Learning Uncertainty Tolerant Plans through Approximation in Complex Domains

Abstract

Abstract : Most artificial intelligence systems have difficulty functioning in complex, uncertain environments. These systems make many implicit assumptions about the world which, if inaccurate, will cause them to fail. For systems to function in these environments requires that explicit approximations be used, that approximation failures be detectable, and that the system has some method for recovering from failures. An architecture for learning approximate plans is introduced based on explanation-based learning. In explanation-based learning, a system uses a single observed example in conjunction with its domain knowledge to construct an explanation for how some goal in that example was achieved. This explanation can be generalized into a plan capable of functioning not only on the specific observed example but a wide variety of examples of that class. Central to explanation-based learning is a concept called operationally which allows a system to rate different plans. We offer a comprehensive definition for this term for use with systems functioning in real-world environments. We demonstrate how the architecture for learning with approximations can be employed to learn uncertainty-tolerant plans. An implemented system called GRASPER is described which embodies the approximation architecture and learns uncertainty-tolerant plans for grasping blocks in the robotics domain.