Artificial Intelligence has extended its realms to almost all fields, and we find its applications in nearly all spheres of life. In several computational tasks, AI systems have even outperformed humans, marking significant strides in technological advancement. But AI systems, just like humans, also make mistakes and errors, especially when subjected to an unseen scenario. This happens as AI depends on the amount of data and computation available. Consequently, ongoing research strives to mitigate these limitations, enhancing AI’s adaptability and robustness across diverse situations.
Nevertheless, AI systems can beat professional players in tricky and challenging games such as chess, poker, etc. These AI systems use Reinforcement Learning, which can enable them to learn from trial and error and gain more knowledge. But despite these AI chess systems being robust and powerful, they still need to attain the optimal level yet. They are prone to adversarial attacks and may also hallucinate.
To tackle this issue, the researchers at Google DeepMind have developed a new work, Diversifying AI: Towards Creative Chess with AlphaZero. They conducted extensive research to explore how artificial intelligence can leverage the creative problem-solving mechanisms observed in human intelligence. They devised a way to train a group of different high-quality AI agents. They represented each player by a latent variable. Each agent is based on AlphaZero (AZ), but they are brought together using a special structure(latent) that helps them work as a team. AlphaZero can play logical games such as chess and shogi from scratch. AlphaZero can play these even if it has no prior knowledge of them. It can also play creative moves and can beat human professionals too.
To solve chess puzzles, the researchers set AlphaZero-based Agent AZdb in a face-off against a more uniform AZ group. They found that AZdb outperformed the AZ group by solving the toughest puzzles, including the challenging Penrose positions, at a rate twice as fast. A central aspect of their study was to determine if this amalgamation of AI systems could generate a higher quantity of innovative ideas as a collective entity in comparison to the output of a single AI system.
The researchers emphasized that AI can improve its accuracy from creative problem-solving mechanisms. The researchers tried to focus on AI?s capability to solve problems creatively. They defined this term as searching for an original and previously unknown solution to a problem.
The study demonstrated that AZdb’s diverse approaches to playing chess led to improved puzzle-solving abilities as a collective, surpassing the performance of a more uniform team. Analysis of their chess games revealed that AZdb participants exhibited specialization in various openings.
The researchers concluded that despite this AI system performing well, there is still a gap between humans and machine intelligence. Still, the researchers hope this work might serve as a foundation for further research.