Learning to behave: an investigation of connectionist approaches to behaviour-based control in autonomous agents

Most reinforcement learning applied to autonomous agents has relied on a coarse discretization of the control space. This paper presents a modular connectionist architecture for the autonomous control of a mobile agent based on a form of continuous reinforcement learning using backpropagation through random number generators. It discusses the potential of this approach as a way of decomposing a complex goal so that the structural credit assignment problem is made tractable and the complexity of the neural network topology necessary for solving a problem is reduced