The influence of search components and problem characteristics in early life cycle class modelling

This paper examines the factors affecting the quality of solution found by meta-heuristic search when optimising object-oriented software class models. From the algorithmic perspective, we examine the effect of encoding, choice of components such as the global search heuristic, and various means of incorporating problem- and instance-specific information. We also consider the effect of problem characteristics on the (estimated) cost of the global optimum, and the quality and distribution of local optima. The choice of global search component appears important, and adding problem and instance-specific information is generally beneficial to an evolutionary algorithm but detrimental to ant colony optimisation. The effect of problem characteristics is more complex. Neither scale nor complexity have a significant effect on the global optimum as estimated by the best solution ever found. However, using local search to locate 100,000 local optima for each problem confirms the results from meta-heuristic search: there are patterns in the distribution of local optima that increase with scale (problem size) and complexity (number of classes) and will cause problems for many classes of meta-heuristic search.