On conditional optimality of a class of learning automata in random environments (Q800091)

The author analyzes the asymptotic behavior of a new class of learning automata in Q-model environments (i.e., random environments with a finite number of responses). The reinforcement scheme adopted has just the same form as the Bayesian learning scheme, and is specialized to coincide with the \(\beta\) -model scheme introduced by R. D. Luce for an environment with binary responses (a P-model environment). A learning automaton is optimal, by definition, if it converges to its goal with probability one as time goes to infinity. Under some restrictions on the environments, the author proves the optimality of the defined automata. A method for constructing a special type of learning automata is presented, which is based on the use of the maximum-entropy principle. By applying this method to any Q-model environment and suitably choosing the output function, we can get an effective learning automaton with simple structure. The optimality of the resulting automaton is also investigated.