Using estimation of distribution algorithm to coordinate decentralized learning automata for meta-task scheduling

Learning automaton (LA) is a reinforcement learning model that aims to determine the optimal action out of a set of actions. It is characterized by updating a selection probability vector through a sequence of repetitive feedback cycles interacting with an environment. Decentralized learning automata (DLAs) consists of many learning automata (LAs) that learn at the same time. Each LA independently selects an action based on its own selection probability vector. In order to provide an appropriate central coordination mechanism in DLAs, this paper proposes a novel decentralized coordination learning automaton (DCLA) using a new selection probability vector which is combined with the probability vectors derived from both LA and estimation of distribution algorithm (EDA). LA contributes to the own learning experience of each LA while EDA estimates the distribution of the whole swarm's promising individuals. Thus, decentralized LAs can be coordinated by EDA using the swarm's comprehensive knowledge. The proposed automaton is applied to solve the real problem of meta-task scheduling in heterogeneous computing system. Extensive experiments demonstrate a superiority of DCLA over other counterpart algorithms. The results show that the proposed DCLA provides an effective and efficient way to coordinate LAs for solving complicated problems.