Runtime Analysis of a Co-Evolutionary Algorithm: Overcoming Negative Drift in Maximin-Optimisation

Co-evolutionary algorithms have found several applications in game-theoretic applications and optimisation problems with an adversary, particularly where the strategy space is discrete and exponentially large, and where classical game-theoretic methods fail. However, the application of co-evolutionary algorithms is difficult because they often display pathological behaviour, such as cyclic behaviour and evolutionary forgetting. These challenges have prevented the broad application of co-evolutionary algorithms.

We derive, via rigorous mathematical methods, bounds on the expected time of a simple co-evolutionary algorithm until it discovers a Maximin-solution on the discrete Bilinear problem. Despite the intransitive nature of the problem leading to a cyclic behaviour of the algorithm, we prove that the algorithm obtains the Maximin-solution in expected O(n^1.5) time. However, the algorithm quickly forgets the Maximin-solution and moves away from it. Along the way, we present new mathematical tools to compute the expected time for co-evolutionary algorithms to obtain a Maximin-solution. We are confident that these tools can help further advance runtime analysis in both co-evolutionary and evolutionary algorithms.