Researchers at DeepMind have used AlphaZero to examine the game of chess, looking for areas where improvements can be made. The AI was used to play around with a variety of different rules alongside help from Vladimir Kramnik – ex world champion of chess – to come up with new variations of the game.
AlphaZero is an adaptive machine learning system, which has the ability to learn new rules from scratch. This system was used to test outcomes from 9 variants of the board game – these variants made with the help of Vladimir Kramnik.
Chess is a board game which has evolved considerably across the centuries, new variants having been added to both improve issues and add new elements into the competitions. Changing the rules of chess has tremendous impacts on the dynamics of the game, including strategy and playability.
In the past, the results of integrating new chess variants were only able to be obtained over an extended period of time, watching how human players adapt to such alterations. What AlphaZero, and the DeepMind research overall, has done is achieve observations into how new rules would be taken to more effectively than ever before.
DeepMind’s team of researchers, an AI solutions provider, set out to find the best revisions they could make to this classic board game, using AI to help achieve this.
Tens of thousands of games were played for each of the 9 variants – all played by AlphaZero against itself. Once this was complete, researchers involved in the study, alongside Kramnik, assessed what real-life, human games would look like if playing such variants of the game, using this to help them discern if certain rules could help in improving the game.
DeepMind researchers comment: “Training an AlphaZero model under these rules changes helped us effectively stimulate decades of human play in a matter of hours”
Utilising this AlphaZero model, the research team further commented, helped them “answer the ‘what if’ question: what the play would potentially look like under developed theory in each chess variant.”
The system is reported to have played 10,000 games in each of the 9 variants at a second per move, whilst another 1,000 games of each variant were played again at a minute per move. The researchers looked at numerous different factors to objectively determine how such changes impacted the quality of the game.
The researchers found that most of the variants increased the number of potentially decisive results. In addition to this, the time controls were also found to have an impact on the decisiveness of the game – those played at a second per move being less likely to result in draws than when played at a minute per move.
The abstract for the DeepMind publication states:
“By learning near-optimal strategies for each variant with AlphaZero, we determine what games between strong human players might look like if these variants were adopted. Qualitatively, several variants are very dynamic. An analytic comparison show that pieces are valued differently between variants, and that some variants are move decisive than classical chess. Our findings demonstrate the rich possibilities that lie beyond the rules of modern chess.”