D-rules: learning & planning

Abstract

One current research goal of Artificial Intelligence and Machine Learning is to improve the problem-solving performance of systems with their own experience or from external teaching. The work presented in this paper concentrates on the learning of decomposition rules, also called d-rules, i.e., given some examples learn rules that guide the planning process, in new problems, by determining what operators are to be included in the solution plan. Also a planning algorithm is presented that uses the learned d-rules in order to obtain the desired plan. The learning algorithm includes a value function approximation, which gives each learned rule an associated function. If the planner finds more than one applicable d-rule, it discriminates among them using this feature. Decomposition rules have been learned in the blocks world domain, and those d-rules have been used by the planner to solve new problems.