Hierarchical Neural Network (HNN) for Closed Loop Decision Making: Designing the Architecture of a Hierarchical Neural Network to Model Attention, Learning and Goal Oriented Behavior.

摘要

The objectives of the project were to design and evaluate a hierarchical neural network (HNN) capable of real time learning and decision making in closed loop. In the initial stages of the project the problem was defined and the relating state of the art methods were surveyed. Later control of a robotic system was used as the prototypical task and a HNN was designed and compared with the state of the art adaptive control techniques. During this project the concept of exploratory schedules (ES) was developed. ES is defined as system trajectories internally generated by the HNN for the purpose of efficient learning. This concept was implemented in an open loop fashion for the control of robotic manipulators. A theorem was proved that gives constructive conditions for stable learning in closed loop. The third technique yielded improved transients in tracking desired trajectories in comparison with adaptive control methods. HNN architecture was applied as a controller for a class of nonlinear systems linear in control. It was shown to have guaranteed asymptotic stability. HNN architecture was employed with partial success in areas of pattern recognition and control.