There is a long tradition of using games as AI performance indicators. Search and learning-based approaches performed well in various perfect information games, while game theory-based methods performed well in a few imperfect information poker variations. By combining directed search, self-play learning, and game-theoretic reasoning, the AI researchers from EquiLibre Technologies, Sony AI, Amii and Midjourney, working with Google’s DeepMind project, propose Student of Games, a general-purpose algorithm that unifies earlier efforts. With its high empirical performance in big perfect and imperfect information games, this algorithm is a significant step toward developing universal algorithms applicable in any setting.
**What is SoG – “Student Of Games”?**
Combining search, learning, and game-theoretic analysis into a single algorithm, SoG has many practical applications. SoG comprises a GT-CFR technique for learning CVPNs and sound self-play. In particular, SoG is a reliable algorithm for optimal and suboptimal information games: SoG is guaranteed to generate a better approximation of minimax-optimal techniques as computer resources improve.
**Why is SoG so effective?**
SoG employs a technique called growing-tree counterfactual regret minimization (GT-CFR), which is a form of local search that may be performed at any time and involves the non-uniform construction of subgames to increase the weight of the subgames with which the most important future states are associated.
**Summary of Algorithms**
The SoG method uses acoustic self-play to instruct the agent. GT-CFR is a two-stage process that begins with the present public state and ends with a mature tree. Training data for the value and policy networks is generated throughout the self-play process.
The research team believes it has the potential to thrive at other sorts of games due to its ability to teach itself how to play nearly any game, and it has already beaten rival AI systems and humans at Go, chess, Scotland Yard, and Texas Hold ’em poker.