Hi Hans Jürgen, thanks for asking, I am more than happy to explain.

I am aware of this (recent) compile offered by JA. It is indeed rather "normal" as Jim gives the instruction to set the level for Maia to 1 node, thru the GUI setting. I guess the added value lies in the ready-to-use package of all the required components.

Time ago I experimented this 1 node setting and found out Maia was far weaker than expected with regards to the promise made by the net names. Well, there is some misunderstanding in such an expectation: a Maia net mimics probable moves from human players ranked accordingly, with some accuracy and success - but there is no promise to play a whole game at this level. And indeed, some moves played can fit the level as far as positional patterns are concerned, but are tactical blunders that ruin the game.

I don't own a Chessnut EVO, yet I read in the forum the device offers Maia with depth search levels added - a solution to the above concern. It lets the LC0 engine use its Monte-Carlo Tree Search while still using the Maia net for evaluation. This means less tactical blindness whilst still aiming at selecting moves that would fit the target level, from a positional standpoint.

My implementation is slightly different, though targeting the same: instead of fixed depth levels, I use NPS throttling. Depth reached is therefore "free", the limit is the allowed thinking time. This way another concern is solved: the instant play you get using the 1-node setting (which induces a human player to think less and move fast). NPS limit is a standard LC0 UCI setting.

Of course the NPS limit needs to be drastic to avoid reaching too high a depth with regards to the expected level of play, turning Maia into a tactical monster is not suitable. Therefore, there is some tuning effort required to set the NPS limit value for each target level of play. Maybe the outcome of this tournament will drive some changes in the settings I use!

mit freundlichen Grußen,
Eric