Keywords Intelligence, Image, Logic, Decision making, Cybernetics, Management. Abstract Image theory has been used, in numerous studies, as a basis for ...
The current issue and full text archive of this journal is available at http://www.emerald-library.com/ft
Kybernetes 30,2
166
Intelligent agent behavior based on organizational image theory David G. Schwartz and Dov Te'eni
Graduate School of Business Administration, Bar-Ilan University, Israel Keywords Intelligence, Image, Logic, Decision making, Cybernetics, Management Abstract Image theory has been used, in numerous studies, as a basis for understanding and describing the decision-making activity of managers in both cooperative and competitive environments. The fundamental division of duties prescribed by image theory ± namely adoption decisions and progress decision ± maps very well to the adaptability requirements of intelligent agents. The issues of adaptive planning and execution monitoring in agents can be well served by applying the empirical lessons learned from the application of image theory across groups of decision makers. This paper explores the concepts of adoption and progress decisions in the context of image theory and provides a basis for creating image-theoretic agents. This paper sets the foundation for an interdisciplinary bridge between Beach and Mitchell's Image Theory for human decision making, and the construction of intelligent agents. We begin by presenting image theory and describing its use among human decision makers. We then show how the mechanisms of image theory can be implemented in an agent-based architecture to implement both execution monitoring and adaptive planning. This is done through the image-theoretic constructs of progress decisions and adoption decisions. We conclude by presenting logic-programming implementation of the Imaginal Agent Architecture that supports the adaptive planning and execution monitoring of agents through the use of meta-level constructs for adoption and progress decisions.
Kybernetes, Vol. 30 No. 2, 2001, pp. 166-178. # MCB University Press, 0368-492X
Introduction A decision-making agent is an agent whose primary activities involve the selection of suitable goals, the determination of plans through which those goals can be achieved, and the implementation of such plans in the pursuit of effective decision making. As found in many agent architectures, a decisionmaking agent must make use of some form of values (beliefs), goals (desires), and plans (Georgeff and Lansky, 1987; Suchman, 1987; Shoham 1993). It is the dynamic modification of those values, goals, and plans that will allow an agent to exhibit adaptive characteristics. Agents exhibiting adaptive characteristics are capable of dynamically adapting their problem-solving methodologies in order to achieve superior performance in their decision-making tasks. This paper presents a treatment of image theory (Beach and Mitchell, 1987, 1990, 1998) as a basis for creating adaptive intelligent agents. Image theory is a theory of organizational behavior and managerial decision making that has heretofore been applied exclusively to the modeling of human decision-making activity. Previous work on image theory as a basis for agent architectures has been limited to determining a set of agent design principles based on the decision-making framework suggested by the theory (Schwartz and Te'eni, 1996). In contradistinction to de Raadt (1991), who proposed a cybernetic
approach to information systems and organizational learning, we are Intelligent agent developing an organizational learning approach to elements of cybernetics. behavior Image theory has been used, in numerous studies, as a basis for understanding and describing the decision-making activity of managers in both cooperative and competitive environments. The fundamental division of duties prescribed by image theory ± namely adoption decisions and progress 167 decision ± maps very well to the adaptability requirements of intelligent agents. The issues of adaptive planning and execution monitoring in agents can be well served by applying the empirical lessons learned from the application of image theory among groups of decision makers. This paper places particular emphasis on the concepts of adoption and progress decisions in the context of image theory and provides a basis for creating image-theoretic agents. We begin by presenting image theory and describing its use among human decision makers. We then show how the mechanisms of image theory can be implemented in an agent architecture to improve both adaptive planning and execution monitoring. This is done through the image-theoretic constructs of adoption decisions and progress decisions. We conclude by presenting logicprogramming implementation of the Imaginal Agent Architecture that supports the adaptive planning and execution monitoring of agents through the use of meta-level constructs for adoption and progress decisions. Background and related work Intelligent agents Trying to present a consensus definition of intelligent agents has long been abandoned, and rightly so. Russel and Norvig (1995) provide what is perhaps the most general of accepted definitions, considering an intelligent agent to be any entity that perceives its environment through sensors and effects its environment through some form effector. Agent architectures Agent architectures have been approached from a number of different perspectives ranging from game theoretic models (Rosenschein and Zlotkin, 1994) to biological models (Beer, 1990). In between we find the use of shared mental models (Sycara and Lewis, 1991), and social interaction models (Gasser, 1991; Hewitt, 1991; Star, 1989; Werner, 1989). The adaptive nature of agents has received considerable attention from the perspectives of traditional learning and discovery, knowledge-based modification (Imam, 1996), and dynamic plan generation (Pollack, 1992). Sedbrook (1994) draws parallel between dynamic group processes and genetic learning. Moore et al. (1998) propose a brokered agency architecture with four types of agents: (1) user agents to handle the UI and initial problem partitioning to one or more broker agents; (2) broker agents to divide and subcontract out different subproblems;
