Bayes Adaptive Reinforcement Learning versus Off-line Prior-based Policy Search: an Empirical Comparison

This paper addresses the problem of decision making in unknown finite Markov decision processes (MDPs). The uncertainty about the MDPs is modeled using a prior distribution over a set of candidate MDPs. The performance criterion is the expected sum of discounted rewards collected over an infinite length trajectory. Time constraints are defined as follows: (i) an off-line phase with a given time budget can be used to exploit the prior distribution and (ii) at every time step of the on-line phase, decisions have to be computed within a given time budget. In this setting, we compare two decision-making strategies: OPPS, a recently proposed meta-learning scheme which mainly exploits the off-line phase to perform policy search and BAMCP, a state-of-the-art model-based Bayesian reinforcement learning algorithm, which mainly exploits the on-line time budget.
We empirically compare these approaches in a real Bayesian setting by computing their performances over a large set of problems. To the best of our knowledge, it is the first time that this is done in the reinforcement learning literature. Several settings are considered by varying the prior distribution and the distribution from which test problems are drawn. The main finding of these experiments is that there may be a significant benefit of having an off-line prior-based optimization phase in the case of informative and accurate priors, especially when on-line time constraints are tight.