Towards the autonomous control of mobile robots by connectionist experts

This paper concerns an investigation into the form of mobile robot control known as reactive navigation. It describes the development of control architectures that are layered, as in the well known subsumption architecture, but constructed on connectionist principles. Thus individual navigational competences are not hand-crafted but derive from the learning capabilities of the nets employed. In these experiments the learning is not supervised and the networks are loosely coupled to promote selective interpretation of the common sensory data sets. Trial and error learning occurs at each level as the robot navigates towards a series of goals. A brute force control strategy is compared with a probabilistic gating network, with nonlinear memory to exploit temporal context information, as alternative means of encouraging individual competences to develop in each module