Computer scientists at Google DeepMind have developed a new approach to AI systems by incorporating diversity and collaboration, resulting in improved problem-solving abilities.
When the COVID-19 pandemic forced people into isolation, computer scientist Tom Zahavy rediscovered his passion for chess. As he delved deeper into the game, Zahavy became intrigued by the limitations of chess programs and their inability to solve complex puzzles. Inspired by mathematician Sir Roger Penrose’s puzzles, Zahavy and his colleagues at Google DeepMind set out to develop a new approach to AI systems that could tackle tough problems by incorporating diversity and collaboration. Their findings not only revolutionize chess programs but also offer insights into the potential of AI systems to think creatively and solve a wide range of complex problems.
The Limitations of Chess Programs and the Power of Chess Puzzles
Chess has long been a testing ground for artificial intelligence, with programs like Deep Blue defeating human grandmasters. However, puzzles designed by Penrose and others revealed that chess programs struggled to solve complex and contrived positions. Zahavy’s interest in these puzzles led him to explore the potential of AI systems to tackle challenging problems that go beyond winning a game of chess.
The Birth of a Diversified AI System
Zahavy and his colleagues developed a new AI system that combined multiple decision-making AI systems, each optimized for different strategies. By weaving together up to 10 diverse systems, including DeepMind’s powerful chess program AlphaZero, the team created a program that showed enhanced problem-solving abilities and creativity. The system could seamlessly switch between different approaches when faced with a roadblock, leading to improved performance in solving Penrose’s puzzles.
The Power of Diversity in AI Systems
The use of diverse AI systems provides a range of perspectives and strategies to tackle complex problems. Allison Liemhetcharat, a computer scientist at DoorDash, explains that a population of agents increases the probability of finding solutions within the domain of at least one of the agents. This approach allows AI systems to efficiently tackle hard problems beyond the game of chess.
Chasing Failures and Embracing Creativity
Zahavy’s research revealed that traditional AI systems based on reinforcement learning often struggle with generalization and fail to recognize failure. By introducing diversity and rewarding creativity, Zahavy’s diversified AlphaZero system learned to explore new strategies, resulting in improved problem-solving abilities. The ability to fail and recognize failure is crucial for creative problem-solving, as it allows AI systems to consider alternative approaches and solutions.
The Implications Beyond Chess
The success of the diversified AlphaZero system extends beyond the game of chess. Antoine Cully, an AI researcher at Imperial College London, highlights that this approach mirrors human brainstorming sessions and leads to creative and effective solutions. Julian Togelius, a computer scientist at New York University, emphasizes the potential of diversity in training physical systems, drug discovery, and stock trading strategies. The diversified approach opens doors to a wide range of applications and problem-solving scenarios.
Conclusion:
The research conducted by Zahavy and his colleagues at Google DeepMind demonstrates the power of diversity and collaboration in AI systems. By incorporating multiple decision-making systems optimized for different strategies, the diversified AlphaZero program showcased enhanced problem-solving abilities and creativity. This approach not only revolutionizes chess programs but also offers insights into the potential of AI systems to think creatively and solve complex problems in various domains. As AI systems continue to evolve, embracing diversity and rewarding creativity may pave the way for groundbreaking advancements in artificial intelligence.
Leave a Reply