Experimentation-Driven Operator Learning

Expert-provided operator descriptions are expensive, incomplete, and incorrect. Given the assumptions of noise-free information and an completely-observable state, OBSERVER can autonomously learn and refines new operators through observation and practice (Wang 1995). WISER, our learning system, relaxes these assumptions and learns operator preconditions through experimentation utilizing imperfect expertprovided knowledge. Our decision-theoretic formula calculates a probably best state S’ for experimentation based on the imperfect knowledge. When a robotic action is executed successfully for the first time in a state S, the corresponding operator’s initial preconditions are learned as parameterized S. We empirically show the number of training examples required to learn the initial preconditions as a function of the amount of injected error. The learned preconditions contain all the necessary positive literals, but no negative literals. Unless given a rule like arm-empty --+ +&ding(X), a robot may believe a state, {(arm-empty), (holding bo~j}, to be possible due to unavoidable perceptual alias. The plan execution in this state is unreliable. To make a planner robust in a noisy state, WISER learns constraining negative preconditions by interconnecting two types of logic systems used in planning systems. The state representation uses two-value logic plus the Closed-World Assumption and the operator representation uses three-value logic. Let L represent all the predicates known to WISER and P the predicates true in S. By CWA, N, the predicates not true in S, is defined as {L P}. WISER induces preconditions from S* = {P U TN}, which prevent inconsistent actions in a noisy state. Let the instantiated operators, iopi and iopj, be obtained from an operator op by instantiating each parameter of op to the same object respectively except that one parameter is instantiated to different objects of the same type, say a for iopi and b for iopj. iqj is obtained by substituting b for a. In any state S, if the preconditions for iopi and iopj are both satisfied (or not satisfied) and their respective actions have the same (parameterized) effects on S, the two objects a