Coevolutive adaptation of fitness landscape for solving the testing problem

摘要

A general framework, called Uniform Coevolution, is introduced to overcome the testing problem in evolutionary computation methods. This framework is based on competitive evolution ideas where the solution and example sets are evolving by means of a competition to generate difficult test beds for the solutions in a gradual way. The method has been tested with two different problems: the robot navigation problem and the density parity problem in cellular automata. In both test cases using evolutive methods, the examples used in the learning process biased the solutions found. The main characteristics of the Uniform Coevolution method are that it smoothes the fitness landscape and, that it obtains "ideal learner examples". Results using uniform coevolution show a high value of generality, compared with non co-evolutive approaches.