A two-level neural network system for learning control of robot motion

Abstract

A two-level neural net system is proposed as a learning controller for a mobile robot. The lower-level subsystem adapts to environmental changes while the higher-level subsystem maintains a library of connection weights for a variety of distinct environments and loads the appropriate set of coefficients to the lower level following the recognition of the current environment. The conceptual design of the system is presented, as well as a qualitative analysis of the lower-level subsystem convergence performance using simulation results. The simulation results show that, rather than random initial weights, a prototype set obtained from a simple analytical model could markedly reduce the number of iterations. The proposed two-level neural net structure, by recalling from a library the appropriate set of connection weights, can bring down the number of iterations below 10, given that the recalled weights are within approximately 15% of the steady-state values.<>