessTab, http://www.ida.liu.se/~her/JessTab/, 2010 [13] D-Y Cheng, K-M Chao, C-C Lo, C-F Tsai, A User Centric ServiceOriented Modeling Approach, World Wide Web: Internet and Web Information Systems, in press, 2011. [14] DEHEMS, http://www.dehems.eu/ [15] C. W. Nielsen, D. A. Few, D. S. Athey, Using mixed-initiative humanrobot interaction to bound performance in a search task, ISSNIP 2008. International Conference on In International Conference on Intelligent Sensors, Sensor Networks and Information Processing, pp. 195-200. 2008 [16] M. H. Burstein, D. E. Dillerr, Cooperative Information Sharing Among Mixed-initiative Human/Agent Teams, Proceedings and Presentations of the Workshop on Mixed-Initiative Intelligent Systems, pp 17-22, 2003