Search Heuristics, Case-Based Reasoning and Software Project Effort Prediction

Abstract

This paper reports on the use of search techniques to help optimise a case-based reasoning (CBR) system for predicting software project effort. A major problem, common to ML techniques in general, has been dealing with large numbers of case features, some of which can hinder the prediction process. Unfortunately searching for the optimal feature subset is a combinatorial problem and therefore NP-hard. This paper examines the use of random searching, hill climbing and forward sequential selection (FSS) to tackle this problem. Results from examining a set of real software project data show that even random searching was better than using all available for features (average error 35.6% rather than 50.8%). Hill climbing and FSS both produced results substantially better than the random search (15.3 and 13.1% respectively), but FSS was more computationally efficient. Providing a description of the fitness landscape of a problem along with search results is a step towards the classification of search problems and their assignment to optimum search techniques. This paper attempts to describe the fitness landscape of this problem by combining the results from random searches and hill climbing, as well as using multi-dimensional scaling to aid visualisation. Amongst other findings, the visualisation results suggest that some form of heuristic-based initialisation might prove useful for this problem.