In this meeting of the AI Seminar Series, Dr. Mario Andrés Muñoz will give a short talk on the selection of algorithms for continuous optimisation.

Title:A decision support system for the automatic selection of algorithms in the black-box continuous optimization domain

Speaker:Dr. Mario Andrés Muñoz

When: Tuesday, 20th May 11 AM - 12 NOON

Where: Doug McDonnell-10.05

It is a non-trivial task to select the most appropriate algorithm to solve a black-box continuous optimization problem. Such problems lack of an algebraic expression; have non-calculable or useless derivatives; or exhibit uncertainty or noise. In this work, we investigate the efficacy a machine learning based decision support system for the automatic selection of algorithms. In this approach, we construct a mapping between the set of continuous optimization problems, and the set of algorithms designed to solve such problems. Three different models are evaluated. Their inputs consists of exploratory landscape analysis measures and a subset of algorithms, and the output is an algorithm recommendation.

We carried out comprehensive numerical experiments using a suite of black-box optimization problems and well-known optimization algorithms. To estimate the accuracy of the system, we implemented the hold-one-instance-out and hold-one-problem-out validation approaches. In addition, we proposed a method to automatically select algorithm subsets using concepts from voting systems theory. We also examined the effects that the size and randomness of the input sample have on the landscape measures.

Our results demonstrate that the decision support system is accurate using landscape measures calculated with a small sample size. Subsequently, we have identified a number of associated trade-offs in the system performance. However, the results indicate that the decision support system is a robust approach to the algorithm selection problem in the black-box continuous optimization domain. Although the allocated computational budget should be a few orders of magnitude larger than the cost of calculating the landscape measures.

Dr. Mario Andrés Muñoz completed his at the Mechanical Engineering Department, The University of Melbourne, in 2013. He is currently a Research Fellow at the School of Mathematical Sciences, Monash University. His work focuses on the analysis of black-box optimization problems and empirical modelling of algorithm performance. His research interests include randomized heuristics for search and optimization, intelligent control systems, and applications of machine learning in engineering.