Evolution of Communication and Cooperation

In the wild, spotted hyenas have been observed to chase lions away from a recent kill. This is a high risk, high reward behavior that requires significant teamwork and decision making skills. Modeling this behavior and creating algorithms that can improve evolutionarily may lead to more adaptable artificial systems for robotics and other cooperative artificial agents. Previous research has shown that having a lead or "flag bearer" hyena can significantly improve evolution. Thus, the complex social dynamics and coordination abilities required for this problem make it interesting artificial intelligence task. This also suggests that the type and encoding of the sensory inputs has a significant effect on the evolutionary trajectory and overall success at the task. Additionally, in the wild genetic diversity is driven by the migration of young males between packs, which leads to interesting evolutionary questions. To address the role of input encodings we introduce two evolutionary neural network variants, one using absolute headings as inputs/outputs and one using relative headings as inputs/outputs (headings defined relative to environmental elements). Our results show that the networks with relative inputs and outputs evolve significantly faster and result in better performance, suggesting that a critical difference is the existence of easily accessible, problem relevant, references for defining movement vectors. Our results also show that the inclusion of a leader in the team structure can improve the rate at which cooperative behaviors are evolved, but does not lead to better overall behaviors. In addition, we examine the emerging behaviors as the teams go from random behavior to a circling pattern to an aggressive charge towards the goal.