Predictions for the Future of Optimisation Research

Abstract

The global effort to find good optimisation methods is an evolutionary algorithm (note “is”, not “is analogous to”). A team’s research effort is an individual, or ‘chromosome’, and peer review is a (very) noisy and multiobjective fitness metric. Genetic operators (new directions and ideas for research efforts) are guided partly by discussions at conferences, maybe even sometimes guided by plenary talks. In this talk I will predict what kind of research in optimisation I expect to have the highest fitness scores in the next several years. They will be, mainly, combinations of learning and optimisation that are theoretically justified, or simply justified by their excellent results, and they will be works concerned with generating algorithms that quickly solve a distribution of problem instances, rather than one at a time. These combinations of learning and optimisation will be informed by the (slow) realisation that several separate studies, emerging from different subfields, are converging on very similar styles of approach. A particular point is that, in this way, we see that theoretical work on optimisation is slowly beginning to understand aspects of methods used by nature. Finally, these are predictions, and certainly not prescriptions. The overarching evolutionary process that we serve cannot succeed unless lots of diversity is maintained. So, please ignore what I say.