Kybernetes 30,2
168
(3) domain specialists who handle actual subproblem resolution; and (4) database wrappers who deal with the integration, access, and update of legacy data required by the other agents. Classic adaptive agent algorithms address adaptability exclusively in terms of changing the plan to be executed. The issues of adapting fundamental agent behavior, tendencies, strategies, and priorities have taken a back seat to the decision-theoretic approach. The application of organizational behavior and managerial decision-making models has yet to receive thorough treatment in an agent architecture. Human decision makers In selecting an appropriate theoretical basis for an agent architecture, it is important to consider the performance characteristics of the theory. If the theory is ``new'', i.e. with no basis in contemporary literature to support its effectiveness, it becomes difficult, if not impossible, to justify the theory as a basis for agent behavior. It is perhaps the difficulty of this task that has resulted in many agent architectures lacking a firm theoretical basis. Humans have been shown to adapt well to changing conditions (Payne, 1982). A series of simulations and laboratory experiments have shown that adapative rules for adopting decision rules can describe decision behavior well, and perhaps more importantly, that this behavior is efficient (Bettman et al., 1990; Payne et al., 1988). In fact, it has been argued that some of the human tendencies or biases that are often the basis for assuming suboptimal behavior (e.g. Kaneman and Tversky, 1972) are human ways of adapting in a changing environment (see Einhorn and Hogarth, 1981). For example, Ben-Bassat and Te'eni (1984) simulated human heuristics, such as discounting unfamiliar attributes, and found that human heuristics did no poorer than Bayesian decision rules. One of the main factors in the high quality of human decision making is the adaptive nature of the human decision maker. We believe that by examining the theoretical bases of effective human decision making, agent researchers can build more robust decision agents with effective mechanisms for both execution monitoring and adaptive behavior. This approach is consistent with the directions taken by Suchman (1987) in her work on situated action, as well as Georgeff and Lansky (1987), Rao and Georgeff (1993) and others. Image theory This paper focuses on a specific theoretical basis for human decision making known as image theory. Image theory was developed by Lee Roy Beach and Terence R. Mitchell in the late 1970s and has since served as a basis for numerous empirical studies of managerial decision making (Dunegan, 1995; Beach and Strom, 1989; Beach et al., 1988). The following section explains image theory and presents its suitability as a basis for adaptive agent architectures. This is followed by a discussion the
implementation of image theory and its main algorithmic components that Intelligent agent produce the adaptive behavior of image-theoretic agents. behavior Image theory Image theory has been developed along two complementary tracks: for personal decisions (Beach and Mitchell, 1990, 1987); and for decisions within organizations. Our primary concern is with the latter given its focus on individuals acting as decision agents on behalf of organizations. Types of images In organizational image theory, a decision maker can be profiled in terms of his or her organizational images. There are four such images, as summarized below (Beach and Mitchell, 1990, 1998). Each type of image is a representation of the decision maker's perception of some aspect of the organization. (1) Organizational self-image. The organizational self-image consists of the beliefs, morals, ethics, values, norms, etc. that are generally accepted across the organization. These images are formed irrespective of the individual's opinions and are meant to provide an accurate reflection of the organization's principles. It is these principles that will provide the basis upon which new goals are generated and candidate goals are evaluated. (2) Organizational trajectory image. The organizational trajectory image contains the goals and goal markers that comprise the organization's agenda for the future. These goals can be concrete events, abstract states, or interim non-goal states (markers) that are milestones on the path to a goal. (3) Organizational action image. The organizational action image is made up of a set of plans associated with each of the goals in the organizational trajectory image. Each plan consists of tactics that implement the specific behavior of a given plan. (4) Organizational projected image. The organizational projected image is a forecast of anticipated events and states that are expected as result of implementation of the organizational action image. Adoption and progress: fundamental activities for the adaptive agent In image theory, the decision maker modifies his image on an ongoing basis. This dynamic adaption is considered fundamental to the activity of human decision makers. Each of the self, trajectory, and action images are serviced by adoption decisions whose purpose is to adopt or reject candidate goals and plans from these three images. A second decision category, progress decisions, services the organizational action image by determining which of its plans are progressing in a satisfactory manner towards its goals.
