Evolving SoccerBots : A Retrospective

摘要

In the RoboCup97 robot soccer tournament, we entered a team of softbot programs whose player strategies had been entirely learned by computer. Our team beat other human-coded competitors and received the RoboCup97 Scientific Challenge award. This paper discusses our approach, and details various ways that, in retrospect, it could have been improved. 1 Introduction The RoboCup simulation tournament pits teams of softbot programs against each other in simulated soccer matches,promoting artificial intelligence strategies that are robust in dynamic, noisy environments. For the RoboCup 1997 simulation tournament, most teams refined well-understood robotics techniques in order to win the competition. Almost every team used hand-coded decision algorithms, though some optimized a few low-level functions (like ball interception) with backpropagation or decision trees. In contrast, we ([Luke et al 1998a]) entered a softbot team whose high-level strategies were entirely learned through genetic ...