When developing an agent model, a number of aspects need to be considered. They are: how to model agent behavior, agent-agent interactions, and the environment.
Like in individual-based modeling, agent behavior is typically modeled as a collection of heuristic, context-dependent rules that are iterated over one or more discrete time steps. Foraging behavior, for example, might be modeled as a random walk over the landscape, or as a more complex search routine in which the animal assesses its environment and moves in a deliberate fashion to seek out preferred food sources. One of the challenges of agent-based modeling is the elaboration of such rules. Empirical observations of animal
Box 2 What is the difference between individual-based and agent-based models?
There is considerable confusion in the ecosystem modeling literature concerning the difference between individual-based models (IBMs) and agent-based models. Individual-based modeling has a long tradition in ecological modeling and many authors use the terms individual-based and agent-based interchangeably. Adaptive agents are similar to individuals in IBM, with the exception that agents are provided with mechanisms by which they can adapt, learn, or evolve. In the agent literature, one generally differentiates between proactive and reactive agents. A proactive agent is motivated and goal oriented, whereas a reactive agent simply reacts to stimuli. Individuals in IBMs are usually more similar to reactive agents.
behavior can serve as a basis for rule generation. For example, a study of white-tailed deer foraging habits might show that they preferentially eat new shoots of a specific type of vegetation, and this preference could be incorporated into a rule. When the agents represent humans, modelers will typically interview subjects or distribute surveys as a means of compiling information. Another common approach is the use of role-playing games, in which subjects are asked to describe or act out how they would respond in a certain situation. These scenarios are then used to elaborate behavioral rules.
Agents are, by definition, goal oriented. They may seek to satisfy a basic need, such as meeting basic energetic requirements, or they may have a more complex goal, such as having a pleasant experience while visiting a wilderness area (in the case of human agents). Attaining this goal is often not straightforward, and may require several intermediate transactions with other agents (e.g., trade, combat/competition, or cooperation in the form of information or resource sharing). Many agent models simulate agent decision making via simple context-dependent rules as described above. Other approaches may use neural networks or other techniques in artificial intelligence and machine learning to model this aspect of agent behavior.
A common conceptual model that is applied in agent-based modeling is the beliefs-desires-intents (BDI) model. 'Beliefs' represent the background knowledge held by the agent (e.g., food is good to have); 'desires' represent its goals (e.g., get more food); and 'intents' represent the set of actions it intends to carry out to meet its goals (e.g., move toward the smell of food). Using a conceptual model such as BDI helps a modeler to organize the representation of an agent's goal-oriented behavior and to break down the decision-making process into parcels that are easier to deal with.
Adaptation, memory, and learning in agents may be simulated as part of an agent's behavioral and decision-making routines. Over longer time spans (e.g., multiple generations of agents), these characteristics are typically modeled using an approach based on genetic algorithms (see Evolutionary Algorithms). Genetic algorithms are binary (or numeric) sequences that represent an agent's 'genome'. The genome may contain parameters used to simulate agent behavior, including information about how the agent interacts with its environment and with other agents. When an agent reproduces, its genome is passed on to future generations. During reproduction, mutation (slight copy errors) may occur, mimicking the natural process of genetic mutation. Reproduction may be sexual or asexual. In the case of sexual reproduction, the offspring's genome is a combination of its parent's genomes.
A central tenet of agent-based modeling is the idea that group behavior is emergent from agent-to-agent interactions. Such interactions may be direct, via communication (message passing) or competition for a resource, or indirect, via modifications to a shared environment. Direct communication can be accomplished through a system that allows agents to compose and send messages that are subsequently interpreted by receiving agents. This type of communication is most often used in models where the agents represent humans. For example, in a model in which fisherman agents search for fish, agents may send messages to tell other agents the locations of good fishing sites. An agent may send a message specifically to one or more agents, or it may globally broadcast the message for all agents in the system to read using a blackboard system. Direct communication can also be modeled by adaptive agents that display 'tags' that can be decoded by the agents with which they come in contact. For example, an animal agent that is searching for a mate may display a special code that can be decoded by other animal agents in its local environment. This approach mimics the ecological mechanism of pattern display on animals. Indirect communication between agents often takes place via modifications to the environment. For example, in a model of ant foraging behavior, agents (ants) may leave pheromone trails that are subsequently followed and reinforced by other ant agents. In all of these examples, communication between agents reinforces and amplifies the effects of local interactions, giving rise to group-level dynamics.
Adaptive agents are locally situated in an environment, which may be represented in a number of different ways. Most commonly, the environment is modeled as either an »-dimensional lattice or an interaction network, although other representations are possible. The key is to constrain agent-to-agent interactions locally so as to model situations for which mean-field theory or other models based on assumptions of global mixing are inadequate.
In the case of an interaction network, each agent occupies a node in the network, and links between nodes designate avenues through which information or resources may be passed. An agent's location in the network thus determines the other agents with whom it may communicate. Interaction networks are often used to model communities of social agents. For example, the iterated prisoner's dilemma has been played on different network structures (each player is a node that plays with the nodes to which it is connected) in order to study the evolution of cooperation in social networks. Since networks are used to represent trophic interactions between species, a food web may also serve as an environment for an agent-based model in which agents represent species.
While network-based environments are useful for studying the flow of resources through social or physical systems, in agent-based ecological models, a spatially explicit environment is usually more pertinent due to the predominance of space in ecological interactions. In such spatially explicit models, each agent has a specific spatial location at a given time and is aware of its immediate surroundings, including the presence of other agents. The environment may be heterogeneous and can include passive objects as well as different types of agents. While there are many different ways that the spatial environment may be modeled, it is common to use a two-dimensional lattice. With this approach, each cell in the lattice may have a series of characteristics (e.g., resource availability) that can be accessed and modified by the agents. The lattice approach has a number of commonalities with cellular automata (see Cellular Automata). Models that seek to represent a more realistic environment may use a geographic information system (GIS) to provide initial conditions from cartographic data.
A widespread conclusion derived from simulations with agent-based models has been that an individual's environment has a substantial influence on its behavior and, subsequently, on the overall dynamics of the population. For this reason, most agent-based models contain a detailed representation of an environment that is variable, yet sufficiently regular for learning and adaptation to occur.
Was this article helpful?