Agent-Based IS

This section points to selected resources for research and teaching on agent-based information systems. We invite you to forward any material of interest and relevance, such as Web links, so that it can be added appropriately. Please send material by e-mail to our Webmaster. The categories of teaching tools and e-bibliography are still under construction.

Please notice that research papers and Web sites that fit with our Research Areas are listed separately. Every research project that is described in the Research Area section concludes with a paragraph on project-specific references. They are sorted in alphabetic order.

Simulation
Simulation in general is understood as “a technique for understanding and predicting the behavior of systems” (Simon 1996, 13). At its core, it refers to the creation of the artificial (as opposed to the natural) things that are synthesized by human beings, may imitate appearances in nature, and can be characterized in terms of functions, goals, and adaptation.

Benefits
The use of simulation experiments has been viewed as a promising approach (Courtney et al. 1997, 71, 78), particularly, because conventional analytic tools fail to solve even moderately complex economic models (Chaturvedi and Mehta 1999, Rust 1996). Most mathematical and statistical tools, for example, rely on the assumption of linearity (a function is linear if its value is a weighted sum of the values of its arguments: .)
Unfortunately, many phenomena are of a non-linear nature.

Often, the behavior of a whole system is more than a simple sum of the behavior of its parts. Organizations, for example, are considered to be complex, dynamic, non-linear, adaptive, and evolving systems (Prietula et al. 1998). Tools - such as trend analysis or the calculation of equilibria - have proven very useful for generalizing observations of linear phenomena into theory. However, when applied to nonlinear problems, these tools may produce approximations that fail to provide explanations.

Although computer-based simulation is no more valuable than its underlying assumptions, significant benefits can be realized by its use (Simon 1996, 15–16):

Even in poorly understood environments, simulation yields beneficial insights (16–17).

While the aggregate outcome of a phenomenon such as the state of a nation's economy appears to be difficult to explain or to predict, this is not necessarily the case for the behavior of individual components, such as firms and consumers. Therefore, creating a simulation system composed of sufficiently understood building blocks and relationships appears to be a promising research strategy to analyze complex, dynamic or emergent structures.

The use of multi-agent modeling and simulation extends well beyond the academic community. The U.S. Department of Justice, for example, has replaced structural analysis with simulation to analyze mergers in differentiated product industries. Business executives use it to complement conventional analytic tools to make better strategic decisions (Courtney et al. 1997). Simulation and computational experiments have been particularly helpful in areas that require an understanding of the dynamics of change and competitive interaction. This is typically the case where executives have to respond to deregulation, disruptive technologies, or market entry. Simulation-based analysis can allow decision makers to "play-out" different scenarios and study implications for firm profits (local performance) and industry profitability (global performance). Furthermore, once a simulation system has been designed and implemented it can be reused and modified to evolve with a changing competitive landscape and thus provide ongoing decision support.

Chaturvedi, A. R., and S. R. Mehta. 1999. Simulation in Economics and Management. Communications of the ACM 42(3) (March): 60-61.

Courtney, H., J. Kirkland, and P. Viguerie. 1997. Strategy Under Uncertainty. Harvard Business Review (November-December): 67-79.

Prietula, M. J., K. M. Carley, and L. Gasser. 1998. A Computational Approach to Organizations and Organizing. In: Prietula, M. J., K. M. Carley, and L. Gasser (eds). Simulating Organizations: Computational Models of Institutions and Groups. The MIT Press: Cambridge, MA: xiii-xix.

Rust, J. 1996. Dealing with the complexity of economic calculations. July 31-August 3 Workshop on Fundamental Limits to Knowledge in Economics held at the Santa Fe Institute, Santa Fe, NM.

Simon, H. A. 1996. The Sciences of the Artificial. 3rd ed. MIT Press: Cambridge, MA.