169
Kybernetes 30,2
170
Adoption decisions Adoption decisions determine whether to augment current goals and strategies or use them as they are. Adoption decisions rely on two types of tests: compatibility and profitability. The compatibility test is a yes/no test and is non-compensatory. It tests the fit between a candidate decision policy and the decision maker's set of images. A candidate is immediately adopted if it is the only candidate that fits ± where fit is defined as being within some threshold value to the attributes in the decision maker's images. If several candidates fit the decision makers images, the profitability test is applied to select the best candidate. The profitability test evaluates the potential pay-offs (both positive and negative) of a given candidate goal with respect to alternative candidates. There are many ways in which decision strategies can be compared, ranging from the effortless random choice up to highly analytic procedures such as maximum expected utility (see the cost-benefit model defined in Beach and Mitchell (1987). During the adoption phase, these two tests are applied in sequence, the compatibility test filtering out unacceptable candidates, and the profitability test selecting the best from the remaining compatible candidates. One aspect of image theory's adoption decisions that is of significant interest is its emphasis on the do-nothing reaction. The tendency of human decision makers towards protecting the status quo has been shown to be an important factor in the efficiency of human decision. This tendency, when mapped to an adaptive agent architecture, can result in more efficient agent activity. Progress decisions Progress decisions monitor the processes underway to determine whether the plans being implemented still fit the goals or require modification. They do this by performing a comparison between the organizational trajectory image and the organizational projected image. If there is a reasonable fit between the two images, it is indicative that the organizational action image contains suitable plans for achieving the trajectory image's goals. If the trajectory and projected images are found to be incompatible, it is indicative of either unsuitable goals in the trajectory image, inappropriate plans in the action image, or both. In either case, it requires some form of adaptive response. Image theory does not prescribe the exact procedure for generating the adaptive response when events generate conditions that are not specified in the current images. We suggest that the reflective nature of effective decision makers should also be modeled into the image theory framework. Reflection is needed when the current way of thinking breaks down and new components of the images must be developed. Effective reflection is characterized by systematic transition between levels of abstraction, between the activities of examining current images and testing new ones, and between the different parts of the problem (Srinivasan and Te'eni, 1995). In the image theory framework, the levels of abstraction are given by the hierarchy of the images, and the activities are shown in Figure 1. When the trajectory images cannot
Intelligent agent behavior
171
Figure 1. Decision cycle of an image theoretic agent
explain the unexpected results of action, and new plans must be developed, they must be developed by consulting the higher levels of the self-image. But developing new plans also means testing them by forecasting the results of action ± a simulation, if you will. In terms of intelligent agent architectures, progress decisions play a fundamental role in providing execution monitoring. This execution monitoring, together with the recursive application of adoption decisions to modify both the trajectory and action images, provide a complete adaptive planning mechanism. Implementing the Imaginal Agent Architecture Motivation Implementation of the Imaginal Agent Architecture uses the Prolog language. The choice of Prolog had two primary motivations: (1) the clarity provided by the declarative representation of Logic programming makes the mapping from Beach and Mitchell's loosely algorithmic description of image theory more intuitive; (2) the meta-level characteristics of Prolog provide an elegant and effective framework for implementing the different layers of decision-making activities that image theory demands of the imaginal agents. Object and meta are relative terms. Two languages O and M, the object language and meta language respectively, are in an object-meta relationship if there is a representation of the language O in M. Meta-programs are programs written in meta-language which manipulate and/or reason about an object language. The object-meta relationship is of particular interest in the design of adaptive agents in that it provides an opportunity for reflection ± jumping between the object and meta language levels. The reader is referred to (Schwartz, 1995) for a discussion of the use of Prolog and meta-interpreters as a
