Regresar

Towards training swarms for game AI

Abstract:

This paper presents a game design pattern that can inspire the creation of new games featuring swarms of hostile non-player characters. It is driven by reinforcement learning through insane-Q: an on-l[in]e, [sa]mpling-based, [n]on-[e]pisodic [Q]-Learning variant. To unify the swarm’s behaviors, individuals query a shared insane-Q instance. To unchain the individuals from behaving identically—that is, to create sub-classes in the swarm—the data they feed to insane-Q is entangled with a programmable handle, which enables individuals to, indirectly, manipulate how insane-Q interprets its shared learning. As a test bed, a challenging racing mini-game was created through this pattern. It pits players against a reckless stampede of eighty four autonomous driving carts staring two sub-classes: greedy carts always take the path with the least amount of obstacles, while lazy carts always take the shortest path towards their next destination.