Neuroevolution for Realtime Strategy Game Micromanagement
AdvisorLouis, Sushil J.
Computer Science and Engineering
AltmetricsView Usage Statistics
We explore the use of neuroevolution to evolve control tactics for groups of units in real-time strategy (RTS) games. RTS games have units with different characteristics in weaponry, movement and abilities, which makes effective coordination of units in real time a challenging problem. We focus on micro control, which requires quick planning and decision making during skirmishes between two groups of opposing units. We created a new representation that can effectively map game state to the neural input domain and evolved the network using neuroevolution of augmenting topologies (NEAT) to control movement and attack commands for each unit. We applied this approach to a group of ranged units skirmishing with a group of melee units in StarCraft II, an RTS game. The evolved neural networks performed well and lead to kiting behavior for the ranged units, which is a common tactic used by professional players in ranged versus melee skirmishes in RTS games. The evolved neural network also generalized well to other starting positions and numbers of units. We then move to a more natural two-objective fitness evaluation that maximizes damage done and minimizes damage received and show that we can use NEAT to evolve a variety of high-performance micro tactics along the Pareto front. Finally, we show that we can generate even more robust micro tactics by evolving against the best performing networks trained using normal evolution. We believe these results indicate that our neuroevolutionary approach can generate effective coordinated distributed control for agents in RTS games and in other distributed control environments.