A Theoretical Comparison of Models

Abstract

We have seen in the previous chapter that Markov Decision Processes can be consid- ered an “ideal” approach to the implementation of intelligent agents. Even though assigning utilities to states and probabilities to transitions between states might be regarded as a questionable way to solve the problem of preference, there are many situations in which this is acceptable. Once we have accepted that the problem is cor- rectly formulated in terms of the probabilities of actions having particular effects, and certain states having higher rewards than others, the MDP solution algorithms yield MEU-optimal policies. By this we mean mappings of states into actions that tell the agent what to do in each state, based on the probable outcomes of every possible action.