Methodological Issues of Agent-Based Modeling in Social Sciences

Elisa Vuori

Tampere University of Technology

elisa.vuori@tut.fi

Abstract

Agent-based modeling is used in many areas of science. Agent-based modeling has originated in AI, and its initial purpose of use differs from many applications it has nowadays. This shift in the usage of agent-based thinking causes a discontinuation point in the philosophical background of the method. This paper discusses two issues arising from this setting: (2) the history of agent-based modeling and (1) the relation of agent-based modeling to other scientific methods of creating new knowledge, and its tasks and aims. This discussion includes the pondering of the research questions that can be answered by agent-based modeling and the nature of those answers. The study is carried out as a literature study.

Contact:

M.Sc. (Eng.) Elisa Vuori

Institute of Business Information Management

Tampere University of Technology

Tampere, FI-33101, Finland

Tel: +358 3 3115 3732

Fax: +358 3 3115 4680

Email: elisa.vuori@tut.fi

Key Words: Agent-based modeling, agent, artificial intelligence, scientific method, social science

Acknowledgement: The work for this paper has been supervised by the director of the research programme, Dr. Marjatta Maula.

Support: The author of this paper is a researcher in TIP (Knowledge and Information Management in Knowledge-Intensive Services) Research Program, and this paper is part of her research on organization populations containing knowledge-intensive services and the agent-based modeling of them. TIP research program covers topics of organizational research, knowledge management inside and outside of an organization, and innovation studies. The topics are approached from the field of complexity science, aiming at understanding the social systems as complex, evolving systems. The research program takes place at Tampere University of Technology, Institute of Business Information Management and it is funded mainly by the National Technology Agency of Finland.

Methodological Issues of Agent-Based Modeling in Social Sciences

Elisa Vuori

Introduction

Agent-based modeling has increased its popularity in the area of social sciences. The origin of agent-based modeling is in the development of artificial intelligence, and the Complexity science adopted it quickly. Nowadays agent-based modeling is exploited in various areas of scientific research and even in enterprise management. Agent-based modeling is related to agent-based software engineering, because both of them have originated from the same research area of artificial intelligence. Agent-based modeling and software engineering are also intertwined. They have mutual concepts, and their development is partially interdependent because they utilize each other’s results. Both of them are interesting also from the point of view of complexity science. This paper concentrates on agent-based modeling and its application to scientific research.

Since the agent approach was first adopted in the research area of artificial intelligence, it has experienced up- and downhill in popularity. There has been a clear reason for this oscillation: the expectations of the possibilities of agent-based approach have been too high for the reality to live up to them. Now the expectations are high in the area of social sciences, which gave the impulse to write this paper. Understanding right the nature of agent-based modeling as a scientific research method may prevent some disappointments. The utility of agent-based modeling and simulation for different research goals, and methodological assumptions are not clear [David, Marietto, Sichman & Helder, 2004]. The clarification of them requires a study of the history and principles of agent-based modeling and also a survey on scientific methodology. The questions to which this paper seeks to answer are: - What is the relationship between artificial intelligence, agent-based software development and agent-based modeling?

- What is the position of agent-based modeling in scientific methodology?

In the second chapter of this paper the agent-based paradigm is approached from the point of view of history and by discussing two branches of it, namely agent-based software engineering and agent-based modeling. The similarities and dissimilarities of these two braches are covered. Also the connection to complexity science is noted. In the third chapter we turn to scientific methodology and the position of agent-based modeling and simulation in it. Chapter four concludes.

The Agent-Based Paradigm

As any human innovation, agent-based thinking did not develop in an empty space. According to Jennings, Sycara and Woolridge, there had been related discussion in several areas, but the most important fields of research for the birth of agent-based thinking were artificial intelligence, object-oriented programming and human-computer interface design [Jennings, Sycara & Woolridge, 1998]. However, artificial intelligence was the most significant contributor.

The emergence of agent-based design in artificial intelligence took place in 1985. It was named “autonomous agent research”, “behavior-based research”, “bottom-up AI” or “animat approach” (animat = ’artificial animal’[Kang, Waisel & Wallace, 1998]. [Maes 1994] These names all describe the characteristics of new kind of AI. It is an approach, where systems are designed bottom-up, and thus the ‘autonomous’ components, namely agents, act according to their internal schema instead of external commands. This produces emergent behavior, which to some extent resembles life, and thus is closer to natural intelligence than the behavior the previous approaches produce. In this paper the term autonomous agent research is used to refer to the agent-based design in AI.

Figure 1 describes how this autonomous agent research has produced two new fields of research, namely agent-based modeling and agent-based software design. Both of these fields hold the same design principles as autonomous agent research did. They base on the bottom-up approach and assume the acting agents to be autonomous to some extent. The difference between them exists in the research purposes. Agent-based modeling is a modeling method, and is often exploited to build a model and run simulations on it to study certain phenomena. Agent-based software design utilizes the agent approach to design decentralized software, and autonomous agent research aims at imitating natural intelligence and in some places, replacing it. However, the borderlines between these three research areas are blurred.

