Artificial intelligence (AI) has demonstrated remarkable proficiency in conquering complex games, marking significant milestones in the field. However, these models are typically designed for specific challenges. A breakthrough DeepMind algorithm, capable of handling a broader spectrum of games, is seen as a stride toward achieving more generalized AI, according to its creators.
The use of games as a benchmark for AI prowess has a rich history. Milestones like IBM’s Deep Blue defeating chess world champion Garry Kasparov in 1997 and DeepMind’s AlphaGo triumphing over Go champion Lee Sedol in 2016 fueled enthusiasm about AI’s potential. AlphaZero, a subsequent model from DeepMind, mastered various games, including chess and shogi, but primarily excelled in perfect information games.
Perfect information games, such as chess and Go, involve complete visibility of all game details to both players. In contrast, imperfect information games, like poker, conceal certain details from opponents. While models have emerged to beat professionals in imperfect information games, they often employ different approaches than those used in perfect information games.
DeepMind’s researchers have now amalgamated elements of both approaches, creating a model named “Student of Games.” This model exhibits proficiency in chess, Go, and poker, marking a significant breakthrough. The researchers believe that this achievement could accelerate the development of more general AI algorithms capable of solving a diverse array of tasks.
Perfect information game AI often relies on tree search, exploring potential sequences of moves. Imperfect information games usually involve game theory, using mathematical models for rational solutions. DeepMind’s new algorithm combines tree search, self-play, and game theory, enabling it to handle both perfect and imperfect information games. The researchers reported that the algorithm outperformed the best publicly available poker-playing AI, Slumbot, and played chess and Go at a professional human level.
While specialized algorithms like AlphaZero still excel in specific games, the ability to address both perfect and imperfect information games signifies a crucial step toward more generalized AI algorithms applicable to diverse environments. Despite the significance of this achievement, caution is urged not to extrapolate excessively, given the controlled nature of game environments compared to the complexities of the real world. Nevertheless, this breakthrough provides a blueprint for future models with enhanced capabilities and broader applications.
By Impact Lab