Diffuse versus true coevolution in a physics-based world

We compare two types of coevolutionary tournaments, true and diffuse, in contests using a general-purpose, physics-based simulator. Previous work in coevolving agents has used true coevolution and found that populations tend to enter mediocre states. One hypothesis for alleviating these problems is to use diffuse coevolution. Our results show that agents evaluated with diffuse tournaments are more generalized than those evaluated with true tournaments.