In games like chess, poker, and other strategic endeavors, artificial intelligence has demonstrated its ability to outperform individual players. Google DeepMind, the creator of the AI called Student of Games, sees this as a significant advancement towards achieving artificial general intelligence capable of excelling in diverse tasks.
The development of the Student of Games (SoG) model can be attributed to two key projects, as noted by Martin Schmid, a former DeepMind AI researcher now working at EquiLibre Technologies. One project involved DeepStack, an AI from the University of Alberta that bested professional casino players, while the other project was AlphaZero from DeepMind, known for its victories in chess and Go.
These projects differ in their focus, with DeepStack designed for imperfect-knowledge games like poker, where players lack full information about each other’s hands, and AlphaZero tailored for perfect information games like chess. To address both game types, Intel brought together the DeepStack team, resulting in the creation of SoG.
Schmid describes SoG as a starting “blueprint” for learning various games, improving through practice and exposure to different strategies. This adaptable model can learn and evolve by playing against itself, mastering new tactics and enhancing its gameplay across different types of games.
SoG underwent testing in games like Go, Texas Hold ‘em poker, Scotland Yard, Leduc Hold’Em Poker, and a modified version of Scotland Yard. The results showed that SoG outperformed many AI models and individual players, demonstrating its versatility and effectiveness across multiple games.
While SoG may have a slight performance trade-off compared to more specialized algorithms, it remains capable of defeating top players in the majority of games it learns. By learning from self-play and exploring various game scenarios, SoG showcases its ability to excel even in games with limited information.
Despite these advancements, Michael Rovatsos from the University of Edinburgh emphasizes that there is still significant progress needed before AI can be deemed truly intelligent, as games provide a controlled and defined environment unlike the complexities of the real world.
It is essential to recognize that the controlled nature of game environments, with precise rules and outcomes, presents a unique challenge distinct from real-world scenarios, where ambiguity and unpredictability are inherent.
