Subject: IJCAI-97 WORKSHOP REVIEW FORM Comments for Author Title: Assumption Hunting as Developing Method for Problem-Solving Methods Number: 1. How RELEVANT is this paper to AI researchers? (mark one box) [x] Very relevant [ ] Moderately relevant [ ] Not relevant Please explain your rating: Fits exactly with the CFP. Hmmm... let me see. who wrote this paper and the CFP? 2. How SIGNIFICANT is this paper? [x] Very significant [ ] Moderately significant [ ] Not significant Please explain your rating: next step in KA: managing change 3. How ORIGINAL is this paper? [?] Very original [x] Moderately original [ ] Not original Please explain your rating: having read the pos/Akkermans paper, it is unclear to me if this is a bold new direction or part of a large "assumption-management in PSM" group containing Benjamin, Fensel, Pos, ... Or am I doing you an injustice? Are you in fact leading the pack and they are following your lead? 4. Is this paper technically SOUND? [x] Yes [ ] Seems valid, but did not check completely [ ] Has minor errors [ ] Has major errors If there are errors, please identify them: 5. How well is this paper PRESENTED? Good Average Poor Overall organization [x] [ ] [ ] English [x] [ ] [ ] Readability [x] [ ] [ ] If there are presentational problems, please identify them: small point: should the title be: Assumption Hunting for Developing Methods for Problem-Solving Methods --------------------^^^-----------------^ 6. Further comments, advice or explanations (Please be specific and constructive, especially with respect to any negative judgments above. Point to the section(s) where an error occurs, cite omitted references, etc.) Use as much space as you need. My full response to this article is an article in itself, one that i hope dr Fensel will help me to write. however, while that later article grows, let me say: 1) it is not clear to me how the assumptions are manipulated? automatically? or by a knowledge engineer browsing the output of KIV. this should be clearer. note that both manipulation options are valid: one is a decision support system and the other is an automatic expert system 2) Abduction is usually defined as two equations: EQ1: Theory and Assumptions ==> Goals EQ2: Theory and Assumptions (does not ==>) contradictions Now your browsing of KIV to find the assumptions is totally consistent with EQ1. However, the fun really begins when we move to EQ2. Suppose KIV is pursuing multiple proofs and the assumptions made by those proofs are inconsistent. Then EQ2 kicks in and things get really fun because you now have to trade off assumptions supporting one desired proof with assumptions that support other desired proofs while blocking this one. Formally, this trading off process implicit in EQ2 moves the system from a single-world reasoning device to a multiple worlds reasoning device. i.e. the "theorem prover" you need here is not a single world deductive device, but a multiple world device. note that such multiple world devices divide into two broad groups: JTMS (doyle's approach) vs ATMS (deKleer's approach). in the JTMS, you leap from world to world but can only hold a single consistent belief set (world) at one time. in the ATMS, all worlds are available for reflection at one time. i.e. the ATMS is parallel worlds development while the JTMS is serial worlds development. My own (untested) conviction is that the ATMS approach is better. It seems to me that if you are going to reflect over different options (worlds) and trade them off, you need to have access to those options. That is ATMS better than JTMS. the down side of this belief is that the ATMS is abduction, np-hard, slower while the JTMS is a extension of deduction, can be faster and lots of current theorem provers can be extended to a JTMS approach. SO, yes, you are doing abduction, but only EQ1 abduction. EQ2 is not in your current thinking. If it was, you would be writing assessment criteria for different assumption sets defining your different worlds. 3) If you are feeling generous, you might want to add an HT4 reference. Menzies' proposal is to focus more on the assumptions generated in the search space traversal rather than in the assumptions found in the PSM. This KIV approach is what I call KL-B based (assumes a explicit PSM which can vary a lot) while HT4 is KL-A based (one "PSM" which can be heavily parameterised. oh, SOAR is another KL-A approach). For more details, see section 2.3 of my submission for more details). ===================================== Comments for Program Committee Members Only (This section of the review will be withheld from the author eehm: Only if he is NOT a member of the organisation committee) Title: Assumption Hunting as Developing Method for Problem-Solving Methods Number: 1. My recommendation is: [x] Accept [ ] Leaning to accept [ ] Leaning to reject [ ] Reject 2. How confident are you in your appropriateness as a referee for this paper? [x] Very confident - I am an expert in this area. [ ] Confident - I have a reasonable knowledge of this area. [ ] Fairly confident - I have some knowledge of this area. [ ] Not confident - I have no significant knowledge of this area. 3. If this paper is marginal or unusual in some way, can you comment on anything else that might help the Program Committee reach a decision? Reviewer: (please fill in your name here) Tim Menzies