Abstract
A behavior network and its learning rule are proposed as an autonomous motion acquisition method for an agent in an unknown environment. The proposed method is applied to a learning of the collision avoidance by a simulation model of Khepera. The motion is decomposed into 8 behavior elements, and the proposed behavior network is a simple one-layered network, which maps a vector in an input sensor space to one output unit corresponding to each behavior element. The weights are obtained by Hebbian type learning according to a given reward, which evaluates the behavior chosen by the network. The structure of the network, which is decomposable into the behavior elements, enables a straightforward expansion of the network by the addition of output units. A simulation shows the successful learning of the agent, and more various behavior patterns are obtained by the expansion of the network. The proposed method is easily applied to other motion acquisition tasks owing to the simplicity of the network structure and the learning rule, and an example is demonstrated by a learning simulation of the reaching task by the same model.
