Optimal Policies for Partially Observable Markov Decision Processes

摘要

this paper, we will be exploring a specific model-based scheme in which decisions need to be made. Even when we cannot assume we have the model, we can use techniques, [2], that allow us to approximate the model and then apply these model-based schemes. The many problems associated with such models will be outlined in a subsequent section. Throughout this discussion, the term agent will refer to the automated process that has to make decisions. A convenient example of an agent is that of an autonomous robot trying to survive in a real world environment. However, the agent can simply be a computer program such as one that does medical diagnosis. In this case, the model of the world might be based upon statistics and medical research.