A team of scientists, led by Max Jaderberg, have taught artificial intelligence to play a popular first-person multiplayer video game (Quake III) with skills similar to those of humans. This is a previously impossible task for this type of technology due to the amount of events and tactics involved, many of them without rules or logical chronologies.
The AI demonstrated a strange ability to develop and use high-level strategies learned independently to compete against human players and cooperate with their team (also formed by humans) in the game environment.
To reach this result, the Jaderberg team trained AI with the reinforcement learning system (RL). Basically it is about teaching a software the most appropriate actions to reach a goal.
The method is not new, since it had previously been used on machines that learned to play Go and chess. The difference is that in these cases they are two players who each have their turns in the game, while in video games, there is collaboration between several players and there are no turns to respect and the reaction to the movement of the opponent must be instantaneous if they want to win.
In contrast to previous studies, where AI received information about the game environment or the status of other players, the Jaderberg team gave each machine its own time to learn to play and create its individual strategy using only what it could “see” (pixels) and the game score. And then he faced several AIs in thousands of randomly generated games. According to the study authors, published in Science, over time, AI independently developed high-level strategies, not very different from those used by expert human players. What’s more, in games with human players, the agents outpaced the humans, even when the reaction times were similar. As if that were not enough, they could also train and cooperate in computer and human teams trained at the time.