An auto-tuning controller with supervised learning using neural nets

Concerns the use of neural nets in the feedback control of systems described by ordinary differential or difference equations. The controller is assumed to be of fixed structure but with free parameters that must be tuned to meet some previously chosen performance specifications. The supervisor consists of a neural net producing the actions and evaluating the success of failure of the selected action. An action is a point in the parameter space that is used to adjust a controller of fixed structure. Supervised learning is accomplished using a performance evaluator which alters conveniently the action network following a connectionist strategy. The analysis is based on the stochastic learning automaton, which selects an action according to a probability distribution that depends on the reinforcement signal produced by the performance evaluator. The feedback is, therefore, provided in the form of an evaluation of the results on the environment. The probability density over the action space is stored internally in the neural net, and the changes are incorporated by adjusting the synaptic weights in the learning phase