Examining the ϵ-optimality property of a tunable FSSA

In this paper, a new fixed structure learning automaton (FSSA), with a tuning parameter for amount of its rewards, is presented and its behavior in stationary environments will be studied. This new automaton is called TFSLA (tunable fixed structured learning automata). The proposed automaton characterizes by star shaped transition diagram and each branch of the star contains N states associated with a particular action. TFSLA is tunable, so that the automaton can receive reward flexibly, even when it accepted penalty according to its previous action. Experiments show that TFSLA converges to the optimal action faster than some older FSSAs (e.g. Krinsky and Krylov) and the analytic examination proofs that the new automaton is ϵ-optimal.