Speedup of evolutionary behavior learning with crossover depending on the usage frequency of a node

摘要

For online robot behavior learning, we propose heuristics using node usage for speedup of evolutionary learning, and verify the utility experimentally. Genetic programming (GP) is an evolutionary way to acquire a program through interaction with an environment. Since behaviors of a robot are described with a program, researches on applying GP to robot behavior learning have been activated. Unfortunately, in most of the studies, the behavior learning is done off-line using simulation, not a real robot. Because convergence of GP is slow, this makes operation of a real robot quite expensive. However, since situations out of simulation easily happens in a real world, the behavior learning with a real robot (called online learning) remains very significant. Thus, in order to make online behavior learning with GP practical, we propose a crossover method for speedup of GP using node usage of a program.