In computational organization theory (COT) and the information systems and artificial intelligence research literature, an agent is understood as a representation of a decision unit or, in general, a knowledge processor and (when implemented) a software module. A recent overview of the computational modeling of organizations is provided in Prietula et al. (1998). An agent is created to perform a task or set of tasks, and it features some combination of the following selected properties:

Properties are defined in terms of a particular task and its environment, which influences an agent and is affected by agent behavior, as well. Depending upon the combination of tasks and the environment, the literature distinguishes between various categories of agents. An overview is provided by Wooldridge and Jennings (1995).

A multi-agent system "allows for the different modules, objects, or processes--i.e., agents--to maintain their local control by treating them as autonomous agents, while at the same time providing them with a means for achieving the desired coordination" (Sikora and Shaw 1998, 72).

A complex adaptive system is a multi-agent system with a particular set of mechanisms and properties in which "a major part of the environment of any given adaptive agent consists of other adaptive agents, so that a portion of any agent's efforts at adaptation is spent adapting to other adaptive agents" (Holland 1995, 10). Holland provided a detailed framework for CAS design that includes three mechanisms and four properties common to all CASs (1995, 10-40):

For demonstration purposes, a building block mechanism and an aggregation property, as well as their interrelations, are described below and illustrated using documentation of a CAS for organizational analysis.

The CAS design mechanism of building blocks refers to the decomposition of a complex scene into a few, distinct categories of components and relationships (Holland 1995, 34-37). Building blocks can be reused and combined to create relevant, perpetually novel scenes. With decomposition and the repeated use of building blocks, novelty arises through combination. Even with only a few sets of categories of components or building blocks and rules for combining them, an exponentially large number of different configurations can be assembled. Consider the following example: With only two different types of agents, such as firms (for example manufacturers and distributors), and ten variations of each type, such as ten different degrees of vertical integration, as many as 2^10 or 1024 different combinations or industry scenarios could be created. Hewitt stated, "Our hypothesis is that by organizing large-scale systems into composable units, each of which provides support for operations, membership liaison, accounting, and management, we will be able to construct large-scale systems that are more scalable, robust, and manageable" (1991, 99-100).
The challenge with building blocks is the decomposition of a complex scene into as few relevant categories of components and rules for combining them as possible. If this can be achieved, building blocks can provide great scalability and efficiency with reconfiguration.

Actual building blocks can be perceived as agents with their embedded functions or methods, while relationships can be considered as dependencies in the context of multi-agent systems (Sikora and Shaw 1998).

Building blocks become even more powerful when applied to aggregated or tiered designs. Aggregation “concerns the emergence of complex, large-scale behaviors from the aggregate interactions of less complex agents” (Holland 1995, 11). This design property allows for patterns of higher-level events to derive from the settings of lower-level building blocks. A higher-level event, such as a change in industry structure, can be observed emerging from lower-level conditions and dynamics of interaction.

Aggregation corresponds to the concept of sub-agents in Sikora and Shaw’s multi-agent system framework (1998).

Hewitt, C. 1991. Open Information Systems Semantics for Distributed Artificial Intelligence. Artificial Intelligence 47: 79-106.

Holland, J. H. 1995. Hidden Order: How Adaptation Builds Complexity. Helix, Addison–Wesley: Reading, MA.

Malone, T. W. 1987. Modeling Coordination in Organizations and Markets. Management Science 33(10) (October ): 1317-1332.

Malone, T. W., and Stephen A. Smith. 1988. Modeling the Performance of Organizational Structures. Operations Research Vol. 36, No. 3 (May-June): 421-436.

Sikora, R., and M. J. Shaw. 1998. A Multi-Agent Framework for the Coordination and Integration of Information Systems. Management Science 44(11) (November): 65-78.

Swaminathan, J. M., S. F. Smith, and N. M. Sadeh. 1998. Modeling Supply Chain Dynamics: A Multiagent Approach. Decision Sciences 29(3) (Summer): 607-632.

