This has to do with this thread: https://lichess.org/forum/lichess-feedback/your-opponent-left-the-game-you-can-claim-victory-in-x-seconds?
So, my idea is that Lichess could make a supervised machine learning model that uses gradient boosting to train an ensemble of small regression/decision trees to estimate the chance of a rage quit or genuine connection loss based on factors such as the evaluation of the chess position during and slightly before the disconnection (i.e. if the position is highly unfavorable for the disconnecting player), network stability, recent move quality, and the player's history of disconnecting based on these factors.
The decision of whether to auto-forfeit the game or wait it out should occur after 30-or-so seconds as a "safe zone" just in case it is a genuine disconnection regardles to wether it is more likely to be a rage quit or not.
One user pointed out that if the countdown reaches 89 seconds (past 30 seconds of the 120-second countdown), then the connected player will know that they are not winning since, if based on eval alone, the ragequitter is most likely losing, thus ending the countdown at 90 seconds. Another user pointed out (with respect to the previous user) that this essentially would be forced cheating, since the eval is practically presented to the connected player. This would predominate in situations like complex endgames, where it is unclear to the player whether or not they are winning or not, while to the engine, it is crystal clear. A way to counter this issue is to train the model on even more factors such as player rating and puzzle performance (among other factors) to determine if it is likely that they comprehend the eval.
My whole message is too long, so I will send the remaining part of it as a response to this post.
#AI idea for Lichess
hot pollen
The "Chess Insights" can also be included for further accuracy.
The model should be trained on games with... well, disconnections. And the time controls that predominate the training data should be blitz and rapid, since ragequitting is more likely to happen in these time controls, while bullet is obviously discarded, and classical can be used subsidiarily.
Training the model shouldn't be expensive, since the platform (Lichess) already produces the data for the model in huge amounts. GBDTs are extremely efficient for tabular data sets (like those mentioned in the first paragraph), quicker and easier to train than artificial neural networks, and are highly accurate. The most expensive part of this project would be hiring a Machine Learning Engineer, a Backend Engineer, and possibly a Data Engineer.
But... of course, there is a high probability that Lichess doesn't want to do all this (which is reasonable). So a simple but boring alternative is to simply use a Reactive Machine AI for this (-_-).
Now, I am not an expert, so my thoughts might get scattered. So, I would like if some knowledgable people or devs check on my idea and refine, as I think it has potential. What do you guys think?