Agent-based software design

“Objects do it for free; agents do it for money.”

[Jennings, Sycara & Woolridge, 1998] In software design, autonomous agents and especially multi-agent systems are a new and intriguing way to analyze, design and implement complex software systems [Jennings, Sycara & Woolridge, 1998]. It is a paradigm, which suits to various kinds of software systems. The important aspects of these systems are complexity and decentralization. It is slightly difficult to find a clear difference between agent-based software design and more general object-oriented design. This difference has been discussed by many authors. Franklin and Graesser conclude that “An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future” [Franklin & Graesser, 1996]. This definition separates autonomous agents from other pieces of code, because an object does not have e.g. an own agenda. It is also stated that agents differ from other programs because they are autonomous, flexible, are considered to have their own thread of control and exhibit a control over their behavior [Jennings, Sycara & Woolridge, 1998]. These definitions of an agent may also be contradicted. Steels states that an agent has got two aims to pursuit: to perform its own function as a part of a larger system and to maintain its own viability [Steels, 1995]. From his point of view, software agents should not be called agents, because they do not have any self-interest at all [Steels, 1995]. In strict sense this may be true, but a software agent may have a built-in selfishness, which ensures that the agents try to maintain themselves and act in the most efficient way to attain its objectives.

One important advantage of agent-based software engineering is the greater robustness of a system under design. This robustness arises from the equality and adaptability of agents, which causes that the loss of an agent does not seriously affect the workability of a system. Agent-based design has been utilized e.g. in the areas of manufacturing, process control, telecommunications, air traffic control, transportation systems, information management, electronic commerce, business process management, games, interactive theatre and cinema, patient monitoring and health care [Jennings, Sycara & Woolridge, 1998]. As we see, all these systems contain complex interactions and many of them require high robustness. They are also research areas, which are already and will be important in the future.

Agent-based modeling

Agent-based modeling does not aim at implementing a system, which would carry out some practical tasks. Rather it is a tool for creating artificial worlds and systems, which may replicate real world systems, or be purely virtual ones. This modeling may help the design of real systems or it may be a tool for understanding them. Agent-based modeling has also been used to create artificial life, which does not directly imitate any form of life on earth, but fulfills the definition of life [Casti, 1997]. The range of possible systems to be modeled has a very wide range, from the anthills to human societies and everything between them.

Agent-based system is one in which the key abstraction used is that of an agent, and thus it does not provide that the system should be implemented in software [Jennings, Sycara & Woolridge, 1998]. Constructing an agent-based model may be useful intrinsically, and there is no imperative to run simulations on it, not by computers or otherwise. However, in many cases the greatest benefit of agent-based modeling is the possibility to simulate the behavior of the system under study. Utilization of simulation is especially useful, if it is cheaper and quicker than testing on a real system [Kang, Waisel & Wallace, 1998]. A computational model is sensible also in situations where real data is not available, it is very expensive to get or testing might be hazardous [Kang, Waisel & Wallace, 1998]. Anyways, a model can never substitute the study of real-world dynamics, but it is to give clues what could happen also in reality [Huberman & Glance, 1998].

As said in the beginning of this paper, agent-based modeling has been used in various areas of research. Examples of model categories on the area of social science are artificial societies, socio-cognitive, socio-concrete and prototyping for resolution [David, Marietto, Sichman & Helder, 2004]. Artificial societies may resemble some real systems, but they may also be purely abstract. Socio-cognitive models are intended to model socio-cognitive or sociological theories and thus test them for refinement, extension and verification. A socio-concrete model describes an observed social system in order to provide insight to the behavior of the system. Prototyping for resolution utilizes agent-based modeling for the design of agent-based software systems. [David, Marietto, Sichman & Helder, 2004]

Agent-based design and complexity

Deguchi states that there have been three revolutions in the research of complex systems, and agent-based modeling will be the fourth breakthrough [Deguchi, 2004]. Thus the development of complexity science and that of agent-based modeling are deeply intertwined: the study of complex systems needs agent-based modeling, and the use of agent-based modeling and design is understandable through complexity science. This is visible also in the theoretical basis of many researches, which utilize agent-based modeling. The ideas of complexity are implicitly or explicitly present, even if the system and phenomena under study are from the areas of social science or economics.

