A formal framework for speedup learning from problems and solutions

摘要: Speedup learning seeks to improve the computational efficiency of problem solving with experience. In this paper, we develop a formal framework for efficient from random problems and their solutions. We apply two different representations learned knowledge, namely control rules macro-operators, prove theorems that identify sufficient conditions in each representation. Our proofs are constructive they accompanied algorithms. captures both empirical explanation-based speedup unified fashion. illustrate our implementations domains: symbolic integration Eight Puzzle. This work integrates many strands experimental theoretical machine learning, including rules, macro-operator Explanation-Based Learning (EBL), Probably Approximately Correct (PAC) Learning.