Arguably, one can differentiate between five broad categories of simulation techniques:
There are situations where particles are considerably more complex and possess considerably more internal ``intelligence'' than in standard physics or engineering applications. Typical examples of such particles are humans, but one can also look at ants or adaptive traffic signals. Such particles can not be described by the methods (i)-(iv). A viable method is however to use rules to describe such particles. Rules can be easily encoded for computers, but they are difficult to treat using the methods of continuous mathematics. For that reason, MAS is a recent method.
MAS can be seen as a very recent merger of techniques from two different disciplines: (i) Complex Adaptive Systems (CAS), and (ii) Distributed Artificial Intelligence (DAI). Influences to CAS are cellular automata (4), neural networks (5), classifier systems/genetic algorithms (6), and ant colony optimization (7). interaction of many simple entities to produce complex behavior; in that sense, the CAS interpretation of a neural network would be that many simple neurons interact in order to produce complex emergent behavior of the neural network. One direction of CAS research has taken those complex entities composed of simple rules to again interact and thus form higher-level emergent complex adaptive systems. An example for this are stock market simulations using a different genetic algorithm for the simulation of each individual trading agent (8).
In contrast, Distributed Artificial Intelligence (DAI) has concentrated on the internal structure of individual agents. Recently, the interactions of several such agents has received much more attention (10,9). Somewhat simplifying, one can say that DAI concentrates on the interaction of few relatively complicated agents, while CAS concentrates on the interaction of many relatively simple agents. Therefore, MAS, as a combination of CAS and DAI, concentrates on the interaction of many, somewhat complicated agents.
MAS codes are best implemented using object-oriented programming languages, since those languages allow the clean encapsulation of internal object structure and object-object interaction. An early project in this direction is SWARM (11) based on ObjectiveC; a recent addition is RePast (12) based on Java. People interested in computational performance will also consider C++.
With respect to transportation simulations, aspects of MAS are slowly entering the field. Let us look at transportation planning. The traditional approach is the four-step process, consisting of Trip Generation, Trip Distribution, Mode Choice, and Route Assignment (e.g. (13,14)). When looking at each of the individual modules, one recognizes that they are supposed to describe human behavior, but they are formulated in terms of aggregated flows. For example, many mathematical formulations of the route assignment problem allow flows to be continuous variables (14).
There are however considerable efforts to put transportation simulation on a better behavioral basis (see (15)). These efforts include for example activity-based demand generation (e.g. (16,17)) and individual route-choice models (e.g. (18)). In addition, traffic simulations which allow one to follow individual travelers (e.g. (20,22,23,19,21)) make it now possible to couple those behavioral demand generation models with plausible traffic dynamics, allowing for logically consistent feedback. For example, it was difficult to consider departure time choice and the resulting activity scheduling problem with static route assignment models which cannot model peak spreading. High-performance implementations of those simulations (e.g. (25,26,24)) allow the simulation of very large scale scenarios, with 10 million or more travelers simultaneously in the system.
Looking further ahead, the agent-based representation is extremely amenable to methodological developments. For example, results from psychology (e.g. (27)), from experimental economics (e.g. (28), (29)), or from travel behavior research (e.g. (1)) can be directly used for such simulations, without the need for further transformations into the variables used by the simulation system. In fact, we are currently in the process of implementing activity-based demand generation within the agent-based framework.