If you’re at all familiar with chess, you know that AI has long been dominant in the game compared to humans. Most Grandmasters and Super Grandmasters of chess now utilize Artificial Intelligence to analyze their own games to find mistakes and areas to improve. But playing and practicing against AI is very different than playing against a person, especially for lower rated players (newer to the game.)
This is because strong chess engines spot difficult combinations very easily and engines have no mating patterns or positional set-ups to rely on or to hold them back. A strong player is inclined to move toward these patterns and set-ups while an engine will be able to spot weird un-positional moves that don’t make much sense to a human player. Engines have no preference for normal or known positions, if it sees something better, it will play it. This can lead to a lot of odd looking moves. Every beginner learns to simplify the board and remove extra pieces to get to a less complicated ending. Engines do not worry about this as they never lose track or blunder pieces in the complexity. Though engines do have some issues with strong defensive positions (fortress) and that will be able to show best between a human player and engine. In time, it can be expected that engines will be able to work through these positions as well.
Introducing Maia Chess, a human-like neural chess engine designed to play the human move. Maia is an engine that learns from human games instead of self-play games, with the goal of making the most human move and not most optimal. Leela and Stockfish are other engines that try to match human moves, but they only reach 43% and 38% respectively. Maia on the other hand predicts up to 53% of player moves and as result is the most natural, human-like chess engine to date. Maia was developed using code adapted from Leela which in turn is an open-source clone of Alpha Zero, a revolutionary AI program created by DeepMind.
There are 9 trained versions of Maia, one for each Elo milestone between 1100 and 1900. Each version is only trained on human players of the same ranking, so Maia 1900 is trained only on games of 1900-rated players. Each version has learned from 12 million human games and is still learning by playing real players at Lichess, a popular online chess server.
In current work, Maia is being pushed to predict the moves a particular human player would make. By starting with a base model and training the model on a specific player’s games they are able to personalize the engine. This has given the engine the highest results and raises the accuracy to 65% for the particular player.
What does it mean?
Chess is a great way to develop and study AI. It’s popular, has well-defined rules, and it has not yet been fully solved. Many poeple know the game and AI researchers use it as a “model system” to study new ideas or techniques.
Chess players have their own playing styles, so it can be more difficult to predict their moves over the best possible move due to the astronomical amount of potential chess positions. Even the same player may make a difference move on a position that they have already seen for any various number of reasons.
Having a chess engine that plays the same way a human does at a specific Elo is a great way for players to practice and improve their own play. A normal chess engine at a specific Elo will still play like a chess engine but just make more intentional mistakes. Maia will play like a real human player and will give players a more genuine experience.
When using a particular players games for Maia to analyze and learn from, it raises the prediction rates even higher. So for more professional players that want to practice against another competitor, it will be possible for them to compile all their online games into the engine and have Maia imitate them. Allowing for a more personal training session than could be given in any other way.
Line of Research
This line of research into an AI that tries to be more human can eventually be used elsewhere as technology improves. Being able to have training/teaching software with an AI that understands which parts humans have most trouble with and what solutions have the best results to raise proficiency. For AI systems to be able to behave in more humanlike ways will allow for more people to understand how AI systems work. It can make it more acceptable for AI systems to be accepted in the future. Instead of the idea that AI systems will replace people, it will be a way for AI to be used to augment humans
Garry Kasparov, the former world chess champion, lost to the IBM’s Deep Blue over 20 years ago and he believes that AI can make us ‘more human.’ Which is definitely a thought provoking statement and would love to hear what others think about it.