Learning feature weights for CBR: Global versus local

Abstract

k-Nearest Neighbour is a popular case retrieval technique in Case-Based Reasoning. It has the disadvantage that its accuracy depends strongly on the weights assigned to the case features. This problem can be addressed by using Introspective Learning to discover appropriate values for feature weights. The basic idea with Introspective Learning (IL) is to examine cases that are similar in order to discover which features are important and which are not. The only problem with this idea is that there are a myriad of ways in which weights can be updated based on this kind of analysis. There are several different cues that can trigger a weight change; there are several ways in which the weights can be changed and there is the added complication that weights can be global or local. In this paper we report some analysis of IL in a CBR system for conflict resolution in Air Traffic Control. We show that local weights are best in this particular domain and we show which update cues are most effective. We also show that overfitting can be a problem with IL and we discuss how it can be avoided.

This research was carried out with the support of Eurocontrol Experimental Centre at Paris, the European centre for research and simulation in Air Traffic Control.