Automatic Behavior Generation in a Multi-Agent System through Evolutionary Programming

The purpose of this project is to develop a system that is able to generate strategies for the multi-agent system proposed by the RoboCup Four-Legged league. This system starts with different soccer match situations established by the user, and uses evolutionary programming as a means to make the process automatic. Given the nature of soccer, it is easy to see that the different situations that make up this game can have radically different local objectives, even if the ultimate goal of the game is one and the same: to score more goals than your opponent. This is why the system will focus on generating strategies for very specific scenarios (for example, the behavior a goalkeeper must take when it is kicking off and it has an opponent player in front of him), that will both allow us to adjust the fitness function as much as possible, and to generate state machines that are as specialized and optimized as they can be for the situation they focus on. The aforementioned system runs over a simulator developed by Vega et al. (2006), which takes behaviors defined in XML as state machines in order to define players from these data. This gives us both the advantage of quickly testing the validity of the results obtained from a given run of the behavior generator, and it also allows us to easily adapt these results on the AIBOs on later stages. The present paper details work done so far to generate complex behaviors with a genetic algorithm and shows that this is possible. At the moment the system is working generating simple behaviors, which will be later the base of complex behaviors