Genetic Programs and Co-Evolution

摘要

A co-evolutionary approach for developing programs for controlling a very simple "robot-like" simulated vehicle is presented. The main goal is to find programs that can generalize and solve other similar problems. Good results are achieved by coevolving the test cases and the simulated vehicles and using locality in both the reproduction and evaluation phases. The fitness of a controller is determined by its performance in competition with its neighbours in the test case population. The fitness of a test case is similarly determined through competition with its neigh bours in the controller population. The co-evolved controllers are more robust and general than a simple hand-designed algorithm or controllers evolved using a fixed training set.