Towards adaptive perception in autonomous robots using second-order recurrent networks

In this paper a higher-order recurrent connectionist architecture is used for learning adaptive behaviour in an autonomous robot. This architecture consists of two sub-networks in a master-slave relationship: a function network for the coupling between sensory inputs and motor outputs, and a context network, which dynamically adapts the sensory input weights in order to allow a flexible, context-dependent mapping from percepts to actions. The capabilities of this architecture are demonstrated in a number of action selection experiments with a simulated Khepera robot, and it is argued that the general approach of generically dividing the overall control task between sequentially cascaded context and function learning offers a powerful mechanism for autonomous long- and short-term adaptation of behaviour