Reinforcement Learning with Time

摘要

This paper steps back from the standard infinite horizon formulation of reinforcement learning problems to consider the simpler case of finite horizon probems. Although finite horizon problems may be solved using infinite horizon learning algorithms by recasting the problem as an infinite horizon problem over a state space extended to include time, we show that such an application of infinite horizon learning algorithms does not make ue of what is known about the environment structure, and is therefore inefficient. Preserving a notion of time within the environment allows us to consider extending the environment model to include, for example, random action duration. Such extentions allow us to model non-Markov environments which can be learned using reinforcement learnign algorithms.