A new breakthrough has happened in game-playing AI. Ever since the creation of the chess playing computer Deep Blue in the 1980s, board games have been a solid foundation for stress testing the complexity and the limits of computer software. But the newest AI developed by DeepMind makes a big leap forward since the board game it has managed to learn is packed with far more strategic uncertainty than chess: this AI has managed to beat several expert players at Stratego.
Deepmind's new game-playing AI
The news on this game-playing AI came from a report on Singularity Hub. The piece opens with some very important context for the development of this AI and why it is so significant. Simply put, games like chess are ones that have all of their pieces and moves out in the open with no complications or caveats. It's the same kind of simple framework that allowed another famous computer intelligence, IBM's Watson, to win at Jeopardy. By contrast, Stratego is a board game with a lot of obfuscation baked in. You don't know if the game piece you're engaging is a bomb, a certain type of unit, or the flag needed to win. Because of that, the game relies a lot more on bluffing, reverse psychology, and strategizing around these unknown quantities.
The people at DeepMind behind the new game-playing AI had a novel approach to this conundrum. They actually managed to interweave an algorithmic strategy based on game theory and model-free deep reinforcement. The result is an AI, named DeepNash, reaching an all-time top three ranking among expert Stratego players. Incidentally, a similar AI breakthrough happened recently at Meta where one of their programs made it into the top 10% of another highly strategic board game based on strategic uncertainty, Diplomacy.
According to an official blog post on the matter, the development of this game-playing AI could signal more for the field going forward.
The value of mastering Stratego goes beyond gaming. In pursuit of our mission of solving intelligence to advance science and benefit humanity, we need to build advanced AI systems that can operate in complex, real-world situations with limited information of other agents and people. Our paper shows how DeepNash can be applied in situations of uncertainty and successfully balance outcomes to help solve complex problems.
While the development of DeepNash does mark a new level of sophistication for AI, this doesn't mean that ethical dilemmas will be solved overnight. As it stands right now, AI of this nature can be used to deal with minor unexpected issues as part of the real world such as self-driving cars being able to navigate traffic better. After all, driving on an empty high way is straightforward; driving on a highway with a bunch of unpredictable human drivers is something else.
