Paper for the SCS/ACM Agent-Directed Simulation Symposium (ADS’09) 22 -27 March 2009, San Diego
Investigation of Information Flow in Hierarchical Organizations Using Agent-Based Modeling Jeff Waters1, James Eitelberg2, Ritesh Patel3, and Marion Ce ruti,4 Ph.D. Space and Naval Warfa re Systems Center, Pacific (SSC Pacific) Code 53621,1 Code 53226,2 Code 53627,3 Code 53624,4 53560 Hull Street, San Diego, CA 92152-5001, USA
[email protected], james.eitelbe
[email protected],
[email protected],
[email protected] Keywords: Information flow, intelligent agents,
organizational structure
Abstract This paper describes an agent-based model for exploring the effects of traditional supervisory meritbased promotions and hierarchical organizational structure on the flow of information. Each agent simulating a worker follows the same rules for handling information and promotion: Agents differ only in the breadth of info rmation they consider important. The model considers three types of information, “red,” “yellow,” and “green.” One-third of the workers (red) consider only red information important. One-third of the workers (yellow) consider both red and yellow information equally important. The final third of the workers (green) consider red, yellow, and green i nformation equally important. Red workers are more focused, restricted, and narrowminded in the type of information they consider important, whereas green workers are more unfocused and unrestricted. Starting with an organizational structure with only one supervisor and moving toward a tiered organizational structure through promotions, the objective of the simulation was to determine how this agent-based model affects the flow of information that workers consider important. The results indicate that the more narrowly focused workers will dominate the organizational structure, thereby suppressing the flow of information they do not consider important. 1. INTRODUCTION One of the goals of network centricity [3], [4], [6] and service-oriented architectures [10] is an increased capability to share information for making improved decision, which sometimes is called “getting the right information to the right person at the right time.” Various systems, processes, and structural improvements are recommended to
achieve this goal. As o rganizations consider new techniques to improve information sharing, consideration of the impact of the traditional organization structure and policies on information flow becomes increasingly important. This paper describes an agent-based model for exploring the impact of traditional promotion and hierarchical organizational structure on information flow.
2. AGENT-BASED MODELING Agent-Based Modeling (ABM) can provide useful insight where a complex system can be treated as a set of interacting autonomous co mponents (agents). ABM is especially useful where each agent follows simple rules, but the system as a whole exhibits emergent behaviors that may or may not be readily apparent from the rules governing the behavior of individual agents. (See, for example, [1], [2], and [5]) In ABM, a key assumption is that the whole is more than the sum of its parts. ABM has been used to explain a wide variety of phenomena, events and observations based on emergent “flocking” behavior from simple individual rules of interaction. For example, this has been used to explain: 1) flocking in birds; 2) timeliness and characteristics of outbreaks of “civil unrest” from simple notions of how aggrieved individuals behave in public asse mblies, and 3) the emergence of an “invisible college” from simple concepts of how college professors collaborate on papers. (See, for ex ample [5].) This paper covers the following topics: • Information flow, treated as a complex process of interacting individuals that may result in emergent behaviors of the system as a whole. • An information-flow model. • The assumptions and simple rules of the model for information flow in an o rganization. • The model implemented in NetLogo, an ABM modeling and simulation system.
• • • •
Discussion of expected values, The results of running the model. Improving information flow. Implication of the results for organizations and
•
communities of interest. Directions for future research, enhancements to the model
The model provides some insight into potential problems with information flow as a result of traditional organizational structure and promotion
policies.
