Backpropagation learning in real autonomous mobile robot using proximal and distal evaluation of behaviour

Murase et al. (1998) have previously proposed a backpropagation algorithm for the development of behaviour based autonomous robots. Although the method was proved applicable to a real mobile robot, there existed some problems such as the convergence to the local minimum and the variability among trials. In order to improve the performance, we introduce a new criterion for selecting the training data set. Coefficients of a multilayer neural network, that determined the sensor-motor reflex of the robot, were first set randomly, and the robot was allowed to behave in environment for some time. Sets of the sensor-motor values were continuously sampled during the free-moving period, and each set was evaluated by the behaviour that occurred after the sampling by using an evaluation function. The behaviour was evaluated by both immediate response to near-by obstacles and long-range navigation capability. The new criterion provided a much faster and stable convergence than the previous one, and far better than the conventional genetic algorithm.