Simple Model-Based Exploration and Exploitation of Markov Decision Processes Using the Elimination Algorithm

摘要

The fundamental problem in learning and planning of Markov Decision Processes is how the agent explores and exploits an uncertain environment. The classical solutions to the problem are basically heuristics that lack appropriate theoretical justifications. As a result, principled solutions based on Bayesian estimation, though intractable even in small cases, have been recently investigated. The common approach is to approximate Bayesian estimation with sophisticated methods that cope the intractability of computing the Bayesian posterior. However, we notice that the complexity of these approximations still prevents their use as the long-term reward gain improvement seems to be diminished by the difficulties of implementation. In this work, we propose a deliberately simplistic model-based algorithm to show the benefits of Bayesian estimation when compared to classical model-free solutions. In particular, our agent combines several Markov Chains from its belief state and uses the matrix-based Elimination Algorithm to find the best action to take. We test our agent over the three standard problems Chain, Loop, and Maze, and find that it outperforms the classical Q-Learning with e-Greedy, Boltzmann, and Interval Estimation action selection heuristics.