Evolutionary robot simulations with competitive-cooperative neural network and adaptive synaptic couplings

Abstract

This paper descn'bes a new approach to control systems for an autonomous mobile robot b y using sandwiches of two different kinds of neuml networks. One is a neuml network for recognizing sensor infonnation with a mechanism of competition and cooperation, where synaptic couplings are fixed. The second is a neural network with adaptive synaptic couplings corresponding genotype in the creature and used for self-learning of wheel controls. I n the computer simulative model with both two parts of neural network, we are successful to obtain typical types of robot with a good performance when going along the carved wall. The first part of the networks play a role to make a decision among sensor signals under noisy enuiwnment, while the second part i s effective to adjust the synaptic couplings through genetic operations so that it m a y tmnsfer the outputs from the first stage with the competition-cooperation neuml network (CCNN) to the rotation of robot's wheel. A test is perfonned to show the superiority of CCNN. A robot with C C N N can enter into a narrow entrance with concaved space and strong robustness against several kinds of noises, while the robot without C C N N cannot enter into the space due to the sumunding noise around the entmnce.