Autonomous mobile robot control using temporal sequence learning

Mike Denham

(University of Plymouth)

This talk will discuss some of the issues involved in designing autonomous mobile robots and describe an approach to the control of such agents which uses learnt sequences over time of pairs of sensory stimuli and motor actions, so-called sensory-action sequences (SAS), to build an internal model of the world in which the robot is navigating. A reward system is used to determine when and if sequences are learnt, based either on the achievement of a particular goal or the detection of novelty. A set of changing goals is therefore what intrinsically drives the robot both to move in such a way as to re-experience some previously encountered sensory experience, e.g., associated with a particular location in the robot's world, whilst also acquiring knowledge about its world through curiosity and responding to novel sensory experiences. Associative links created between SAS provide a dynamic goal-related sensory-action based internal model in memory from which "working memories" can be "activated". The presence of a particular goal determines which set of "working memories" are "active" at a particular time, and stimulates a competition amongst these, which determines the most appropriate action to take at any time. A computational model of the system will be described, based on a particular neural network for learning, recognising and reproducing temporal sequences.
Tuesday 14th November 1995, 14:30
Seminar Room 322
Department of Computer Science