A new class of ε-optimal learning automata

Abstract

New class of P-model absorbing ε-optimal learning automata was presented in this paper. The proposed learning automaton, Discretized Generalized Stochastic Estimator (DGSE) learning automaton, not only possesses the characteristics of the Stochastic Estimator Reward-inaction (SE RI ) learning automaton and the Discretized Generalized Pursuit Algorithm (DGPA) learning automaton, but also converges with a remarkable speed and accuracy. The asymptotic behavior of the DGSE algorithm is analyzed. Furthermore, we stick out the pitfalls in the proof of SE RI algorithm, proved the proposed DGSE algorithm to be ε-optimal, and pointed out that this proof process could be applied to prove SE RI algorithm. It's known that the SE RI learning automaton is the fastest learning automaton up to now, whereas, the proposed DGSE learning automaton is much faster than the SE RI learning automaton. A great number of experiments and simulations verified the propose DGSE learning algorithm is quite efficient when operating in P-model stationary environment. © 2011 Springer-Verlag.