Artificial Intelligence (AI) systems have demonstrated great potential to achieve superhuman performance in a wide range of fields, even win human professional players on complicated games such as chess, Go, poker, and Atari. Nevertheless, these AI systems are still struggling to generalize to new sceneries.
In a new paper Diversifying AI: Towards Creative Chess with AlphaZero, a Google DeepMind research team explores whether artificial intelligence can benefit from creative problem-solving mechanisms identified in human intelligence while pushing to the limits of its computational rationality, and they also deliver AlphaZero-based Agent AZdb, which achieves more creative decision making and strong performance in playing Chess.
This work focus on AI’s capability to solve problems creatively, the team defines this term as “searching for an original and previously unknown solution to a problem”. In particular, they aim at exploring whether a team of diverse chess agents can be more creative then a single super human AI.
The researchers believe that AI can benefit from creative-problem solving mechanisms based on their acknowledge that Reinforcement Learning (RL) agents can solve any problem via trial and error.
To test their hypothesis, they train a league of high-quality, diverse agents, each of them is built upon AlphaZero (AZ) but are combined together by applying a latent-conditioned architecture. Specifically, each player in the league is represented via a latent variable. To encourage diversity, the researchers adopt behavioral and response diversity techniques, with intrinsic motivation to boost behavioral diversity and a matchmaker that samples opponent for each players to improve response diversity.
In their empirical study, the team compared AZdb with a more homogeneous AZ team on solving chess puzzles. Notably, AZdb beats AZ team twice the solve rate on the most challenging puzzles from the data sets, including the challenging Penrose positions.
Overall, this work vindicates that diversity bonuses emerge in teams of AI agents. Despites that there is still a gap between human and machine intelligence, the team hopes their work can serve as foundation to encourage more research on bridging this gap.
The paper Diversifying AI: Towards Creative Chess with AlphaZero on arXiv.
Author: Hecate He | Editor: Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.