Stockfish Deep Game Analysis Failure

In theory, newer versions of Stockfish play better than older versions, but the changes from one version to the next have been incremental. I still have a program on my phone that uses Stockfish 6, but for most positions, it gives the same analysis as newer versions.

Stockfish 12 started using AI for evaluation, but this was improved with Stockfish 16, when the older evaluation methods were completely abandoned. However, for most positions, it still chooses the same moves.

For game analysis, one doesn't have to look very deep with Stockfish to find the correct move, especially in the middlegame. Chess.com used an 18-ply search, which could analyze a whole game in seconds, but now they don't say how deep their analysis function looks. I suspect that it varies depending on the demand on their servers because occasionally Chess.com won't analyze a game, stating that its servers are busy. Hypothetically, there wouldn't be much difference with a 16-ply or a 17-ply search.

I notice very little difference between 20-, 30-, and 40-ply Stockfish analyses, except in the opening. The deeper analysis often prefers different opening moves. For this reason, I would analyze some of my games at 40 ply to see which opening moves it prefers. Rarely, I analyzed even deeper just to see if the opening choice changed.

However, I noticed a problem starting with Stockfish 17.1. According to Google AI, the difference between versions 17 and 17.1 is trivial. However, I noticed that during deep searches, version 17.1 would run into bottlenecks in some positions and get stuck, sometimes taking days to complete the analysis, if it could finish at all. This was especially true in endgames, but it could also happen in the middlegame.

Stockfish is optimized for normal time controls, but deeper searches are where I have been having this problem.

I tested Stockfish 19 to see if it could analyze a whole game at 40 ply. It looked like it might take days, so I switched to Stockfish 17, which analyzed the same game at 40 ply in roughly an hour. I then tested version 17.1 against version 17 with a couple of games, and version 17.1 took almost twice as long.

It seems almost impossible to analyze a whole game at 40 ply with newer versions of Stockfish, so I was using 30 ply instead, although that sometimes failed as well. For one game, I had to drop the depth to 20 ply.

There doesn't seem to be much difference between Stockfish versions 17 and 19 in terms of analysis. Version 19 is reportedly a little stronger. According to ChatGPT, Stockfish 19 is only about 8–15 Elo stronger than Stockfish 17, based on a CCRL rating list at a fast time control. According to the people who make Stockfish, version 19 is 90 points stronger.

For the deep analysis I like to do, I am stuck with version 17.

I was thinking that maybe I am the only person who likes to do deeper analysis, so this has never been an issue. However, I had ChatGPT research this, and it found reports of other people having the same problem.

--
Best wishes,

John Coffey


P.S. Although I use an old version of Fritz to analyze with the Stockfish engine, I sometimes also use this free program: https://www.triplehappy.com/