Alphabet’s DeepMind AI can beat humans at multiplayer games — with ‘team spirit’

• DeepMind, owned by Alphabet and Google’s sister company, has trained its artificial intelligence (AI) agents to beat humans at multiplayer games.
• The AI ‘agents’ aren’t just better gamers but have more team spirit as well.
• Even when the efficiency of the AI was ‘lowered’ to human levels — they still had a much higher win rate than humans.

Artificial intelligence (AI) has been beating humans at games for a while, but their dominance was limited to one-on-one games.

DeepMind, owned by Google’s parent company Alphabet, now has automated ‘agents’ that can beat humans at first-person multiplayer shooter games as well. In fact, AI is so good that it beats gamers even with its efficiency was ‘lowered’ to human levels.

DeepMind ran their experiment on Quake III Arena Capture the Flag and published their results in Science.

Quake III Capture the Flag is a simple game with complex dynamics. There are two teams, each with the objective of capturing the opponent’s flag while protecting their own. If you tag a competing player, they are sent back to the starting point.

The team which captures the most flags in 5 minutes — wins.

Even after 12 hours of practice, the human game testers were only able to win 25% of games against the agent team.


Rise of the machines

The reason that AI wasn’t able to master multiplayer gaming environments till now was because they are more complex than one-on-one game scenarios.

Multiplayer games need bots — or ‘agents’ — to master strategy, tactical understanding and coordination for team play.

In DeepMind’s case, the team used the game Quake III Arena Capture the Flag to train their agents. The first step was to not train each AI agent individually but to treat the team as a single entity instead.

DeepMind calls this ‘multi-agent learning’ where many individual agents will act independently but have to learn how to interact and cooperate with other agents in the game.

The agents, in turn, learn how to turn data that they’re collecting from the game’s raw pixel into action. The team also integrated the functionality that the map layout would change for each match.

So, the agents aren’t clued in on the rules of the game — they learn on their own and develop their own initiative on how to play best.

The agent that’s able to master the game is called ‘For The Win’ (FTW) agent who can play the game at a very high skill level.

AI isn’t just better, it has more team spirit

DeepMind ran a tournament that included 40 human players where the matchups between human and AI players were random — playing on the same team as well as playing against each other.

The resulting data showed that FTW agents had a much higher win-rate as compared to human players and even their strategy application was superior. In fact, they even showed more team spirit than their human counterparts.