Discretized estimator learning automata
J. Kevin Lanctot, B. John Oommen · IEEE Transactions on Systems Man and Cybernetics · 1992
The improvements gained by rendering the various estimator learning algorithms discrete are investigated. This is done by restricting the probability of selecting an action to a finite discrete subset of (0, 1). This modification is proven to be epsilon -optimal in all stationary environments. Various discretized estimator algorithms (DEAs) are constructed. Subsequently, members of the family of DEAs are shown to be epsilon -optimal by deriving two sufficient conditions required for the epsilon -optimality-the properties of monotonicity and moderation. A conjecture about the necessity of these conditions for epsilon -optimality is presented. Experimental results indicate that the discrete modifications improve the performance of the algorithms so that the automata constitute fast-converging and accurate learning automata.>