Bayesian Method-Based Learning Automata for Two-Player Stochastic Games with Incomplete Information
In the field of artificial intelligence, learning automaton (LA) is a self-adaptive decision-maker which plays an important role in reinforcement learning (RL). Games of learning automata are stochastic games with incomplete information that have received frequent usage. Traditional learning automata schemes using in games are parameter-based schemes which exist a tunable parameter (stepsize) changing with different environments. In this paper, we proposed Bayesian method-based parameter-free learning automata (BPFLA) for two-player stochastic games with incomplete information. The parameter-free property indicates that a set of parameters in the scheme can be universally applicable for all configurations of games. Besides, simulation results demonstrate that BPFLA has much faster convergence rate than traditional schemes using games of learning automata with equal or higher accuracy.
KeywordsGames of learning automata Learning automata Reinforcement learning Bayesian inference
- 1.Sutton, R.S., Barto, A.G.: Reinforcement Learning: An introduction, vol. 1, No. 1. Cambridge, MIT press (1998)Google Scholar
- 2.Narendra, K.S., Thathachar, M.A.: Learning Automata: An Introduction. Courier Corporation (2012)Google Scholar
- 3.Thomas, L.C.: Games, Theory And Applications. Courier Corporation (2012)Google Scholar
- 9.Ge, H.: A parameter-free learning automaton scheme (2017). arXiv:1711.10111
- 10.Gupta, A.K., Nadarajah, S. (eds.): Handbook of Beta Distribution and its Applications. CRC Press, Boca Raton (2004)Google Scholar