Learning and Adaptive Hybrid Systems for Nonlinear Control

摘要

Connectionist learning systems are function approximation systems which learnfrom examples, and have received an increase in interest in recent years. They have been found useful for a number of tasks, including control of high dimensional, nonlinear, or poorly modeled systems. A number of approaches have been applied to this problem, such as modeling inverse dynamics, backpropagating error through time, reinforcement learning, and dynamic programming based algorithms. The question of integrating parial a priori knowledge into these systems has often been a peripheral issue. Control systems for nonlinear plants have been explored extensively, especially approaches based on gain scheduling or adaptive control. Gain scheduling is the most commonly used, but requires extensive modeling and manual tuning, and doesn't work well with high-dimensional, nonlinear plants, or disturbances. Adaptive control addresses these problems, but usually can't react to spatial dependencies quickly enough to compete with a well-designed gain scheduled system. This thesis explores a hybrid control approach which uses a connectionist learning system to remember spatial nonlinearities discovered by an adaptive controller. The connectionist system learns to anticipate the parameters found by an indirect adaptive controller, effectively becoming a gain scheduled controller. The combined system is then able to exhibit some of the advantages of gain scheduled and adaptive control, without the extensive manual tuning required by traditional methods. A method is presented for making use of the partial derivative information from the network.