DeepMind Technologies Ltd. researchers developed “DeepNash,” an AI programme that can play the game Stratego at an expert level. The crew discusses its innovative strategy to improve the app’s level of play in an article published in the journal Science.
Stratego is a two-player board game that is notoriously difficult to master. Each player’s aim is to capture their opponent’s flag, which is buried among their first 40 game pieces. Each game piece has a power ranking—in face-offs, higher-ranked players beat lower-ranked ones. The fact that neither player can see the marks on the opponent’s game pieces until they meet face-to-face makes the game more challenging.
Prior study has shown that the game’s complexity is greater than that of chess or go, with 10535 potential game situations. This amount of intricacy makes it incredibly difficult for computer professionals to construct Stratego-playing AI systems. The researchers used a different strategy in their new endeavour, developing an app capable of outperforming most human and other AI systems.
DeepNash, like previous AI system designs, learns to play Stratego by playing itself many times—in this instance, 5.5 billion times—equivalent to hundreds of years of human playing time. The researchers did not make it try to acquire methods from master human players or even compete against other opponents once they learnt how to play.
Instead of perfection, the researchers built an algorithm that worked toward an ideal approach for each step. The algorithm was built on game theory: a perfect system would offer DeepNash a 50/50 chance of winning on every given play, which is significantly better than humans could hope for.
Testing revealed that the team had discovered a means to enhance the chances of an AI software playing Stratego—it obtained an 84% winning record while playing 50 times on an online gaming platform, becoming one of the site’s top three players. And the human opponents were never informed that they were up against a machine.
You might also be interested in reading A new bendable, steerable gadget installed in living brains by a minimally invasive robot.
