Bayes-optimal reinforcement learning for discrete uncertainty domains

  • Authors:
  • Emma Brunskill

  • Affiliations:
  • Carnegie Mellon University, Pittsburgh, PA

  • Venue:
  • Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems - Volume 3
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

An important subclass of reinforcement learning problems are those that exhibit only discrete uncertainty: the agent's environment is known to be sampled from a finite set of possible worlds. In contrast to generic reinforcement learning problems, it is possible to efficiently compute the Bayes-optimal policy for many discrete uncertainty RL domains. We demonstrate empirically that the Bayes-optimal policy can result in substantially and significantly improved performance relative to a state of the art probably approximately correct RL algorithm. Our second contribution is to bound the error of using slightly noisy estimates of the discrete set of possible Markov decision process parameters during learning. We suggest that this is an important and probable situation, given such models will often be constructed from finite sets of noisy, real-world data. We demonstrate good empirical performance on a simulated machine repair problem when using noisy parameter estimates.