Agent-based systems are often studied as wholes and they are considered to be more that the sum of their parts, which is true for all complex systems. One important feature of complex systems is that they do have emergent properties. Emergent property is one, which can not be anticipated on the basis of the interaction between agents. Thus, like the ability to form thoughts is an emergent property of the interaction between different brain cells and consciousness is an emergent property of the interaction of thoughts, emergent properties take place on the system level and are not a property of a single agent. The existence of emergent properties is the motivation also for agent-based modeling and simulation. Often agent-based modeling is aimed at revealing some emergent properties and the process leading to them. The simulation of interaction between agents is often “surprising, because it can be hard to anticipate the full consequences of even simple forms of interaction.” [Axelrod, 1997]

Issues of Scientific Methodology Concerning Agent-Based Modeling and Simulation This chapter reviews quickly those aspects of scientific research methods that raise questions from the point of view of agent-based modeling. First, the relation of simulation to the more general division of research strategies is discussed. That is followed by consideration if agent-based simulation is qualitative or quantitative by nature and if the results of a research utilizing simulation are prone to be descriptive or normative. The logic of reasoning on the basis of an agent-based simulation is contemplated. Finally, the implications of agent-based modeling to verifying and falsifying are discussed.

Simulation is considered to be one kind of a controlled experiment. According to J?rvinen and J?rvinen [2004], controlled experiment is a technique, which is applicable when testing a theory. In the division of the research strategies simulation is basically a reality-related, empirical technique, which tries to answer to the question “what is reality like?” [J?rvinen & J?rvinen 2004]. A simulation may be carried out to test a theory. This is the case if it is possible to construct a model, which represents the assumptions made in theory and if the results of the simulation allow assessing the relevance of those assumptions. However, testing a theory by an agent-based simulation is not as simple as it is by simulation of linear dynamics. The problem is that it may be difficult to demonstrate the cause and the effect in a system, which exhibits non-linear behavior as complex systems do. If we consider agent-based modeling as a tool for a researcher, it is also used for exploration of the system. This may be the case in a situation, where the laws underlying the system under study are not clear, and the studied phenomenon is produced in a simulation by trial and error. Thus the process may be iterative, and the result is not a tested theory but rather a set of assumptions, which may be tested by another method. Thus agent-based modeling and simulation may be suitable for wider variety of tasks than more traditional simulation methods.

Research methods may be divided in many different ways. Two bases of division are treated here, namely the partition to quantitative and qualitative and to descriptive and normative. A coarse division between quantitative and qualitative is to call quantitative study one concerning numbers, where qualitative study handles attributes. In a study, the method for collecting data may be either quantitative or qualitative, and also the method for handling the data gathered may be either quantitative or qualitative. That is, one may gather data through an interview (qualitative) and analyze it by a statistical method (quantitative). Purely quantitative or qualitative studies are infrequent. However, some methods are also difficult to categorize to either qualitative or quantitative, because it may be used for both kind of studies or contain both aspects. Thus the relevance of this basis of division has been sometimes questioned.

Controlled experiments may be either quantitative or qualitative. Simulations are usually considered to be of quantitative nature. This is because running a simulation requires model construction, which means that a conceptual model is translated to the language of computation. However, agent-based models are often implemented on software, and the programming language is something between natural language and pure equations. An agent-based model does not necessarily have any global equations but only local rules and laws, the internal schema, which an agent follows. Thus modeling is neither clearly quantitative, but nor is it a qualitative method.

The division between descriptive and normative research takes a stand on the result of a study. Normative results are those who set norms based on the research and thus their results take a stand for or against something. Normative studies may suggest a certain change on the basis of the study. Descriptive results aim at revealing something previously unknown of the phenomenon under study by describing it. Descriptive studies try to create understanding. This understanding may promote changes, but it is not a direct purpose of the study.

Traditional ways to derive scientific conclusions are considered to be induction and deduction. Induction involves the search for general laws or patterns from a set of observations. Deduction, instead, tries to verify or falsify a certain set of assumptions by examining the system under study, and finding phenomena, which are possible to be justified as a result of the laws previously assumed. Axelrod claims that agent-based modeling is a third way of carry out scientific reasoning [Axelrod, 1997]. He states that agent-based modeling starts with similar assumptions as deduction, but instead of verification or falsification it aims at providing data, which can still be used for induction. Pátkai [2004] classifies human knowledge to two main branches, which are scientific knowledge and rhetorical knowledge (figure 3). According to him, the scientific knowledge is considered to be based on definitions, axioms, theorems and deduction. This suggests that scientific knowledge refers particularly to the knowledge attained by natural science. Rhetorical knowledge is considered to be based on argumentation and analogies. Pátkai sees that computation, which agent-based modeling is a part of, is like a third language besides the scientific knowledge and the rhetorical knowledge. Computation would be like an extension to the scientific knowledge. If we consider agent-based modeling as a method, it is often based on the use of analogy. An agent-based model is a construction, which is analogous to a real-world system. It does not imitate real world meticulously, but it is considered to have characteristics, that make it analogous to the real-world system under study. If the use of analogy is not accepted as a way to make scientific reasoning, there is no way to link the agent-based model to the real-world system, which it is considered to present.

