An intelligent system is a system that has, similar to a living organism, a coherent set of components and subsystems working together to engage in goal-driven activities. In general, an intelligent system is able to sense and respond to the changing environment; gather and store information in its memory; learn from earlier experiences; adapt its behaviors to meet new challenges; and achieve its pre-determined or evolving objectives. The system may start with a set of predefined stimulus-response rules. Those rules may be revised and improved through learning. Anytime the system encounters a situation, it evaluates and selects the most appropriate rules from its memory to act upon.
Most human organizations such as nations, governments, universities, and business firms, can be considered as intelligent systems. In recent years, researchers have developed frameworks for building organizations around intelligence, as opposed to traditional approaches that focus on products, processes, or functions (e.g., Liang, 2002; Gupta and Sharma, 2004). Today’s organizations must go beyond traditional goals of efficiency and effectiveness; they need to have organizational intelligence in order to adapt and survive in a continuously changing environment (Liebowitz, 1999). The intelligent behaviors of those organizations include monitoring of operations, listening and responding to stakeholders, watching the markets, gathering and analyzing data, creating and disseminating knowledge, learning, and effective decision making.
Modeling intelligent systems has been a challenge for researchers. Intelligent systems, in particular, those involve multiple intelligent players, are complex systems where system dynamics does not follow clearly defined rules. Traditional system dynamics approaches or statistical modeling approaches rely on rather restrictive assumptions such as homogeneity of individuals in the system. Many complex systems have components or units which are also complex systems. This fact has significantly increased the difficulty of modeling intelligent systems. Agent-based modeling of complex systems such as ecological systems, stock market, and disaster recovery has recently garnered significant research interest from a wide spectrum of fields from politics, economics, sociology, mathematics, computer science, management, to information systems. Agent-based modeling is well suited for intelligent systems research as it offers a platform to study systems behavior based on individual actions and interactions. In the following, we present the concepts and illustrate how intelligent agents can be used in modeling intelligent systems.
We start with basic concepts of intelligent agents. Then we define agent-based modeling (ABM) and discuss strengths and weaknesses of ABM. The next section appliesABM to intelligent system modeling. We use an example of technology diffusion for illustration. Research issues and directions are discussed next, followed by conclusions.
Intelligent agents, also known as software agents, are computer applications that autonomously sense and respond to environment in the pursuit of certain designed objectives (Wooldridge and Jennings, 1995). Intelligent agents exhibit some level of intelligence. They can be used to assist the user in performing non-repetitive tasks, such as seeking information, shopping, scheduling, monitoring, control, negotiation, and bargaining.
Intelligent agents may come in various shapes and forms such as knowbots, softbots, taskbots, personal agents, shopbots, information agents, etc. No matter what shape or form they have, intelligent agents exhibit one or more of the following characteristics:
• Autonomous: Being able to exercise control over their own actions.
• Adaptive/Learning: Being able to learn and adapt to their external environment.
• Social: Being able to communicate, bargain, collaborate, and compete with other agents on behalf of their masters (users).
• Mobile: Being able to migrate themselves from one machine/system to another in a network, such as the Web.
• Goal-oriented: Being able to act in accordance with built-in goals and objectives.
• Communicative: Being able to communicate with people or other agents thought protocols such as agent communication language (ACL).
• Intelligent: Being able to exhibit intelligent behavior such as reasoning, generalizing, learning, dealing with uncertainty, using heuristics, and natural language processing.
Using intelligent agents and their actions and interactions in a given environment to simulate the complex dynamics of a system is referred to as agent-based modeling. ABM research is closely related to the research in complex systems, emergence, computational sociology, multi agent systems, evolutionary programming, and intelligent organizations. In ABM, system behavior results from individual behaviors and collective behaviors of the agents. Researchers ofABM are interested in how macro phenomena are emerging from micro level behaviors among a heterogeneous set of interacting agents (Holland, 1992). Every agent has its attributes and its behavior rules. When agents encounter in the agent society, each agent individually assesses the situation and makes decisions on the basis of its behavior rules. In general, individual agents do not have global awareness in the multi-agent system.
Agent-based modeling allows a researcher to set different parameters and behavior rules of individual agents. The modeler makes assumptions that are most relevant to the situation at hand, and then watches phenomena emerge from the interactions of the agents. Various hypotheses can be tested by changing agent parameters and rules. The emergent collective pattern of the agent society often leads to results that may not have been predicated.
One of the main advantages ofABM over traditional mathematical equation based modeling is the ability to model individual styles and attributes, rather than assuming homogeneity of the whole population. Traditional models based on analytical techniques often become intractable as the systems reach real-world level of complexity. ABM is particularly suitable for studying system dynamics that are generated from interactions of heterogeneous individuals. In recent years,ABM has been used in studying many real world systems, such as stock markets (Castiglione 2000), group selection (Pepper2000), and workflow and information diffusion (Neri 2004). Bonabeau (2002) presents a good summary of ABM methodology and the scenarios where ABM is appropriate.