including
Figure 1. Information Flow Model at Setup
3. MODEL CONCEPTS This section describes the co ncepts as represented in the model. The model was developed with some simplifying assumptions about information flow in an o rganization. Every agent represents a “worker” in the organization. The model includes simple representations of “workers,” “supervisors,” (sometimes called administrators) “information passing,” “information types,” and “promotions.” The model arbitrarily considers three types of information, labeled red, yellow, and green. Not all types of information are considered important by all supervisors. The model also features three types of workers and supervisors, also labeled “red,” “yellow,” and “green.” Red supervisors consider only red information important. Yellow supervisors
consider both red and yellow information equally important, but do not value “green” information. Green supervisors co nsider all three types of information (red, yellow, and green) equally
important. The agent workers also differ only in the breadth of interest in what they consider impo rtant. One-third of the workers are labeled “red” because they produce only red information and consider only red information impo rtant. One-third of the workers are labeled “yellow” because they produce both red and yellow information and consider both red and yellow information equally important. The final third of the workers are labeled “green” because they produce red, yellow and green information and consider these colors of information equally important. Like their
supervisory counterparts, red workers are more focused, restricted, and narrow-minded in the type of information they value, whereas green workers are more unfocused, unrestricted, and open minded. The objective of the model is to explore the effects on the organization over time under conditions described above. (a) The most valued workers are promoted. (b) Each supervisor assesses value based on useful information that each subordinate worker passes to the supervisor. (c) Workers are co nsidered to have different breadth of interest in types of
information considered useful. Workers can be promoted to become supervisors (administrators). Initially, only one supervisor is present, simulating an entrepreneurial-like initial beginning with a relatively flat organizational structure. Figure 1 is a screen dump that depicts the information-flow model at the start of a simulation. With each time step in the simulation, each worker produces and passes a piece of information to the worker’s supervisor. If the supervisor considers the information important, the worker's organizational value is increased by one. Periodically, a promotion step occurs. Each supervisor promotes a specified number of workers to be subordinate supervisors based on their organizational value. The workers that are promoted retain the original colors that they had as non-supervisory workers after they become supervisors. Thus, a red worker will become a red supervisor and a green worker, a green supervisor. The non-supervisory workers are divided equally among the new supervisors and their organizational value is reset to zero reflecting the fact that they need to re-earn their o rganizational value under this new supervisor. The effect of the promotion is to establish a new level of management in the supervisory hierarchy. Up to four levels are allowed, including the lowest worker level and three upper levels of increasingly powerful administrators. The time steps and promotions continue until eventually every supervisor has a specified number of su bordinate workers (in this case, four), so no new promotions occur until periodically, a supervisor “retires.” When a worker is promoted to a higher level, another worker of the same color replaces the promoted worker. A retiring supervisor is replaced by the supervisor's most highly valued subordinate. If this subordinate were a supervisor previously, the subordinate's old position is filled in the same
manner. If the subordinate were a non-supervisor worker, a new worker is created to fill the empty spot to maintain the same total number of workers. Organizational value is determined by the number of times the su bordinate passed “useful” information to the immediate supervisor. The information is deemed “useful” if that supervisor finds that type of information impo rtant. Workers who pass information that the supervisor does not co nsider important do not increase their organization value. They are less valued and less likely to be promoted. Due to the structure of the four-level hierarchy, the requirements that each supervisor shall manage at most four workers, and the requirement that the number of each color workers remain about equal, there are not enough workers of any single color to hold all the supervisory positions at the three upper levels. However, the structure of the simulation ensures that two of the three colors will be represented at the supervisory levels. 4. EXPERIMENTAL SECTION 4.1. Agent Rules
Each worker, regardless of information-color preference, follows the same simple rules for passing and accepting information and for handling promotions: • Pass to your immediate supervisor information that you think is impo rtant. • Accept information that you consider impo rtant from your immediate subordinates, if any. • Increment the organizational value of any subordinate who passes you information that you consider impo rtant. • At the appropriate time, promote up to four of the highest valued subordinates to be supervisors. 4.2. Running the Model The model maintains a roughly equal number workers of each information type (red, yellow and green). Workers are designed to represent different narrowness of interest in multiple types of information. Some workers (e.g. Green workers) think multiple types of information are equally important. At the other extreme, Red workers think only one type of information is important. The color of the information produced at each step is chosen randomly. The model enables exploration of any emerging behavior that occurs in the organization
when traditional promotion and hierarchical organizational structure are combined with different breadths of interest in the various types of information. In this model, workers do not use adaptation, which includes the use of feedback as inputs to modify their behavior. The model was implemented in NetLogo, which is an open-source, simple but powerful modeling-andsimulation environment [9], [7]. Figure 2 shows a screen capture of the model after it was allowed to run, thus illustrating the four levels that were allowed in the organizational hierarchy. The boxes represent
workers or supervisors.
5. INITIALLY EXPECTED VALUES Based on the model described above, yellow supervisors consider two types of information important and green, consider three types of information important, the lowest common denominator being six. Therefore, let six “points” be allocated to each supervisor, the distribution of which depends on the color(s) of information that each supervisor values. Thus, initially, the red supervisor has six red points, the yellow supervisor has three red points and three yellow points, and the green supervisor has two red points, two yellow and two green.
