Abstract
Coevolution is the only process known to produce human-level intelligence. Prior research demonstrates that coevolutionary learning paradigms like multi-agent reinforcement learning are fertile grounds for arms-races that yield agents capable of superhuman performance on a set of tasks. These methods, however, are computationally expensive and difficult to scale. In this thesis, we explore approaches towards scalable and efficient multi-agent learning from the perspective of the environment, algorithm, and neural representation. First, we study modifications to a foraging environment that enables agents to achieve previously-unreachable cooperative behaviors. Next, we introduce a simple method of estimating interactions between coevolving populations of agents that reduces the computational cost required to run a given generation of coevolution. Then, we explore a backpropagation-inspired indirect encoding method for neuroevolution. Finally, we integrate these methods into a single framework and demonstrate their combined effectiveness. These approaches each make progress towards reducing the wall-clock time required for multi-agent learning at scale.