The team focused on a capture the flag mode, one in which the map changes from match to match. Its AI agents had to learn general strategies to be able to adapt to each new map, something humans do easily. The agents also needed to both cooperate with team members as well as compete against the opposite team, and be able to adjust to different enemy play styles.
“Our agents must learn from scratch how to see, act, cooperate, and compete in unseen environments, all from a single reinforcement signal per match: whether their team won or not,” wrote the researchers in a blog post. They trained a population of AI-powered agents that learn by playing the game, much like we do. Each individual agent is motivated by an internal reward signal, which reinforces them for attaining their goals, like capturing the flag. The agents play each other and human opponents in both fast and slow matches to better improve their memory and stay consistent in their behaviors. The researchers found that the AI agents win more often than humans, which makes sense, but that they also are more collaborative than people. The agents also learned human-like behaviors like following teammates and camping in enemy bases.
If nothing else, the researchers say, AI like this could be extended to more complex games like StarCraft II and Dota 2, leading to both allies and opponents who play much more like other humans. It’s not hard to envision a future where eSports teams use the tech to improve their outcomes, too.