Figure 2. Information Flow Model in Operation Assuming that all colors are equally likely among starting supervisors, the total number of points would be 18, inclu ding 11 red points, 5 yellow points and 2 green points. If promotions are based on the passage of information with these weighting factors, initially,
one could expect to observe 11/18 red promotions, 5/18 yellow and 2/18 green. Therefore, it is not surprising that 30% of the promotions were yellow during the number of time steps in which the model
was run. That is about 5/18, which is 28%. As for the green, the figure is about 11%. The first top-level supervisor is green One would expect to see 11% of the promotions for green supervisors, 28% of the promotions for yellow supervisors, and the rest would go to red supervisors. However, the initial expected values are not the same as the end-state expected value because as the simulation progresses, even after the first promotion cycle, the numbers of green promotions begins to decline below the initial value.
6. RESULTS Figure 3a shows the number of supervisors or administrators of various colors as a function of simulation time step. Promotions are important because they determine the types of information that the organization values and co nsiders important at any given time during the simulation. Red begins to predominate on the average, even during the first
promotion cycle.
a red
yellow
green
Thus, the initial probabilities of information flow by information type differ from the values in the middle and end of the simulation. The middle-and end-state probabilities that various information types will be valued (and therefore, will flow) also change as a function of the number of supervisors of each color. Over time, the organization will extinguish the lessvalued types of information, in this case, green. This suggest that the flow of any less-valued information, including yellow, could tend toward zero over time as more and more red supervisors populate the organization and pass on their opinions regarding the lesser-valued non-red info rmation. However, the rules of the simulation do not permit enough red workers to be available to fill all the administrative positions. Therefore, some yellow supervisors always will be available to promote yellow workers. Figure 3b shows the percent of information flow by information type, the graph shows a bold line at the top indicating that red information is optimally flowing at 100% as the red information passed upward is accepted and valued. Because red is always valued it flows most freely. The graph suggests that the acce ptance and valuation of yellow information averages around 30%. According to the initially expected value calculation in section 5, Figure 3b should show 28% yellow information. Eventually, at the end of the simulation period, no green supervisors remain in the organization. Therefore, the flow of green info rmation will extinguish to 0% as the simulation progresses. In figure 3a, this occurred after 7625 time steps, whereas in another run of the simulation not shown
here, it occurred after 3430 steps.
b red yellow
green
Subordinates are valued according to how much useful information they pass to supervisors. Supervisors promote those they value most. As the model runs, the more narrow-focused red workers dominate the hierarchy resulting in less yellow and green information flowing, as shown in Figure 3. All supervisors will promote workers who pass red information but only green supervisors will promote workers who pass green information. As the number of green supervisors falls to zero, no supervisor is
left to value green information. 7. DISCUSSION
Figure 3. a) Number of supervisors of each type as a function of time. b) Percentage information flow by information type.
7.1. Model and Simulation Green information flow extinguishes at some point if the simulation is allowed to progress through a
sufficient number of time steps. The rate at which workers are promoted influences the rate at which green information is extinguished. The faster the rate of promotions, or alternately, the fewer time steps between promotions, the faster the attrition of the green information. However, this exact point cannot be predicted due to the stochastic mechanisms built into the simulation. This result is expected to differ if the number of red and yellow supervisors is limited to ensure that the number of possible supervisory positions exceeds the maximum combined number of red and yellow workers. In that case, green information will not extinguish due to the presence of green supervisors. With adaptation, agents are expected to learn very quickly which color of information to pass to maximize the probability of promotion. Thus, adaptation is expected to drive the suppression of green info rmation much more rapidly to zero, as agents pass red information, which is always valued, and neglect anything risky, such as green. The bias that favors red-information flow was introduced when the rules were formulated. As discussed in section 6, a green supervisor could promote a worker who passes only red information but the red supervisor will never promote any worker who passes only green information. The assumption of overall equal information value was not built into the model. Therefore, all supervisors will not all value equally information of different colors. Green is not considered as useful as red and yellow information or else the red and yellow supervisors would value it. The behavior of the o rganization is consistent with the rules for individual behavior. Nowhere in the model does the validity or correctness of the info rmation appear. If information is not highly valued, it will not flow within the organization, regardless of whether or not the information is correct. If information is valued and workers are r ewarded for passing it, it will flow freely, regardless or how valid or co rrect it is. This is not necessarily a weakness of the model but it could be considered a weakness of the organizations that the model was designed to represent. 