Verifying and falsifying theories and laws is a necessary part of increasing scientific knowledge. Induction and deduction are established ways to prove a theory. What is the role of analogy or agent-based modeling in this assignment of scientific research? Agent-based modeling is noted to be “a way of doing thought experiments” [Axelrod 1997]. The strength of agent-based modeling and simulation is the possibility to test different settings and gain understanding on the system. It is not an established way to prove theories or laws, and thus it should not be used for that purpose without acceptable justification. Luckily, a computer simulation allows a researcher to examine also the intermediate states and not only the results. Thus, a simulation may help to prove a cause and effect relation, if the process can be made explicit step by step. Even if the logic behind a certain emergent phenomenon can be demonstrated by a simulation, one have to remember that it is still just an analogy of the real system. Thereby the ability of agent-based simulation to prove a law or a theorem is not very strong.

Conclusions

Agent-based modeling as a research method should be studied on its own right. This does not mean that agent-based modeling would not be related to the previous scientific methods. Agent-based modeling and simulation differ to some extent from traditional simulation. It performs multiplier functions than traditional simulation, but it also holds less validation power than traditional simulation. Agent-based modeling is used to test theories, but also to suggest new theories.

It has been suggested, that agent-based modeling is a new way to create scientific knowledge or to prove theorems. A new way to create knowledge it may be, but a tool for proving theorems it may not be. The propositions created by agent-base modeling must be verified by other means. The research utilizing agent-based simulation is likely to be descriptive by nature. This is because simulation is used for understanding and providing ideas of the dynamics of the system under study and not for validation. Thus, on the basis of pure observations it is difficult to set normative rules.

The use of agent-based modeling and simulation as a research tool for a real system implicitly holds the assumption of existing analogy between the reality and the simulation. This assumption means that the researcher supposes that the system under study and the model build of it resemble each other in relevant aspects. Further, this assumption contains the idea that the emergent properties in those two systems are similar.

Agent-based modeling and simulation are good tools when used properly, remembering that the results obtained from a simulation are bound by certain restrictions. These restrictions are the bounded relevance of the use of analogy, and the difficulty to prove the cause and effect relations in non-linear systems.

References

[Axelrod 1997] Axelrod R., 1997, The Complexity of Cooperation: Agent-Based Models of Competition and Cooperation. Princeton University Press, New Jersey.

[Casti 1997] Casti, J. L., 1997, Would-Be Worlds. John Wiley & Sons. New York.

[David, Marietto, Sichman & Helder, 2004] David N., Marietto M. B., Sichman, J. S. & Helder, C., 2004, The Structure and Logic of Interdisclipinary Research in Agent-Based Social Simulation. Journal of Artificial Societies and Social Simulation, Vol 7(3).

[Deguchi 2004] Deguchi, H., 2004, Economics as an Agent-Based Complex System. Springer-Verlag Tokio.

[Franklin & Graesser 1996] Franklin, S. & Graesser, A., 1996, Is It an Agent or Just a Program?: A Taxonomy for Autonomous Agents. Proceedings of the Third International Workshop on Agent Theories, Architectures, and Languages. Springer-Verlag.

[Huberman & Glance, 1998] Huberman B. A. & Glance, N. S., 1998, Fluctuating Efforts and Sustainable Cooperation. In Prietula, M. J., Carley, K. M. & Gasser, L. (Eds.) Simulating Organizations: Computational Models of Institutions and Groups. American Association for Artificial Intelligence: 89-103.

[Jennings, Sycara & Woolridge 1998] Jennings, N. R., Sycara, K. & Wooldridge, M., 1998, A Roadmap of Agent Research and Development. Autonomous Agents and Multi-Agent Systems, 1. Boston, Kluwer Academic Publishers: 7-38.

[J?rvinen & J?rvinen, 2004] J?rvinen, P. & J?rvinen, A., 2004, Tutkimusty?n metodeista. Opinpajan Kirja, Tampere.

[Kang, Waisel & Wallace, 1998] Kang, M., Waisel, L. B. & Wallace, W. A., 1998, Team-Soar: A Model for Team Decision Making. In Prietula, M. J., Carley, K. M. & Gasser, L. (Eds.) Simulating Organizations: Computa-tional Models of Institutions and Groups. American Association for Artificial Intelligence: 23-45.

[Maes 1994] Maes, P., 1994, Modeling Adaptive Autonomous Agents. Journal of Artificial Life. Vol. 1(1&2): 135-162.

[Pátkai 2004] Pátkai, B., 2004, An Integrated Methodology for Modelling Complex Adaptive Production Networks. Doctoral dissertation. Tampere University of Technology.

[Steels 1995] Steels, L., 1995, When are Robots Intelligent Autonomous Agents? Robotics and Autonomous Systems, Vol 15, pp. 3-9.