Parallel distributed processing: explorations in the microstructure of cognition, vol. 1: foundations
Learning state space trajectories in recurrent neural networks
Neural Computation
Hi-index | 0.00 |
A common underlying design feature of neural network methods is the approximation of a fixed goal through a single I/O characterisation. By contrast, the problems investigated here are those which contain sequences of multiple goals. Such sequences may occur when the environmental or task requirements alter unpredictably to vary the outputs appropriate to the given inputs, or where completion of a task requires a sequence of modes of behaviour.Some of these problems prove impossible for training based on gradient descent due to the goal sequence. Other problem may be found to be increasingly difficult as the sequence extends.A Neural Steering mechanism is presented to improve training which chains subgoals so that actual and subgoal excitations are zipped together towards the goal. The results show that the steering method is able to solve such problems feasibly and robustly.