Learning robust build-orders from previous opponents with coevolution
Christopher A. Ballinger, Sushil J. Louis · 2014
Learning robust, winning strategies from previous opponents in Real-Time Strategy games presents a challenging problem. In our paper, we investigate this problem by using case-injection into the teachset and population of a coevolutionary algorithm. Specifically, we take several winning build-orders we created through hand-tuning or coevolution and periodically inject them into the coevolutionary population and/or teachset. We compare the build-orders produced by three different case-injection methods to the robustness of build-orders produced without case-injection and measure their similarity to all the injected cases. Our results show that case injection works well with a coevolutionary algorithm. Case injection into the population quickly influences the strategies to play like some of the injected cases, without losing robustness. This work informs our ongoing research on finding robust build-orders for real-time strategy games.