Robots can defeat humans in multiplayer games.

2019-11-22

Recently, researchers at the Massachusetts Institute of Technology (MIT) have developed an AI-powered robot called DeepRole that can defeat human players in complex multiplayer online games. Many game robots are created to keep up with human players, and some examples can attest to this point: DeepMind's AlphaGo made headlines in 2016 for defeating professional Go players, and earlier this year, a team from Carnegie Mellon University developed the world's first robot capable of beating professional players in multiplayer poker games. However, in these games, the robots "understand" their opponents from the beginning. What sets DeepRole apart is that it has no prior knowledge of its opponents and in multiplayer online games, players' moves are uncertain and there are no fixed rules. DeepRole has no previous examples to refer to, but even under these challenging circumstances, it still outperforms human players. The name of the multiplayer game is "Resistance: Avalon," in which players need to deduce their teammates' secret roles while hiding their own. Researchers had DeepRole play over 4,000 rounds against human players, and whether as a teammate or opponent, DeepRole consistently outperformed human players. This game is similar to "Werewolf," and the key to winning is to deduce who the resistance members or spies are and vote for your collaborators, which is much more complex than chess and poker. DeepRole uses a game planning algorithm called "Counterfactual Regret Minimization" (CFR) combined with deductive reasoning to create a decision at every point in the game, describing the possible actions each player may take in the future. It then conducts billions of game simulations, noting which actions increase or decrease the chances of winning, and continuously adjusts its strategy. Ultimately, it formulates an optimal strategy that can hold its own against any opponent, even in the worst-case scenario. The robot cannot communicate smoothly with other players, which is usually a crucial part of the "Avalon" game. In response to this, the researchers stated, "It turns out that our robot can participate in team collaboration simply by observing players' actions. This is interesting because people might think that such a game requires complex communication, but the robot doesn't need it at all." In the future, researchers may enable the robot to communicate with human players through simple text during the game, such as stating whether a player is good or bad, acting as a "wise one" to simplify the game, or to confuse and increase the difficulty.

RECENT AI NEWS

RECENT AI TOOLS