7.2. Implications for Organizations The results suggest that as the age of an organiz ation or established community increases, the organization or community may become more “set in its ways” and encourage the same type of thinking that corresponds to that of the leadership. This status-quo
bias is well known and has been documented [8]. Newer organizations that have not yet achieved the suppression of rarely valued information are more likely to allow the emergence of thinking that does not always agree with the majority of its highly
promoted leaders. This shows that merit has some drawbacks as a promotion strategy. From an information flow perspe ctive, leaders of organizations might find this problematic. It may not be the intention of an organization to suppress the flow of rarely valued information. However, this simulation shows that by selecting workers for promotion based on the value of their work as defined by their supervisors (i.e. merit), the effect indeed may be to extinguish the flow of info rmation that does not match the predominant ideas in that organization. This approach to promotions is considered by many organizations to be the most fair, reasonable, and logical. The obvious flaw in the “merit” system is the subjectivity regarding what constitutes merit. Organized communities without well-defined boundaries, such as scientific, medical, or legal communities, tend to exhibit behavior similar to that of formally established organizations in which the boundaries are clearly defined. This is especially true when the organization that emerges in communities determines the flow of information. Thus, ideas are promoted or suppressed much in the same way that workers are promoted or denied promotion in a formal organization per se. For example, in an organization or community dedicated to research and development (R&D), the suppression of new ideas, approaches, and applications that may conflict with old ways of thinking and o perating is the antithesis of innovation. However, established members of such co mmunities often seek to preserve the status quo by controlling the information that is valued. This is evident in the review of papers submitted to journals and conferences. If a paper is describes observations that conflict with the predictions of a theory for which a reviewer is a proponent, that reviewer will tend to suppress the paper even if it may be right and the theory may be based on a wrong assumption or at least different assumptions from the paper. Whereas no one wants faulty information to become accepted as established facts, the consequence of merit-based evaluations is that the same filter that prevents flawed studies from beco ming accepted also may
screen out revolutionary new and different ideas that may prove to be useful and valuable in the future.
election of the peers might produce an organizational behavior that differs from that of the merit system.
Compared to the “merit” system, information will flow more freely in an organization or community where promotions are based on seniority. Whether the free flow of unusual ideas in an organization is seen as an advantage or a disadvantage d epends on how well individuals can convince their peers and supervisors that an u nusual idea has merit. This is the same subjectivity associated with the merit system. However, one obvious advantage of seniority-based promotions is that given a clear definition of what constitutes seniority, the rank order of individuals within the organization is objective.
In a future study, the relative availability of the different colored information can be varied. If all information is equally available, with adaptation all agents will learn to select red, but if red is scarce, this would produce a different result.
Another way to determine promotions is by election, as is done in some organizations, where appointed or elected administrators elect the workers to be promoted, including but not limited to the supervisor to head the entire organization. Assuming that the enumeration of ballots is an unambiguous process, and that the determination of voter intent is, likewise, unambiguous, elections also constitute a definitive method to determine promotions. The subjectivity and potential bias in this case resides in the minds of the electors, who may use various criteria in selecting the candidate for whom to cast their votes. The object of this paper is not to argue either for or against the retention of merit as the prime criterion for promotion or idea acceptance. It does, however, reveal a disadva ntage of any merit-based system that may not be obvious to the leaders of the organizational or community. The results of this study underscore the disproportionate power that toplevel supervisors and community leaders have in merit-based organizations and communities as they promote workers of like thinking or ideas of which they approve, and neglect individuals who think that other ideas are impo rtant.