Kybernetes 30,2
172
basis for agent architectures. In the interest of brevity, we have shown only the primary predicates used in implementing the theory. In the implementation of the imaginal agent architecture, the image itself is considered a component of the object level and not of the meta level as one may expect. The reason for this becomes clear when you consider that the actual images are agent-specific and are tightly bound to the specific goals and plans that an agent must execute. It is the monitoring and modification of these images that takes place at the meta-level. Mapping the theory to agent constructs and behaviors Figure 2 presents the knowledge-base constructs that comprise the object-level of an image-theoretic agent. Here we find representation of each of the image components required by the decision maker. The aggregate agent_image/4 predicate is used as a container for the four organizational images. The plan/1 predicate is a surrogate for any number of possible implementations of a plan representation. This is another advantage of the meta-interpreter approach ± we can rely on Prolog's clause resolution mechanism to provide the meta-level with goal-size pieces of the planning predicates, irrespective of their complexity. In other words, the generality of meta-interpretation lets us avoid a prescriptive plan representation and allows different types of agents to use their own plan representation. The image-theoretic meta-interpreter appears in the Appendix. The metainterpreter takes the agent's action image as input and proceeds recursively down the plan elements in the action image. As each planning step is implemented, the meta-interpreter detects new events. These events can be both internally and externally generated. For each step in plan implementation, the image_theory/3 predicate is invoked with the current context, current image representations, and current event list.
Figure 2. Image theoretic knowledge-base constructs
If a suitable policy is found for an equivalent context, then this policy's actions Intelligent agent are performed. If no suitable context was found, we then forecast the result of behavior doing nothing. If the resulting status quo is compatible with the decision maker's images, then no action is taken and the event passes with no response. The forecast/4 predicate is used to project the results of an event occurring in the current context with no action being taken on the part of the agent. The 173 compatible/2 predicate applies a compatibility test of the forecast results with the current image. Execution monitoring takes the form of ongoing comparisons between the organizational trajectory image and the organizational projected image ± the progress decisions. If the forecast of a do-nothing decision results in the creation of a trajectory image that is incompatible with the projected image, then intervention is required. It is here that adaptive behavior begins to take place ± the adoption decisions. After intervention, the new goals and plans are integrated with the existing image to form a new image which is used in future goal processing. The modification of the decision maker's images will have a direct effect on future progress decisions through the compatibility test, thus adapting the agent to new circumstances and changing its projected future reactions. The decision cycle, as illustrated in Figure 1, continues until the action image is devoid of plans. Future directions Modeling decision scenarios Following completion of the Prolog implementation, the Imaginal Agent Architecture must be tested in a variety of decision scenarios. One such scenario, the hiring decision, described in terms of image theory constructs by (Schwartz and Te'eni, 1996), is currently under implementation. Another candidate scenario is the project management decision presented in (Dunegan, 1995). This example is of particular interest as it has been applied in two different empirical experiments testing the use of