Reinforcement Learning for MDPs with Constraints

In this article, I will consider Markov Decision Processes with two criteria, each defined as the expected value of an infinite horizon cumulative return. The second criterion is either itself subject to an inequality constraint, or there is maximum allowable probability that the single returns violate the constraint. I describe and discuss three new reinforcement learning approaches for solving such control problems.

• Type: Conference
• Year: 2006
• Where: ECML
• Authors: Peter Geibel
• Keywords: ECML 2006, Horizon Cumulative Return, Inequality Constraint, Machine Learning, Maximum Allowable Probability