Exploiting Relevance through Model-Based Reasoning

摘要: Since omnipotent reasoning is hard to perform it natural look for shortcuts that (sometime) well. We say some data relevant a task if supports an efficient computation performs correctly on the task. explore few aspects of relevance and show model-based can support these representations tasks. (1)Reasoning within context way use only information situation when arriving conclusions. present two approaches in which "context" information, incorporated with reasoning, makes computational problems easier. Using techniques intelligent agent construct its view world incrementally by pasting together many "narrower" views from different contexts. (2) In cases, where relatively simple environment very complex, modeling exactly would overload superfluous information. deduction tasks, maintaining partial (in form least upper bound theory) suffices, correct be done efficiently cases. (3) suggest machine learning approach performance measures depend learns in. this framework tasks are traditional sense become tractable. This, sense, captures intuition experience easier future Although presented separately, all based interwoven work together, contributing understanding relation between tractability AI problems. This builds recent results (Khardon & Roth 1994b; 1994a). Technical details omitted