Wooldridge, M. J., and N. R. Jennings. 1995. Agent theories, architectures, and languages—A survey. In: ECAI-94 Workshop on Agent Theory, Architectures, and Languages. Springer: Heidelberg, Germany, 1–21.

One important issue with scientific research is rigor. In choosing laboratory experiments as a research method and simulation as a research instrument, the problems with validity of research are twofold.
First, simulation is a new instrument, and therefore it should be exposed to an explicit test of instrument validity (i.e., Boudreau et al. 2001, 12; Carley 1995; Carley and Prietula 1994). One approach that is suggested in the literature is the use of a framework of validity of computational models in organization sciences (Burton and Obel 1995). Furthermore, validation can be complemented by docking (Burton 1998), the alignment of simulation models (Axtell et al. 1996). Overall, validation and docking could best be supported by basing the architecture of a simulation application on previously validated frameworks.
However,--and this is the second problem--architectural compliance with theories and previously validated frameworks is limited by the means of implementation available. Without appropriate software capabilities, certain theoretically required elements and conditions may be impossible to implement.

Lately, the development of object-oriented programming has greatly enhanced the capabilities of simulation tool kits and has therefore provided an opportunity to implement architectures that can comply with validated frameworks.

Object-orientation has two primary objectives, which fit well with the properties of multi-agent system frameworks: (1) the development of reusable components and (2) the development of implementations with greater flexibility for adaptation. The latter advance corresponds well with the requirements of CAS. In order to achieve these objectives, object-oriented languages and tools support the following three key concepts:

One simulation system that takes advantage of object-orientation is the Swarm system, a software package originally developed at the Santa Fe Institute. Swarm is written in Objective C, an object-oriented derivative of the C++ programming language, and it has been specifically designed for multi-agent simulation of CAS (Minar et al. 1996). It allows, for example, simulations of collections of concurrently interacting agents. The fit between MAS/CAS modeling requirements and Swarm implementation capabilities has been confirmed in an exploratory study by Lin et al. (1999). The figure below provides only a high-level overview.

In choosing simulation-based laboratory experiments, researchers have to be concerned with software development life-cycle (SDLC) issues. The development of a software application is typically iterative and can be especially time consuming if the system requirements from the MAS model are evolving rather than fixed. The following figure illustrates the sequence of MAS development activities and loops of iterations.

Axtell, R., R. Axelrod, J. Epstein, and M. Cohen. 1996. Aligning Simulation Models: A Case Study and Results. Computational and Mathematical Organization Theory 1(2): 123-141.

Boudreau, M. C., D. Gefen, and D. W. Straub. 2001. Validation in Information Systems Research: A State-of-the-Art Assessment. MIS Quarterly 25(1): 2-16.

Burton, R. 1998. Validating and Docking. In: Prietula, M. J., K. M. Carley, and L. Gasser (eds). Simulating Organizations: Computational Models of Institutions and Groups. The MIT Press: Cambridge, MA: 215-228.

Burton, R. M., and B. Obel. 1995. The Validity of Computational Models in Organization Science: From Model Realism to Purpose of the Model. Computational and Mathematical Organization Theory 1(1): 57-71.

Carley, K. M. 1995. Computational and Mathematical Organization Theory: Perspective and Directions. Computational and Mathematical Organization Theory 1(1): 39-56.

Carley, K. M., and M. J. Prietula (eds). 1994. Computational Organization Theory. Lawrence Erlbaum Associates: Hillsdale, NJ: 1-18.

Lin, F.-R., G. W. Tan, and M. J. Shaw. 1999. Multiagent Enterprise Modeling. Journal of Organizational Computing and Electronic Commerce 9 (1): 7-32.

Minar, N., R. Burkhart, C. Langton, and M. Askenazi. 1996. The Swarm simulation system: A toolkit for building multi-agent simulations. Technical Report 96-04-2, Santa Fe Institute, Santa Fe, NM.