A Framework for Meta-heuristic Parameter Performance Prediction Using Fitness Landscape Analysis and Machine Learning

Abstract

The behaviour of an optimization algorithm when attempting to solve a problem depends on the values assigned to its control parameters. For an algorithm to obtain desirable performance, its control parameter values must be chosen based on the current problem. Despite being necessary for optimal performance, selecting appropriate control parameter values is time-consuming, computationally expensive, and challenging. As the quantity of control parameters increases, so does the time complexity associated with searching for practical values, which often overshadows addressing the problem at hand, limiting the efficiency of an algorithm. As primarily recognized by the no free lunch theorem, there is no one-size-fits-all to problem-solving; hence from understanding a problem, a tailored approach can substantially help solve it.

To predict the performance of control parameter configurations in unseen environments, this thesis crafts an intelligent generalizable framework leveraging machine learning classification and quantitative characteristics about the problem in question. The proposed parameter performance classifier (PPC) framework is extensively explored by training 84 high-accuracy classifiers comprised of multiple sampling methods, fitness types, and binning strategies. Furthermore, the novel framework is utilized in constructing a new parameter-free particle swarm optimization (PSO) variant called PPC-PSO that effectively eliminates the computational cost of parameter tuning, yields competitive performance amongst other leading methodologies across 99 benchmark functions, and is highly accessible to researchers and practitioners. The success of PPC-PSO shows excellent promise for the applicability of the PPC framework in making many more robust parameter-free meta-heuristic algorithms in the future with incredible generalization capabilities.