Minimizing Necessary Observations for Nondeterministic Planning

Abstract

Autonomous agents interact with their environments via sensors and actuators. Motivated by the observation that sensors can be expensive, in this paper we are concerned with the problem of minimizing the amount of sensors an agent needs in order to successfully plan and act in a partially observable nondeterministic environment. More specifically, we present a simple greedy top-down algorithm in the space of observation variables that returns an inclusion minimal set of state variables sufficient to observe in order to find a plan. We enhance the algorithm by reusing plans from earlier iterations and by the use of functional dependencies between variables that allows the values of some variables to be inferred from those of other variables. Our experimental evaluation on a number of benchmark problems shows promising results regarding runtime, numbers of sensors and plan quality.