Evolving Individual Behavior in a Multi-Agent Traffic Simulator Ernesto Sanchez 1 , Giovanni Squillero 1 , Alberto Tonda 1 1 Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129, Torino, Italy {ernesto.sanchez, giovanni.squillero, alberto.tonda}@polito.it Abstract. In this paper, we illustrate the use of evolutionary agents in a multi- agent system designed to describe the behavior of car drivers. Each agent has the selfish objective to reach its destination in the shortest time possible, and a preference in terms of paths to take, based on the presence of other agents and on the width of the roads. Those parameters are changed with an evolutionary strategy, to mimic the adaptation of a human driver to different traffic conditions. The system proposed is then tested by giving the agents the ability to perceive the presence of other agents in a given radius. Experimental results show that knowing the position of all the car drivers in the map leads the agents to obtain a better performance, thanks to the evolution of their behavior. Even the system as a whole gains some benefits from the evolution of the agents’ individual choices. Keywords: Multi-agent systems, Evolution, Traffic simulation 1 Introduction Road traffic congestion is a crucial problem, the short-range consequences of which can vary from delays to decreased throughput of vehicles. Long-range consequences include reduced safety, environmental pollution, and reduced economic competitiveness. This problem is becoming more intense, not only in western cities but also in countries where the presence of cars, once scarce, is growing at an alarming rate. From websites displaying the current traffic conditions [1] to collections of traffic control strategies [2] available online, information technology is playing a vital role in the development of new approaches to traffic control, even by simply providing the means to evaluate innovative methodologies, by means of sensors, databases and data mining. In this context, simulations are heavily employed to test the possible outcome of new strategies, and often multi-agent systems are chosen as a simulation tool. An agent is defined as a complex software entity that is capable of acting with a certain degree of autonomy in order to accomplish tasks. A multi-agent system (MAS) is a collection of software agents that work in conjunction with each other. They may cooperate or they may compete, or some combination of the two, but there is some common infrastructure that result in the collection being a ‘system’, as opposed to simply being a disjoint set of autonomous agents [3]. Each agent in the MAS tries to achieve some individual or collective task.