A neural network which rapidly learns to perform a multistep task

This research develops a neural network which can learn to perform a multi-step task. The particular task chosen is the autonomous navigation of a vehicle. The task is taught in three phases: to approach a given goal location in the absence of obstacles; to avoid collisions with obstacles during the goal approach; and to move through groups of obstacles while approaching the goal. The network architecture and learning rules for the first phase support a rapid acquisition of domain concepts from demonstrated examples and from learning-by-doing. Large weight changes early during the learning process permit a fast approximation of the target behavior. Each weight change is limited by the weight's plasticity value which represents the network's belief that the weight is incorrect. One of the subnetworks into which the network is broken is a very simplified short-term memory. It provides the contextual information for the next step which is crucial for the network to be able to perform the multi-step task correctly. The next two phases employ additional components which learn how to influence the overall behavior by changing the previously learned spread of activation among components rather than within components. Previous behavior is overlaid by the new knowledge, but remains intact. The new components of phases two and three enhance the network's perceptual capabilities. Additional features can be detected which contribute to the overall improvement of the network's behavior. The network is implemented on a Cray YM-P in FORTRAN.