Comparing Stochastic Planning to the Acquisition of Increasingly Permissive Plans

Abstract

Difficult problems confront learning and planning efforts for real–world domains. These domains are inherently complex and can only be represented by approximate models subject to error. Furthermore, actions taken in the world are subject to control errors and sensor readings are only approximate. Several techniques have been developed to address these problems including reactive planning and stochastic and decision–theoretic approaches. In this paper, we introduce and compare a new technique called permissive planning with stochastic planning methods. Permissive planning uses an approximate domain theory to project future states. It is designed to improve the projection process for real–world domains by tuning plan schemata to tolerate deviations between the model and reality while maintain the accuracy of projection. Empirical comparisons of stochastic and permissive planning are performed with an implemented robotic system for achieving part orientations.