Researches at the University of Toronto have created Maia, a chess engine that aims to unite the barriers separated by artificial intelligence and human thought itself. Other engines with similar functioning, like AlphaZero or Leela, have a conception of chess that is definitely far from the human capacity possessed by each player. On the other hand, Maia has the ability to predict the moves that would occur to a human of a certain level.
The Canadian engine learning system is similar to that of its AlphaZero and Leela counterparts. All of them have the ability to learn by themselves by watching games. Thanks to the analysis of human games online you can be able to predict the most humane move in each position. Researchers evaluate the "move matching accuracy", that is, how many times the move Maia has advanced coincides with the move finally made by the human.
There are currently 9 versions of the engine, one for each ELO between 1,100 and 1,900. For each of the versions are used items with a range below that of its version. For example, for Maia 1.500 only the games of players with an ELO lower than 1.500 are analyzed. Each one of the versions has needed about 12 million human games to capture by means of algorithms which are the main errors that these players make.
Here is an example provided by the website:
In the previous position, Stockfish, which at the moment has the capacity to discover better moves than Maia, assures that the best move is bxa6. However, none of Maia's engines up to 1,500 would make the same move, as it considers that players with ELO below 1,500 could make the mistake of not playing it.
Stockfish and Leela Zero do not have the ability to predict human movements in the same way that Maia does although, as I mentioned earlier, their potential is currently much greater. The Canadian engine was run on 500.000 positions and the main conclusion was that each version of Maia is able to predict the moves of players of its level, as the following graph shows:
Something similar happens when it comes to finding human errors. Maia is able to detect human error with a 25% probability. Errors, moreover, are more difficult to locate than hits, so this capacity has special merit and can be very useful, for example, in the preparation of an opening.
The project aims to create an engine for learning in the practice of chess. In fact, currently you can play against Maia 1.100, Maia 1.500 and Maia 1.900 in the platform Lichess, whose database is the one used by the researchers to develop their engine.
The full article on Maia, published at the 2020 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, can be read here.
Do you want to know more?
Follow us on our social networks to keep you informed.