image theory by human decision makers. Dunegan's results provide us with an interesting benchmark with which to evaluate the performance of image-theoretic agents. Examining group decision making In this paper we have examined image theory from the perspective of an individual decision maker within an organization. This was done with the intent to focus our adaptive agent architecture on the functions and requirements of individual image-theoretic agents. Extending this discussion from the individual decision maker to group decision making within an organization is a natural next step. Examining additional theory-based decision models The management and decision science literature is an excellent source of other models that are of interest when designing adaptive agents. Of
Kybernetes 30,2
174
particular interest is the Vroom and Yetton Model of Leadership (1973). This model is interesting as it places great emphasis on the decision maker switching between different decision-making styles. Vroom and Yetton define five decision models with accompanying metrics for switching between models. The ability to switch between decision models is both a strong indication of adaptability and an excellent match for the meta-level architecture. We are currently considering the Model of Leadership theory as a possible augmentation for image theory in moving from individual decision to group decisions, given the Model of Leadership's emphasis on multiple participants. We believe that further investigations into the application of managerial decision models to the design of intelligent agents can both benefit from, and contribute to, the two disparate disciplines. Conclusions In this paper we have presented the use of Beach and Mitchell's image theory as a basis for building adaptive intelligent agents. The resulting architecture, called the Imaginal Agent Architecture, is a framework that is built around a hierarchy of images and the activity at the different levels of the hierarchy. The Prolog implementation shows at a general level the feasibility of building such an architecture. Implementing image theory as a meta-interpreter provides us with a framework in which to experiment with the different progress and adoption tests that can be developed for different agents, without requiring any change to the underlying image-evaluation mechanism. Our main goal in presenting this framework has been to set the foundation for an interdisciplinary bridge between management decision theory and the construction of intelligent agents. The adaptive nature of human decision behavior seems to have potential use in the design of intelligent agents. Effective decision behavior rests on being adaptive in both decision strategy adoption and in decision strategy implementation. Image theory provides the theoretical basis for explaining adaptive decision behavior by employing both progress decisions on the implementation side and adoption decisions on the decision strategy side. We extend image theory to include reflective human heuristics for progress decisions. Moreover, the human decision maker's tendency to do nothing, i.e. maintain the status quo, finds expression within image theory and in the consequent imaginal agent architecture. This paper attempted to specify the processes of decision behavior according to the extended image theory, in order to come closer to an architecture of adaptive intelligent agents that is consistent with the theory. The implementation of managerial decision scenarios, as well as the extension of the Imaginal Agent Architecture to support group decision making, provide ample interesting directions for the continued exploration of image-theoretic agents.
References Beach, L.R. and Mitchell, T.R. (1987), ``Image theory: principles, goals, and plans in decisionmaking'', ACTA Psychologica, Vol. 66 No. 3, pp. 201-20. Beach, L.R., and Mitchell, T.R. (1990), ``Image theory: a behavioral theory of decision making in organizations'', in Staw, B.M. and Cummings, L.L. (Eds), Research in Organizational Behavior, JAI Press, Greenwich, CT, Vol. 12, pp. 1-41. Beach, L.R. and Mitchell, T.R. (1998), ``The basics of image theory'', in Beach, L.R. (Ed.), Image Theory: Theoretical and Empirical Foundations, Lawrence Erlbaum, Hillsdale, NJ, pp. 3-18. Beach, L.R. and Strom, E. (1989), ``A toadstool among mushrooms: screening decisions and image theory's compatibility test'', ACTA Psychologica, Vol. 72, pp. 1-12. Beach, L.R., Smith, B., Lundell, J. and Mitchell, T.R. (1988), ``Image