Assessing the performance of different behavior selection architectures in a large and complex virtual environment

We compare the performance of autonomous agents with three different behavior selection architectures (static-threshold, winner-takes-all and voting-based) in terms of survival in a large and complex dynamic virtual environment. Experiment results indicate both advantages and disadvantages when applying each architecture in such environmental conditions, and also shows that the performance of voting-based architecture is significantly sensitive to the balance of rates at which various environmental resources can serve to satisfy an agent's physiological needs. Some problems with behavior selection architectures are identified and possible solutions are proposed.