Experimentation-driven learning of planning operators.

摘要

Knowledge engineering for planning is expensive and the resulting knowledge can be imperfect. To autonomously learn a plan operator definition from environmental feedback, an operator learning system WISER explores an instantiated literal space using a breadth-first search technique. Each node of the search tree represents a world state, a unique subset of the instantiated literal space. A state at the root node is called a seed state. WISER can generate seed states with or without utilizing imperfect expert knowledge. WISER experiments with an operator at each node. A positive state, in which an operator can be successfully executed, constitutes initial preconditions of an operator. We analyze the number of required experiments as a function of the number of missing preconditions in a seed state. We introduce a naive domain assumption to test only a subset of the exponential state space. Since breadth-first search is expensive, WISER introduces two search techniques to reorder literals at each level of the search tree. We demonstrate performance improvement using the naive domain assumption and literal-ordering heuristics. Unlike other systems, WISER can model the effects on objects, which are not bound to any operator parameter, by introducing a variable. We show that a machine-generated definition of effects is often simpler in representation than expert-provided definitions. To refine imperfect operators when the internal expectation is not consistent with the external reality, WISER monitors and searches through an instantiated literal space during plan execution. Compared to some systems, WISER needs less monitoring and learns multiple missing literals without relying on a previous success case. Assuming that the system can transition to any user-specified state to perform experiments can be unrealistic. To overcome this shortcoming, we expand WISER to experiment with operators in a current state, monitor results, and experiment again from the naturally-generated new state. we present the empirical results showing that our new approach performs better than our previous approach.