Finding robust strategies to defeat specific opponents using case-injected coevolution

Finding robust solutions that are also capable of beating specific opponents presents a challenging problem. This paper investigates solving this problem by using case-injection with a coevolutionary algorithm. Specifically, we recorded winning strategies used by a human player against a coevolved strategy and then injected the player's strategies into the coevolutionary teachset. We compare the strategies produced by case-injected coevolution to strategies produced by a genetic algorithm that only evaluated against the player's strategies. In this paper, our results show that genetic algorithms do not work well against sufficiently difficult opponents. However, coevolution eventually learns to defeat these opponents by first bootstrapping strategies that work well in general, which drives the population closer to strategies that can defeat the challenging opponent. This work informs our research on finding robust real-time strategy game players that also defeat specific opponents.