8. FUTURE RESEARCH Future stu dies based on different promotion models can be considered, in addition to merit-based promotions, such as seniority and election. One would expect improved information flow from organizational structures or promotion strategies that do not depend only on what the immediate supervisor thinks of a worker. As discussed above, promotion by seniority might be expected to have better breadth of information flow than the traditional merit approach. Also promotion by
9. CONCLUSION The information-flow model described in this paper suggests that organizations with traditional hierarchical structure and traditional promotion strategies are likely to have a restricted breadth of information flow because narrowly focused workers are tend to be promoted into upper management more so than workers who think other information is important. Although the actual functioning of organizations is much more complex than that represented here, the model may provide some insight and will be enhanced in future versions to incorporate a variety of promotion strategies and worker adaptations. The information-flow model described in this paper is available free of charge. Email requests to
[email protected]. 10. ACKNOWLEDGEMENTS The authors thank the Office of Naval Research for their support of this work. This paper is the work of U.S. Government employees performed in the course of their employment and no copyright subsists therein. 11. REFERENCES [1] Behrman, R. and Carley, K. Modeling the Structure and Effectiveness of Intelligence Organizations: Dynamic Information Flow Simulation. In Proceedings of the 8th International Co mmand and Control Research and Technology Symposium, ICCRTS (Vienna, VA, USA, 2003.) CCRP press. [2] Castiglione, F. Large-scale agent-based mo dels perspectives and requirements. In Proceedings of the IMA “Hot Topics” Work shop: Agent Based Modeling and Simulation (Minneapolis, MN, USA, Nov. 3-6, 2003.) [3] Ceruti, M.G. Mobile Agents in Network-Centric Warfare,” Inst. of Electronics Information and Communication Engineers Trans. on Communications (Tokyo, Japan. 2001) Vol. E84B, No.10, 2781-2785. [4] Ceruti M.G. and Powers, B.J. Intelligent Agents for FORCEnet: Greater Dependability in Network-Centric Warfare In Suppl. Vol. of the
Proc. of the IEEE International Conf. on Dependable Systems and Networks (DSN 2004) (Florence, Italy, June 28 – July 1, 2004) 46 -47. [5] Macal, C.M. and North, M.J. Introdu ction to Agent-based Modeling and Simulation,” MCS LANS Informal Seminar, Nov. 2006. http://www.mcs.anl.gov/~leyffer/listn/slides06/MacalNorth.pdf [6] Mayo, R.W. VADM, USN and Nathman, J. VADM, USN. Sea Power 21 Series – Part V: ForceNet: Turning Information into Power. Naval Institute Proceedings, vol. 129, no. 2, (Feb. 2003. USNI press.) 42 -46. [7] NetLogo http://ccl.northwestern.edu/netlogo/ Craig Reynolds http://www.red3d.com/cwr/ibm.html [8] Samuelson, W. and Zeckhauser, R. Status Quo Bias in Decision Making. In Journal of Risk and Uncertainty, 1988, 1, 7 -59. [9] Wilensky, U. 1998. NetLogo Flocking model.
http://ccl.northwestern.edu/netlogo/models/Flocki ng. Center for Connected Learning and Computer-Based Modeling, Northwestern University, Evanston, IL. Inspired by Craig Reynolds’ Boids Simulation: http://www.red3d.com/cwr/boids/ [10] Wilcox D.R. and Ceruti, M.G. A Structured Service-Centric Approach for the Integration of Command and Control Components. In Proceedings of the IEEE International Conference on Service Compu ting (SCC 2008). (Vol. 2, July, 7 -11, 2008, Honolulu, HI, USA) 5 12. Mr. Jeff Waters is a scientist in the Co mmand and Control Technology and Experimentation Division of the Command and Control Department at SSC Pacific. He earned a B.S. degree in computer science from Iowa State University in 1985. He also earned a B.A. in English and a J.D. in law. Mr. Waters is the principal investigator of a research project designed to study the effects of information velocity on command-center performance. His current research interests include dynamic command and control, agent-based computing, service-oriented architecture, and global information-systems interoperability. Mr. Waters received the “Top Scientist of the U.S. Navy” award in 2007 and the Meritorious Civilian Service Award in 2008.
Mr. James Eitelberg is an engineer in the Command and Control Command and Intelligence Systems Division of the Command and Control Department at SSC Pacific. He earned a B.S. degree in mathematics from Western Washington University in 2000 and a M.S. in Systems Engineering from the Naval Postgraduate School in 2007. His research interests include service-oriented architecture, Web services and open standards. Mr. Ritesh Patel is a systems engineer in the Command and Control Technology and Experimentation Division of the Command and Control Department at SSC Pacific. He has completed many projects in software and systems development and engineering, systems interoperability, networks, and sensor systems. One of his most recent projects concerns the operational implementation of the Joint Warning and Reporting Network, which is a program of record in Chemical and Biological defense. Dr. Marion G. Ce ruti is a scientist in the Command and Control Technology and Experimentation Division of the Command and Control Department at SSC Pacific. She earned the Ph.D. in 1979 from the University of California at Los Angeles. Dr. Ceruti's professional activities include information systems research and analysis for command and control decision-support systems, modeling and simulation, intelligent agents, information dynamics, command and control, sensor-data fusion, and research management. She is the author of over 110 journal articles, conference proceedings, monographs and book chapters on various topics in science and engineering. Some of her awards include 2002 Publication of the Year, the SSC Pacific Exemplary Achievement Award, and the SSC Pacific Award for Excellence in Science and Technology. Dr. Ceruti is a senior me mber of the IEEE and a member of the IEEE Computer Society, the Association for Computing Machinery, the Armed Forces Communications and Electronics Association, the International Society for Computers and Their Applications, the Liophant Simulation Club, and the New York Academy of Sciences.