Applicability of Reinforcement Learning
Paul E. Utgoff, Paul R. Cohen · 1998
We describe our experiences in trying to implement a hierarchical reinforcement learning system, and follow with conclusions that we have drawn from the difficulties that we encountered. We present our objectives before we started, the problems we encountered along the way, the solutions we devised for some of these problems, and our conclusions afterward about the class of problems for which reinforcement learning may be suitable. Our experience has made it clearer to us when and when not to select a reinforcement learning method. Introduction We describe our experiences in trying to implement a hierarchical reinforcement learning system. This includes our design objectives, the problems we encountered, the solutions we devised for some of these problems, and our conclusions afterward. These conclusions are somewhat lengthy, making the paper fall into two larger parts: the technical aspects of what we tried to do, and the thinking that our difficulties led us to do. Before proceeding...