ABM is, however, not immune from criticism. Per Bonabeau (2002), “an agent-based model will only be as accurate as the assumptions and data that went into it, but even approximate simulations can be very valuable”. It has also been observed that ABM relies on simplified models of rule-based human behavior that often fail to take into consideration the complexity of human cognition. Besides, it suffers from “unwrapping” problem as the solution is built into the program and thus prevents occurrence of new or unexpected events (Macy, 2002).
ABM FOR INTELLIGENT SYSTEMS
An intelligent system is a system that can sense and respond to its environment in pursuing its goals and objectives. It can learn and adapt based on past experience. Examples of intelligent systems include, but not limited to, the following: biological life such as human beings, artificial intelligence applications, robots, organizations, nations, projects, and social movements.
Walter Fritz (1997) suggests that the key components of an intelligent system include objectives, senses, concepts, growth of a concept, present situation, response rules, mental methods, selection, actions, reinforcement, memory and forgetting, sleeping, and patterns (high level concepts). It is apparent that traditional analytical modeling techniques are not able to model many of the components of intelligent systems, let alone the complete system dynamics. However, ABM lends itself well to such a task. All those components can be models as agents (albeit some in abstract sense). An intelligent system is thus made of inter-related and interactive agents. ABM is especially suitable for intelligent systems consist of a large number of heterogeneous participants, such as a human organization.
Agent-based modeling for intelligent systems starts with a thorough analysis of the intelligent systems. Since the system under consideration may exhibit complex behaviors, we need to identify one or a few key features to focus on. Given a scenario of the target intelligent system, we first establish a set of objectives that we aim to achieve via the simulation of the agent-based representation of the intelligent system. The objectives of the research can be expressed as a set of questions to which we seek answers (Doran, 2006).
A conceptual model is created to lay out the requirements for achieving the obj ectives. This includes defining the entities, such as agents, environment, resources, processes, and relationships. The conceptual modeling phase answers the question of what—what are needed. The design model determines how the requirements can be implemented, including defining the features and relevant behaviors of the agents (Brown, 2006).
Depending on the goals of a particular research, a model may involve the use of designed or empirically grounded agents. Designed agents are those endowed with characteristics and behaviors that represent conditions for testing specific hypotheses about the intelligent systems. When the agents are empirically grounded, they are used to represent real world entities, such as individuals or processes in an organization. Empirically grounded agents are feasible only when data about the real world entities are available. Similarly, the environment within which the agents act can be designed or empirically grounded. In practice, a study may start with simple models, often with designed agents and environments, to explore certain specific dynamics of the system.
The design model is refined through the calibration process, in which design parameters are modified to improve the desired characteristics of the model. The final step in the modeling process is validation where we check the agent individual behavior, interactions, and emergent properties of the system against expected design features. Validation usually involves comparison of model outcomes, often at the macro-level, with comparable outcomes in the real world (Midgley, el at., 2007). Figure 1 shows the complete modeling process. A general tutorial on ABM is given by Macal and North (2005).
ABM for Innovation Diffusion
We present an example of using agent-based intelligent system modeling for studying the acceptance and diffusion of innovative ideas or technology. Diffusion of innovation has been studied extensively over the last few decades (Rogers, 1995). However, traditional research in innovation diffusion has been grounded on case based analysis and analytical systems modeling (e.g., using differential and difference equations). Agent-based modeling for diffusion of innovation is relatively new. Our example is adopted from a model created by Michael Samuels (2007), implemented with a popular agent modeling system—NetLogo.
Figure 1. Agent-based modeling process
The objective of innovation diffusion modeling is to answer questions such as how an idea or technology is adopted in a population, how different people (e.g., innovators, early adopters, and change agents) influence each other, and under what condition an innovation will be accepted or rejected by the population. In the conceptual modeling, we identify various factors that influence an individual’s propensity for adopting the innovation. Those factors are broadly divided into to two categories: internal influences (e.g., word-of-mouth) and external influences (e.g. mass media). Any factor that exerts its influence through individual contact is considered internal influence.
Individuals in the target population are divided into four groups: adopter, potential (adopter), change agent, and disrupter. Adopters are those who have adopted the innovation, while potentials are those who have certain likelihood to adopt the innovation. Change agents are the champions of the innovation. They are very knowledgeable and enthusiastic about the innovation, and often play a critical role in facilitating its- diffusion. Disrupters are those who play an opposite role of change agents. They are against the current innovation, oftentimes because they favor an even newer and perceived better innovation. The four groups of agents and their relationships are depicted in Figure 2. It is common, although not necessary, to assume that those four groups make up the entire population.
