On the Behavior of Learning Automata and Its Applications

摘要

The report is concerned with the behavior of a class of learning automata operating in random environment which basically is assumed to be discrete and stationary. Based on the strategies of two-armed bandit problems, two sequential learning models and having deterministic and stochastic transition rules, are proposed. Automata using those strategies have shown certain advantages over the linear-strategy models. They are easy to implement since no further calculation is necessary except counting. The optimality can only be achieved asymptotically if the memory capacity increases indefinitely; hence, the number of trials must be infinitely large as well. A sharper bound of iterated logarithm inequality is derived. The learning composite fuzzy automaton operating in a random environment has shown the desired property of asymptotic optimality. Extension from P-model to S-model has been made. This can be immediately applied to multi-modal optimum-seeking problems. (Author)