# Diversifying AI: Towards Creative Chess with AlphaZero

**Authors:** Tom Zahavy, Vivek Veeriah, Shaobo Hou, Kevin Waugh, Matthew Lai, Edouard Leurent, Nenad Tomasev, Lisa Schut, Demis Hassabis, Satinder Singh  
**arXiv:** [2308.09175](https://arxiv.org/abs/2308.09175)  
**Date:** 2023-08-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.09175v1/

## Key points

- Artificial Intelligence (AI) systems have made significant advancements and surpassed human intelligence in computational tasks.
- AI systems have limitations, including making mistakes, having blind spots, struggling to generalize, and sometimes hallucinating.
- This study explores whether AI can benefit from creative decision-making mechanisms when pushed to its limits.
- Researchers investigate if a team of diverse AI systems can outperform a single AI by generating more ideas and selecting the best ones.
- The game of chess is chosen as the research focus due to its significance in AI research.
- AlphaZero (AZ), an advanced AI system known for exceptional chess performance, is extended with a new architecture called AZ_db.
- AZ_db represents a league of agents within AlphaZero, allowing for greater diversity and idea generation using behavioral diversity techniques.
- Sub-additive planning is employed to select the most promising ideas generated by AZ_db.
- Extensive experiments show that AZ_db plays chess differently compared to a homogeneous team and solves more puzzles as a group.
- AZ_db outperforms the single-agent version of AlphaZero (AZ), solving twice as many challenging puzzles including complex positions like Penrose positions.
- Players within AZ_db specialize in specific openings, leading to an improvement of 50 Elo points over AZ alone by strategically selecting players using sub-additive planning techniques.
- Diversity bonuses emerge within teams of AI agents similar to teams of humans.
- Diversity is valuable in solving computationally hard problems and significantly enhances the performance of AI systems.

## Summary

In recent years, Artificial Intelligence (AI) systems have made significant advancements and have even surpassed human intelligence in various computational tasks. However, like humans, AI systems also have their limitations. They make mistakes, have blind spots, struggle to generalize to new situations, and sometimes even hallucinate. This study aims to explore whether AI can benefit from creative decision-making mechanisms when pushed to the limits of its computational rationality. The researchers specifically investigate whether a team of diverse AI systems can outperform a single AI in challenging tasks by generating more ideas as a group and then selecting the best ones. To conduct this investigation, they choose the game of chess as it is often considered the "drosophila" or model organism for AI research. To build upon existing AI capabilities, the researchers extend AlphaZero (AZ), an advanced AI system known for its exceptional performance in chess. They develop a new architecture called AZ_db that represents a league of agents within AlphaZero. This architecture allows for greater diversity among the agents and enables them to generate a wider range of ideas using behavioral diversity techniques. The team then employs sub-additive planning to select the most promising ideas generated by AZ_db. Through extensive experiments and evaluations, they find that AZ_db plays chess in diverse ways compared to a more homogeneous team. As a result, AZ_db is able to solve more puzzles as a group and outperforms the single-agent version of AlphaZero (AZ). Notably, AZ_db demonstrates remarkable performance by solving twice as many challenging puzzles as AZ including complex positions like Penrose positions. Additionally, when playing chess from different openings players within AZ_db specialize in specific openings. By strategically selecting a player from AZ_db for each opening using sub-additive planning techniques there is an improvement of 50 Elo points over the performance of AZ alone. These findings suggest that diversity bonuses emerge within teams of AI agents just as they do within teams of humans. The researchers conclude that diversity is a valuable asset in solving computationally hard problems and can significantly enhance the performance of AI systems.