Figure 2. Agents and influences
In a traditional diffusion model, such as the Bass model (Bass, 1996), the diffusion rate depends only on the number of adopters (and potential adopters, given fixed population size). Characteristics of individuals in the population are ignored. Even in those models where it is assumed that potential adopters have varying threshold for adopting an innovation (Abrahamson and Rosenkopf, 1997), the individuality is very limited. However, in agent-based modeling, the types of individuals and individual characteristics are essentially unbounded. For example, we can divide easily adopters into innovators, early adopters, and late adopters, etc. If necessary, various demographic and social-economic features can be bestowed to individual agents. Furthermore, both internal influence and external influence can be further attributed to more specific causes. For example, internal influence through social networks can be divided into traditional social networks that consists friends and acquaintances and virtual social networks formed online. Table 1 lists typical factors that affect the propensity of adopting an innovation.
An initial study of innovation diffusion, such as the one in Michael Samuels (2007), can simply aggregate all internal influences into “word-of-month” and all external influences into mass media. Each potential adopter’s tendency of converting to an adopter is influenced by chance encounter with other agents. If a potential adopter meets a change agent, who is an avid promoter of the innovation, he would become more knowledgeable about the advantages of the innovation, and more likely to adopt. An encounter with a disrupter creates the opposite effect, as a disrupter favors a different type of innovation.
Tablel. Typical internal and external influences
|Internal influence||External influence|
|Laws, policies and regulations|
|Social networks (online/ offline)||RSS|
In order for the simulated model to accurately reflect a real-world situation, the model structure and parameter values should be carefully selected. For example, we need to decide how much influence each encounter will result; what is the probability of encountering a change agent or a disrupter; how much influence is coming from the mass media, etc. We can get these values through surveys, statistical analysis of empirical data, or experiments specifically designed to elicit data from real world situations.
TRENDS AND RESEARCH ISSUES
As illustrated through the example of modeling the diffusion of innovation in an organization, industry, or society, agent-based modeling can be used to model the adaptation of intelligent systems that consist of intelligent individuals. As most intelligent systems are complex in both structure and system dynamics, traditional modeling tools that require too many unrealistic assumptions have become less effective in modeling intelligent systems. In recent years, agent-based modeling has found a wide spectrum of applications such as in business strategic solutions, supply chain management, stock markets, power economy, social evolution, military operations, security, and ecology (North and Macal, 2007). As ABM tools and resources become more accessible, research and applications of agent-based intelligent system modeling are expected to increase in the near future.
Some challenges remain, though. Using ABM to model intelligent systems is a research area that draws theories from other fields, such as economics, psychology, sociology, etc., but without its own well established theoretic foundation. ABM has four key assumptions (Macy and Willer, 2002): Agents act locally with little or no central authority; agents are interdependent; agents follow simple rules, and agents are adaptive. However, some of those assumptions may not be applicable to intelligent system modeling. Central authorities, or central authoritative information such as mass media in the innovation diffusion example, may play an important role in intelligent organizations. Not all agents are alike in an intelligent system. Some may be independent, non-adaptive, or following complex behavior rules.
ABM uses a “bottom-up” approach, creating emergent behaviors of an intelligent system through “actors” rather than “factors”. However, macro-level factors have direct impact on macro behaviors of the system. Macy and Willer (2002) suggest that bringing those macro-level factors back will make agent-based modeling more effective, especially in intelligent systems such as social organizations.
Recent intelligent systems research has developed the concept of integrating human and machine-based data, knowledge, and intelligence. Kirn (1996) postulates that the organization of the 21st century will involve artificial agents based system highly intertwined with human intelligence of the organization. Thus, a new challenge for agent-based intelligent system modeling is to develop models that account for interaction, aggregation, and coordination of intelligent agent and human agents. The ABM will represent not only the human players in an intelligent system, but also the intelligent agents that are developed in real-world applications in those systems.
Modeling intelligent systems involving multiple intelligent players has been difficult using traditional approaches. We have reviewed recent development in agent-based modeling and suggest agent-based modeling is well suited for studying intelligent systems, especially those systems with sophisticated and heterogeneous participants. Agent-based modeling allows us to model system behaviors based on the actions and interactions of individuals in the system. Although most ABM research focuses on local rules and behaviors, it is possible that we integrate global influences in the models. ABM represents a novel approach to model intelligent systems. Combined with traditional modeling approaches (for example, micro-level simulation as proposed in MoSeS), ABM offers researchers a promising tool to solve complex and practical problems and to broaden research endeavors (Wu, 2007).
Agent Based Modeling: Using intelligent agents and their actions and interactions in a given environment to simulate the complex dynamics of a system.
Diffusion of Innovation: Popularized by Everett Rogers, it is the study of the process by which an innovation is communicated and adopted over time among the members of a social system.
Intelligent Agent: An autonomous software program that is able to learn and adapt to its environment in order to perform certain tasks delegated to it by its master.
Intelligent System: A system that has a coherent set of components and subsystems working together to engage in goal-driven activities.
Intelligent System Modeling: The process of construction, calibration, and validation of models of intelligent systems.
Multi-Agent System: A distributed system with a group of intelligent agents that communicate, bargain, compete, and cooperate with other agents and the environment to achieve goals designated by their masters.
Organizational Intelligence: The ability of an organization to perceive, interpret, and select the most appropriate response to the environment in order to advance its goals.