Effects of using different neural network structures and cost functions in locomotion control

Abstract

Effects of using different neural network structures and cost functions in locomotion control are investigated. Simulations focus on refinement and a thorough understanding of an artificial intelligent learning scheme. This scheme uses a neural network controller with backpropagation through time learning rule. Through learning, the controller can generate locomotion trajectory along a pre-defined path. Different issues regarding the scheme have been examined. They include the effects of using different numbers of hidden units, the effects of using only angle parameters in the cost function, and the effects of including an energy criterion in the cost function.