Evolutionary Robot with Competitive-Cooperative Neural Network(Knowledge Processing)

概要

論文の詳細を見る
This paper describes a new approach to control systems for autonomous mobile robots, using sandwiches of two different kinds of neural networks. One is a neural network for recognizing sensor information with mechanisms of competition and cooperation (CCNN), where synaptic couplings are fixed. The other is a neural network controller with adaptive synaptic couplings corresponding to genotypes in creatures and used for self-learning of wheel controls (NNC). In a computer simulation model with both of neural network, we were successful in obtaining typical types of robot that performed well in following a curved wall. The role of CCNN is to decide which sensor signals to select in a noisy environment, while NNC adjusts the synaptic couplings through genetic operations so that it can transfer the outputs from CCNN to the rotation of robot's wheel. A test was performed to show the superiority of CCNN. A robot with CCNN can pass through a narrow entrance to a concave space, and is very robust with respect to several kinds of noises. We also tested a real robot by using the synaptic couplings obtained from the simulation, and showed that the robot performs well in a real environment.
一般社団法人情報処理学会の論文
2003-10-15