Learning to Optimise: A Perspective from Darwinian Evolution

Speaker: A/Prof. Yi Mei
Summary:
Solving complex optimisation problems is hard, especially when facing the real-world challenges such as large problem size, dynamic/uncertain environment, and multiple conflicting objectives. In recent years, “learn to optimise” becomes a trendy research topic, aiming to employ machine learning techniques to automatically design optimisation algorithms.
This talk focuses on a perspective from Darwinian evolution for automatically designing optimisation methods, i.e., heuristics. Briefly speaking, each possible heuristic is an individual in a population. They evolve by crossing over with each other and mutation, and the ones with better fitness are more likely to survive to the next generation. This way, we can effectively and gradient-free search in a broad, flexible, and non-differentiable space of heuristics, optimising multiple objectives simultaneously.
We will show our examples of using genetic programming to learning policies for (1) dynamic job shop scheduling; (2) real-time ambulance dispatching; (3) traffic signal control; and (4) air traffic flow management. We will also show our recent advances on multi-objective learning, interpretability of learned policy, and knowledge transfer between different optimisation scenarios. Last but not least, some discussions on genetic programming vs reinforcement learning will be provided.
Biography:
Dr. Yi Mei is an Associate Professor and the Associate Dean (Research) at the Faculty of Engineering, Victoria University of Wellington, New Zealand. He received his BSc and PhD degrees from the University of Science and Technology of China in 2005 and 2010, respectively. His research interests include evolutionary computation for combinatorial optimisation, genetic programming, automatic algorithm design, explainable AI, multi-objective optimisation, transfer/multitask learning and optimisation.
Yi has over 200 fully refereed publications, including the top journals in EC and Operations Research (OR) such as IEEE Transactions on Evolutionary Computation, IEEE Transactions on Cybernetics, European Journal of Operational Research, ACM Transactions on Mathematical Software, and top EC conferences (GECCO). He received an IEEE Transactions on Evolutionary Computation Outstanding Paper Award, two GECCO Best Paper Awards, a GECCO Human Competitive Award, and an EuroGP Best Paper Award. He is an Associate Editor of IEEE Transactions on Evolutionary Computation, IEEE Transactions on Artificial Intelligence, and of five other international journals. He is the Chair of the IEEE Taskforce on Evolutionary Scheduling and Combinatorial Optimisation and the Chair of the New Zealand Central Section. He is a Fellow of Engineering New Zealand and an IEEE Senior Member.