AI is approaching the point where it will be possible to build autonomous robotic agents capable of performing human-like tasks without direct human control. Such autonomous agents must be able to plan their activities in the face of incomplete knowledge of their environment. This project aims at understanding how such planning works and building implemented systems that accomplish it. Specifically, this investigation is aimed at the construction of an artificial rational agent capable of engaging in decision-theoretic planning in environments of realistic complexity and unpredictability. The design of a system to do automated planning is one of the traditional goals of artificial intelligence research, and some highly successful planning systems have been constructed for use in narrowly constrained environment; however, these systems presuppose that the planner knows everything it needs to know when it is first presented with the planning problem, and most of them further require complete knowledge of all relevant aspects of the agent's environment and knowledge of precisely what will result from performing any relevant act in any circumstance the planner will encounter. While such assumptions might be satisfied by an industrial robot operating in a constrained environment, human beings plan without satisfying any of these conditions. In particular, planning problems often drives the search for new knowledge rather than presupposing that the planning agent knows everything it needs to know from the beginning. And human beings do not assume that they can predict with certainty what will happen when they perform any available action under any conceivable circumstances. In constructing and evaluating plans, people take account of the varying probabilities of different consequences of actions, and they assign values and costs to those consequences before deciding whether to adopt a proposed plan. In other words, they plan decision-theoretically. The objective of this project is to understand how decision-theoretic planning is possible in an agent operating in an uncooperative and only partially predictable environment, and then to build an artificial agent whose planning capabilities more closely approximate those of human beings. This should illuminate some of the structure of rational cognition in both artificial agents and human agents.
