Learning Cohesive Behaviors Across Scales for Semi-Cooperative Agents

The development of automated opponents in video games has been part of game development since the very beginning of the field. The advent of modern AI approaches such as reinforcement learning has opened the door to a wide variety of flexible and adaptive AI opponents. However, challenges in producing realistic opponents persist, namely scalability and generalizability. Scalability is of particular importance when many individual opponents are required to act cohesively over long distances, but this makes learning more difficult. This paper presents a novel architecture applying graph convolutional layers in a U-net with custom pooling operators in order to achieve learning across scales. League play reinforcement learning was used to train competitive agents in a navigation mesh environment.