Empirical Study of the Sensitivity of CACLA to Sub-optimal Parameter Setting in Learning Feedback Controllers

摘要

Continuous Action-Critic Learning Automaton (CACLA) offers an interesting alternative to traditional control approaches to feedback control problems. In this paper, we report results obtained on an inertial model of a feed drive with potentially sub-optimal parameter setting and designer decisions. Namely, we have tested different reward signals, different number of features to approximate value functions and policies, and different learning gains. The results show CACLA to be a very highly robust approach.