Noise tolerance for real-time evolutionary learning of cooperative predator-prey strategies

Learning team-based strategies in real-time is a difficult task, much more so in the presence of noise. In our previous work in the Prey and Predators domain we introduced an algorithm capable of evolving cooperative team strategies in real-time using fitness evaluations against a perfect opponent model. This paper continues our work within the same domain, training a team of predators to capture a prey. We investigate the effect of varying degrees of opponent model noise in our learning system. In the presence of and in the effort to mitigate the effects of such noise we present modifications to our baseline system in the forms of Rescaled Mutation, Conservative Replacement and a combination of the two techniques. The results of the modifications are extremely promising. The combined approach in particular demonstrates a vast improvement and decreased variance in the performance of our team of predators in the presence of opponent model noise. Additionally, the noise-mitigating strategies employed do not adversely affect the performance of the real-time team learning system in the absence of noise.