Selfish vs. Global Behavior Promotion in Car Controller Evolution

Abstract # ↰

We consider collective tasks to be solved by simple agents synthesized automatically by means of neuroevolution. We investigate whether driving neuroevolution by promoting a form of selfish behavior, i.e., by optimizing a fitness index that synthesizes the behavior of each agent independent of any other agent, may also result in optimizing global, system-wide properties. We focus on a specific and challenging task, i.e., evolutionary synthesis of agent as car controller for a road traffic scenario. Based on an extensive simulation-based analysis, our results indicate that even by optimizing the behavior of each single agent, the resulting system-wide performance is comparable to the performance resulting from optimizing the behavior of the system as a whole. Furthermore, agents evolved with a fitness promoting selfish behavior appear to lead to a system that is globally more robust with respect to the presence of unskilled agents.