Optimal learning control via cell mapping

Abstract

In this paper, a new learning method based on the concept of cell-to-cell mapping is developed to generate automatically a controller from experimental data without using a system model. The continuous state space is divided into many discrete cells initially, and then the dyanmics of the system are approximated by the cell mapping. A series of learning signals is applied to excite the system, the experimental data are collected in a data buffer, and are then processed into a control table which is used to store the experienced dynamics of the system. The control table is optimized and updated through learning to improve the controller performance. Some generalized controls called pseudo-experience production rules are used to fill some empty entries in the control table. Finally, the control table is used as a controller via a table look-up scheme for the system.