The statistics of state-spaces

Abstract

The state-space model is a general, powerful, and elegant representation of problem solving. Nevertheless, state-spaces have rarely been used to model realistic environments because conventional state-spaces are inherently deterministic, while the world is not. This paper extends the model toprobabilistic state-spaces (PSS), which elegantly capture many of the world's complexities by regarding state values as generated by random variables. These PSS models, when combined with decision analytic techniques for knowledge elicitation and encoding, should yield realistic representations. The first investigation of states based on random variables derived the expected-outcome model of two-player games, which led to some powerful results about game evaluators. Most of the work done on expected-outcome was quite general, and should extend easily beyond game-trees to arbitrary state-spaces. The second potential application domain, strategic planning for robot controllers in automated assembly plants, is currently under investigation.