Weight-Assignment Last-Position Elimination-Based Learning Automata

Abstract

Learning Automata (LA) is an adaptive decision-making unit under the reinforcement learning category. It can learn the randomness of the environment by interacting with it and adaptively adjust its behavior to maximize its long-term benefits from the environment. This learning behavior reflects the strong optimization ability of the learning automaton. Therefor LA has been applied in many fields. However, the commonly used estimators in previous LA algorithms have problems such as cold start, and the initialization process can also affect the performance of the estimator. So, in this paper, we improve these two weaknesses by changing the maximum likelihood estimator to a confidence interval estimator, using Bayesian initialization parameters and proposes a new update strategy. Our algorithm is named as weight-assignment last-position elimination-based learning automata (WLELA). Simulation experiments show that the algorithm has higher accuracy and has the fastest convergence speed than various classical algorithms.