Skip to main content
  • CIS
    Members: Free
    IEEE Members: Free
    Non-members: Free
    Length: 01:09:24
28 Jun 2021

Xin Yao (The University of Birmingham, UK),
Abstract: Co-evolution has been an old research topic in evolutionary computation, but never very popular. It seems to be a last resort, instead of the first choice, when tackling difficult problems. For example, in early days of co-evolutionary learning of game-playing strategies, e.g., strategies for playing iterated prisoner's dilemma games, we faced a hard learning problem where there is no training data and zero knowledge about the game-playing. Yet we have to learn to play the game. Somewhat unexpectedly, co-evolution rose to the challenge and evolved expert level strategies from a set of random strategies with no training data and zero knowledge.Then research on co-evolution got quiet until new challenges come, e.g., the challenge of scalability for evolutionary optimisation. Evolutionary algorithms (EAs) have been successful in solving optimisation problems in various domains, from car engine design to agricultural land use optimisation. However, the number of real-valued decision variables (i.e., problem dimension) that could be handled by EAs is relatively small, up to a couple of hundreds in most cases. Large-scale global optimisation (LSGO) has been a real challenge. Although parallel and distributed computing could help to alleviate the situation a little, the amount of hardware needed can never catch up the growth of optimisation problem. Seemingly as a last resort, the good old co-evolution was brought back to the front stage and used to tackle LSGO. It has now become one of the most effective approaches to LSGO in the literature. Co-evolution has grown out of laboratories. Similar ideas to those used in tackling LSGO have been used in design optimisation in concurrent engineering. Of course, not only can co-evolution be used for numerical global optimisation, it can also be used for combinatorial optimisation, albeit as a last resort again. It is the last resort because the first reaction people would have when hearing combinatorial optimisation would be classical computer algorithms and mathematical programming methods. Only when they fail would one start exploring heuristics and then meta-heuristics. When they also fail, one might think of co-evolution, as a special technique for meta-heuristics. In this talk, we would like to argue, through the previous examples, that co-evolution should probably be treated as a problem-solving approach, rather than just some algorithms or techniques. It encourages us to approach hard problems differently. Finally, it might be worth mentioning that co-evolution theories are also making progresses although at a rather slow pace.