| I've got a bit of a confession to make... You see I didn't actually make it to Klaus-Dieter's invited talk on "Case-Based Reasoning & Experimental Software Engineering". Well it was on terribly early and some German colleagues of Klaus-Dieter's would not let me go to sleep (K-D: it was Stefan & Wolfgang). I assume the presentation was about the work being done at the Fraunhoffer Institute for Experimental Software Engineering and the use of CBR for organisational learning in software engineering (please somebody correct me if I'm wrong). |
|
|
| After a lot of coffee I did manage to make it to David McSherry's talk and I was very glad that I did. David presented a deceptively simple adaptation heuristic for an estimating problem (house prices). The technique introduced a concept of case dominance which was used to maintain the consistency of the case-base by keeping adapted estimates within known upper and lower bounds. The paper shows how a simple idea that has been well thought out can be very useful. |
|
|
| Ivo Vollrath described the application of CBR to the reuse of electrical designs. The system called READee, can take an abstract high level query and transform it using rules into a lower level more specific query which can be used to retrieve specific designs. |
|
|
| Axel Schairer then described how CBR could be used to reduce effort in calculating proofs for formal software specifications. At first sight this seems like an odd approach since the notion of similarity seems incompatible with formal verification. But, since complex proofs are derived from many simpler sub-proofs it is often possible to avoid recalculating the sub-proofs. The results presented were very encouraging with a 45-70% reduction in user interaction. |
|
|
| The last talk of the morning was by David Sinclair of Dublin City University. He presented a very interesting paper on the use of CBR for chess playing. This has always seemed to me to be a good application for CBR since it is fairly obvious that good chess players can remember board positions and moves. They are not using brute force to search all possible moves. David used Principle Component Analysis to identify 11 predictive features describing board positions from a total set of 56. Thus, he had reduced the problem space from 56 dimensions to 11. He was then able to use a database of 16,728 games (801,840 individual positions) to test his system with good results. In the end his system can identify a small set of "good" candidate moves which can then be searched using a normal search algorithm. Thus, CBR is being used to effectively prune the search space in a way that could be analogous to human Grand Masters. |
|
...and yes, I was bought
some Guinness, thank you :-)