As promised here is the first 10 game match between the Millennium ChessGenius Pro2024 and Chessnut Evo’s revised implementation of the Maia9 bot. More information about the Maia project can be found here https://maiachess.com/

In essence, the project has created nine sets of data each one targeted to a particular ELO rating and were created by training the neural networks with 12 million games for each set. The games come from human play at the chosen ELO rating. The idea behind the Maia project is zero is look ahead, and just using the data to pick the “human move” not necessarily the best move at the respective ELO level. The current highest ELO used for training the network is 1900 but as many of you are aware, the Maia does not reach this level of play and perhaps after making some reasonable moves will frequently blunder pieces and lose the game like a complete beginner and not the expected play from a human player at this level. The expectation might be that a neural network trained with 12 million human games between players rated at 1900 that the neural network would match this level in play, but trained with such data is one thing, playing at the same level is something else. 12 million games sound a large number, but dealing with the game such as chess with such a vast number of permutations and combinations then this figure is not so impressive and obviously leaves large gaps of knowledge in positions. For it to identify and avoid blundering pieces without looking ahead would surely require a magnitude higher level of data? Therefore, why not just implement a selectable ply search that takes just a few seconds? This is what Chessnut have implemented after receiving feedback from its customers such as myself. This feature as many will be aware has been implemented as a slider control allowing ply depths from 0 to 6. So we can have it play as it did before, blundering pieces or set a maximum depth of search. Setting a few ply search still results in near instant response but now Chessnut Evo’s Maia no longer blunders, in fact it plays rather well. It could be argued that it’s no longer following the concept of the Maia project with searching enabled, but I’m happy for the option to be added and have suggested to Chessnut that we have a greater range of ply settings for experimentation.

I decided to use the MCGP2024’s classic book for this match. I’m not sure what book was used for the ELO rating on the Wiki ELO page? I also opted for the Active time control on the MCGP2024. The Maia bot on the Evo does not use the clock for its thinking and just searches to the specified depth, thus for most moves the response is almost immediate or at most just a few seconds. I chose the classic book because with this the ChessGenius machines seems to perform better against other machines. However, here we are dealing with a neuronal engine so it may be a little different especially as Maia plays a little more human-like. The Maia does not have a book opening option, it just makes its moves from the data it was trained with, so in essence, knows the book openings from this and now aided by some searching to avoid pitfalls. I’ve mentioned this in a previous post but the Maia bot will always play 1. P-K4 when playing white, which is disappointing if we want more variety. That being said, it’s doing this because of the 12 million human games it was trained with, so not that unusual.

After that preamble onto the games:

Game 1/10



Although ending in a draw, Maia played very well in this game holding a better position for most of it. For example in the following position black had just moved 16…Qa5:

Black’s Queen has been pushed back to the side and Maia now moved 17.b4! Enclosing Black’s Queen. The position is very good for White with the Bishop on b2 commanding a good diagonal especially after the move d5 that comes later. The Maia bot (Lc0 engine with Maia weights file) is playing very well with good piece positioning and attacking style. However, despite having a Queen and 3 pawns against Rook and 5 pawns in the endgame, it only managed 3-fold repetition. Maybe with more depth of search it would have found a way to win?

Ray