Observation Reduction for State-Action Tables

The problem of planning for reachability goals in full observable nondeterministic domains has been addressed in the literature, where plans are often expressed as state-action tables under the hypothesis that, states of the world can be sensed directly by the robot or agent executing the plans. In most real world domains, however, a robot or agent distinguishes states of the world by acquiring some optional observation information at some cost. In addition, some observation information may be unnecessary to the plan execution. Observation reduction for a general state-action table, aims at reducing the observation information that is unnecessary at execution time. The contribution of this paper is an algorithm that calculates the loops of belief states, and necessary observation information that may arise during the execution of a state--action table; and transforms the state-action table into a structured plan which only branches on the necessary observation information.