Despite games often being used as a testbed for new computational intelligence techniques, the majority of artificial intelligence in commercial games is scripted. This means that the computer agents are non-adaptive and often inherently exploitable because of it. In this paper, we describe a learning system designed for team strategy development in a real time multi-agent domain. We test our system in the game of Pacman, evolving adaptive strategies for the ghosts in simulated real time against a competent Pacman player. Our agents (the ghosts) are controlled by neural networks, whose weights and structure are incrementally evolved via an implementation of the NEAT (Neuro-Evolution of Augmenting Topologies) algorithm. We demonstrate the design and successful implementation of this system by evolving a number of interesting and complex team strategies that outperform the ghosts' strategies of the original arcade version of the game.