Last-Position Elimination-Based Parallel Learning Automata

The updating scheme for Learning Automata (LA) is important. Thathachar and Arvind first proposed parallel operation of leaning automata (LA), which was a promising mechanism that could maintain the accuracy while reducing convergence time. In this paper, we implement this mechanism which helps to break the limit of the convergence speed of single LA. In contrast to existing scheme, the proposed scheme eliminates the worst performed LA in the sequence of interactions, until there is only one LA left. We compare Last-Position Elimination-Based Parallel Learning Automata (LEPLA) scheme with the classic one incorporating two pursuit schemes, discretized generalized pursuit algorithm (DGPA) and discretized pursuit algorithm with reward-inaction (DP_RI), respectively. Simulations prove that the proposed scheme gets an evidently higher accuracy and faster convergence than the classic ones.