Prashant J. Doshi and Piotr J. Gmytrasiewicz
Partially observable Markov decision processes (POMDPs) represent a decision-theoretic framework for optimal planning in partially observable environments. Their well-studied theoretical properties make them an attractive proposition, though their practical usefulness is somewhat diluted due to their high computational complexity. In this paper, we put forward POMDPs as an architectural component for modeling emotional behavior, and consequently use emotions to approximate optimal planning. We use an example toy problem to demonstrate how POMDPs can be engineered to produce several distinct emotional behaviors, and empirically show that the resulting affect-based plans are close approximations of the optimal affectless plans.