In 2020, during the global pandemic, Tom Zakhavi, a computer scientist, reignited his passion for chess. Influenced by Harry Kasparov’s book and popular chess shows, Zakhavi shifted his focus from improving his own game to solving chess puzzles. These puzzles led Zakhavi and his Google Deepmind colleagues to uncover hidden limitations in chess programs.
Chess has long been a testing ground for artificial intelligence. A puzzle created by Sir Roger Penrose exposed that powerful chess programs struggle to evaluate intricate positions as accurately as experienced human players. This observation inspired Zakhavi to explore innovative problem-solving approaches using AI.
At DeepMind, Zakhavi developed a method that combines ten different AI systems, each optimized for distinct game strategies. This approach enabled the creation of a system that outperformed Alphazero, DeepMind’s renowned chess program, at solving Penrose puzzles. Success was attributed to the system’s “self-labor” agents, which swiftly switched methods if one failed to yield results.
Zakhavi’s research at DeepMind focused on deep learning principles with reinforcement, where AI systems learn through experience and receive rewards for successful actions. The AlphaZero program was trained by playing 44 million games of chess against itself, rapidly reaching a level of play surpassing that of any human. However, Zakhavi cautioned that this learning method could lead to “blind spots” where the system struggles with unfamiliar challenges.
Zakhavi’s approach advocated for the integration of diverse AI systems, each trained under varying conditions, enabling the system to generate creative solutions to complex tasks akin to human brainstorming. Testing demonstrated that this unified system could solve significantly more puzzles than Alphazero alone.
The study’s findings highlight the potential effectiveness of utilizing multiple AI systems to tackle complex challenges not only in chess but also in other domains such as drug discovery and stock market trading strategies. Zakhavi proposes that for AI systems to exhibit creative thinking, they simply need to consider a wider range of solutions. The research underscores that intelligence lies in computational capabilities and the ability to choose from a vast array of optimal strategies.