theory: descriptive sufficiency of a simple rule for the compatibility test'', Journal of Behavioral Decision Making, Vol. 1, pp. 17-28. Beer, R. (1990), Intelligence as Adaptive Behavior: An Experiment in Computational Neuroethology, Academic Press, New York, NY. Ben-Bassat, M. and Te'eni, D. (1994), ``Human-oriented information acquisition in sequential pattern classification'', IEEE Transactions on Systems, Man and Cybernetics, SMC-14, pp. 131-8. Bettman, J., Johnson, E. and Payne, J.W. (1990), ``A componential analysis of cognitive effort in choice'', Organizational Behavior and Human Decision Processes, Vol. 45, pp. 11-39. de Raadt, J.D.R. (1991), ``A cybernetic approach to information systems and organizational learning'', Kybernetes, Vol. 20 No. 1, pp. 29-48. Dunegan, K.J. (1995), ``Image theory: testing the role of image compatibility in progress decisions'', Organizational Behavior and Human Decision Processes, Academic Press, New York, NY, Vol. 62 No. 1, pp. 79-86. Einhorn and Hogarth (1981), ``Behavioral decision theory'', Annual Review of Psychology, Vol. 32, pp. 15-41. Gasser, L. (1991), ``Social concepts of knowledge and action: DAI foundation and open system semantics'', Artificial Intelligence, Vol. 47 No. 1-3, pp. 107-38. Georgeff, M.P. and Lansky, A.L. (1987), ``Reactive reasoning and planning'', Proc. AAAI-97, pp. 677-82. Hewitt, C. (1991), ``Open information systems semantics for distributed artificial intelligence'', Artificial Intelligence, Vol. 47 No. 1-3, pp. 79-106. Imam, I.F. (1996), ``Adaptive methodologies for intelligent agents'', in AAAI-96, Technical Report on Intelligent Adaptive Agents, AAAI Press, Portland, OR. Kahneman, D. and Tversky, A. (1972), ``Subjective probability: a judgment of representativeness'', Cognitive Psychology, Vol. 3, pp. 430-54. Moore, C., Baek, S., Liebowitz, J., Kilmer, R. and Minehart, R. (1998), ``Intelligent agent-based information warfare advisor (`Bob-in-a-box')'', Kybernetes, Vol. 27 No. 1, pp. 38-53. Payne, J.W. (1982), ``Contingent decision behavior'', Psychological Bulletin, Vol. 92 No. 2, pp. 382402. Payne, J.W., Bettman, J. and Johnson, E. (1988), ``Adaptive strategy selection in decision making'', Journal of Experimental Psychology, Vol. 14, pp. 534-52. Pollack, M.E. (1990), ``The uses of plans'', in Artificial Intelligence, Vol. 57, pp. 43-68. Rao, A. and Georgeff, M.P. (1993), ``A model-theoretic approach to the verification of situated reasoning systems'', Proc. of IJCAI-93, Chambery, France, pp. 318-24.
Intelligent agent behavior
175
Kybernetes 30,2
176
Rosenschein, J.S. and Zlotkin, G. (1994), Rules of Encounter ± Designing Conventions for Automated Negotiation Among Computers, MIT Press, Cambridge, MA. Russel, S. and Norvig, P. (1995), Artificial Intelligence: A Modern Approach, Prentice-Hall, Englewood Cliffs, NJ. Schwartz, D.G. and Te'eni, D. (1996), ``Imaginal agents'', Artificial Intelligence in Economics and Management, P. Ein-Dor (Ed.), Kluwer Academic Publishers, pp. 50-9. Schwartz, D.G. (1995), Cooperating Heterogeneous Systems, Kluwer Academic Publishers, Dordrecht. Sedbrook, T. (1994), ``Exploring dynamic group processes with GAIA ± groups of adaptive inferencing agents'', Kybernetes, Vol. 23 No. 5, pp. 12-26. Shoham, Y. (1993), ``Agent-oriented programming'', in Artificial Intelligence, Vol. 60, pp. 51-92. Srinivasan, A. and Te'eni, D. (1995), ``Modeling as constrained problem solving: an empirical study of the data modeling process'', Management Science, Vol. 41 No. 3, pp. 419-34. Star, S.L. (1989), ``The structure of ill-structured solutions: boundary objects and heterogeneous distributed problem solving'', in Gasser, L. and Huhns, M. (Eds), Distributed Artificial Intelligence Vol. II, Morgan Kaufmann, San Mateo, CA, pp. 37-54. Suchman, L.A. (1987), Plans and Situated Actions: The Problem of Human-machine Communication, Cambridge University Press, Cambridge. Sycara, K. and Lewis, C.M. (1991), ``Cooperation of heterogeneous agents through the formation of shared mental models'', Workshop on Cooperation among Heterogeneous Intelligent Agents, Pittsburgh, PA, July. Vroom, V.H. and Yetton, P. (1973), Leadership and Decision Making, University of Pittsburgh Press, PA. Werner, E. (1989), ``Cooperating agents: a unified theory of communication and social structure'', in Gasser, L. and Huhns, M. (Eds), Distributed Artificial Intelligence Vol. II, Morgan Kaufmann, San Mateo, CA, pp. 1-36. Appendix
(continued)
Intelligent agent behavior
177
(continued)
Kybernetes 30,2
178