拡張型GPによるロボット行動の自動生成へのアプローチ : ライブラリ導入による部分汎化行動の獲得

概要

論文の詳細を見る
We present a new approach to genetic programming (GP) for robot learning. Our goal is to realize robustness in robot navigation control by means of the automatic discovery of functions. In robot learning by conventional GP, the program is not separated for each partial behavior because the program is represented as a long program. This results in lack of reusability and robustness in the program. Due to the fact that GP can not rearrange the acquired program if the environment is unstable. We propose an approach, called COAST, to improve the robustness, which has a library for storing certain subroutines for reuse. These subroutines are generated by automatically defined functions. We use the wall following problem to demonstrate the efficiency of this mehod.
一般社団法人日本機械学会の論文
1997-05-25