DeepMind’s AlphaZero AI Helps Design New Chess Rules
Overview of the paper “Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess” by N Tomašev et al.
Off-late we have been seeing quite a few papers exploring the process of designing a game as something that goes hand in hand with learning to play that game. This means that after generating a game level, we can quickly assess how easy or difficult it would be for humans to play the game by training an AI to complete that level. This makes it much easier to design games that are the right balance between fun, engaging and challenging at the same time.
Today I want to share a paper that is based on a similar concept. It is titled “Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess” and is published by researchers from DeepMind along with the world chess champion, Vladimir Kramnik!
Let’s say we want to take the classic game of Chess and introduce some new rules to this game, for example, allowing the pawn to move side-ways or back-ways, or disallowing castling, etc. Now, in order to ensure that these new rules do not ruin the balance of the game, and the competition between the first mover and the second mover still remains fair with no unfair advantage, we need to run many tests for analysis.
Many games that look fun to start with become extremely boring once humans master it, take my childhood favorite game Tic-Tac-Toe for example. If both players have sufficient expertise in the game, every single game will end in a draw, and the game is no longer fun to play. This is why introducing new rules to games like chess could ruin the game if expert humans master the game, making it very difficult to quickly and easily assess how new rules will affect the game of chess over a long period of time.
This is where DeepMind is using their AlphaZero AI to quickly train expert game-playing agents on the chess games with new rules, so that we can pit them against one another and simulate thousand of matches quickly to see the average outcome.
Assessing New Rules
For the given two figures below, we have different rule-changes to the classical version of chess on the vertical axis and the number of wins by the black and white players on the horizontal axis.
This type of analysis ensures that one player does not have unfair advantage over the other over multiple games where both the players are experts of the game. On the right is a similar analysis but where the time required to make a move has been increased from one second to one minute. This gives AlphaZero much more time to explore more strategies before making a move, hence majority of the games end up in a draw. This is very good because when these games will be played by humans, it shows that those who have the ability to think deeper will come off as better players, thereby making sure skills of a player actually matter in the game.
This is an interesting way to think about game design, especially for those multiplayer games where we pit one human against another. This kind of work will make it much much easier to come up with new game ideas and quickly test them out to see if people will actually enjoy playing them or get frustrated and stop playing it. Super interesting!
Useful Links
Thank you for reading. If you liked this article, you may follow more of my work on Medium, GitHub, or subscribe to my YouTube channel.
