Reducing Real-world Failures of Approximate Explanation-based Rules

Most current learning and planning systems have been designed to function well in an environment which is a model of the real world. No model of the world can be perfect, however. For a system to actually be able to learn and plan with the real world it must be able to detect problems encountered while acting on the world and to reconcile its model with sensed data. We have constructed an explanation-based learning system called GRASPER which has capabilities to monitor execution of its plans and to tune its model of the world through use of explicit approximations. This paper first characterizes the different kinds of approximations and introduces the use of approximate rules for the purpose of learning uncertainty tolerant plans. Uncertainty tolerant plans offer the important advantage that they can function in spite of errors rather than imposing censors which restrict the generality of a plan. The key issue with uncertainty tolerance approximations is the ability to tune them when the system encounters failures. Our approximations for uncertainty tolerance involve tunable continuous quantities. A new general algorithm is presented which, by creating qualitative representations for the quantitative behavior of an explanation-based rule, can generate explanations as to how to increase the probability of success of the failing expectation through tuning various approximate quantities. A real-world example is given illustrating the tuning process for one of the more common failures occurring with GRASPER operating in the robotics grasping domain. An empirical comparison of failures rates for tuning and non-tuning runs is provided in the task of grasping all pieces to a children's puzzle.