StockFish made a great comeback vs LCO at the End.

[Chessqueen]

But of course we need at least 1000 games to make any sound judgement, since if you beat your closest friend who is close enough in rating in a 100 mere match, he probably will NOT admit of you being stronger unless you beat him in 1000 games, which could be very hard if it was physical fighting to get such a beating. Lets just compare in the Human World Chess Championship GM Caruana can claim that Carlsen is not much better than him since they only played 12 games which came to a draw
https://tcec.chessdom.com/

[Michael Sherwin]

But of course we need at least 1000 games to make any sound judgement, since if you beat your closest friend who is close enough in rating in a 100 mere match, he probably will NOT admit of you being stronger unless you beat him in 1000 games, which could be very hard if it was physical fighting to get such a beating. Lets just compare in the Human World Chess Championship GM Caruana can claim that Carlsen is not much better than him since they only played 12 games which came to a draw
https://tcec.chessdom.com/

Great points but at the same time it misses the most critical points. As programmers we quickly reach our level. Then we get in a groove and spin for decades making slow progress. Then someone or something comes along and leapfrogs us. Most of the time we can understand what is it that was done to leapfrog us and we can incorporate it and keep advancing. However, in the case of AlphaZero and Leela it is a bit hidden what was done. They took an inferior search methodology and added something to make it equal or superior
on vastly more powerful hardware. And now everyone is mesmerized by that.

I have tried to explain this before and was mostly attacked. What allows Leela to play so strong? It is 99% from reinforcement learning. Leela's NN has information from entire games at its disposal. Stockfish only has info from its search. So Leela having w-l-d info from as much as 200 ply or more incorporated into the search has to benefit greatly from that when deciding from otherwise roughly equal moves. Stockfish has no mechanism to "see" any deeper than its search. But, it could have! The Stockfish team could add RL in two ways. After game learning like in my engine RomiChess or real time RL. Realtime RL would require spending a portion of the allotted search time to play say a thousand really fast games or game snippets to gain info very high up in the tree and then using that info to sway the search between otherwise equal moves in the main search. And it will also improve move ordering in the main search so the main search will search deeper on top of the learning. The SFKH's and those in the same group need to pick the needle up and move it to the next song.

[M ANSARI]

Sounds very much like MC ... which is nothing new. For sure SF will have to do something to compete with Lc0, but to be honest I wouldn't be surprised if a new version of SF can gain from some type of NN or MC search and keep its clearly superior part of its play. But then again Lc0 can also use its superior part of its engine (the NN positional understanding) and mate it with some really powerful AB search. I think the next 3 years will be interesting times for chess engines.

[Michael Sherwin]

"Sounds very much like MC"

Not really, imo. MC is random and requires millions of games to generate deep statistics to base decisions on at the root. And still at the end of that process it is a haphazard approach. It is like making a random dartboard with irregular shaped regions with random values for the regions. How are you going to find the average score per dart? You could throw a million darts at it and figure the average. But, that varies with the skill of the dart thrower. Or a hundred random dart throwers could throw 10 darts each and then evaluate the different types of patterns they produce around the bullseye and extrapolate from that the average score of the dartboard per dart. In the case of the dartboard I do not know which is best. The point is though there is more than one approach. A purely random one needing a huge sample size and a more intelligent one requiring a smaller sample size. So let's say that instead of millions of MC games we only can search 1,000 AB games searched to depth 4 to 6. I do not know what the depth would be. That would need experimentation. The point though is that first AB game is worth far more than a multitude of MC games. In RL the moves of the winning side get a bonus and the moves of the losing side get a penalty. But moves of drawn games all get a small penalty. So any result will cause the AB games to vary their play and search the AB tree very intelligently instead of randomly. The stats returned to the main search are therefore in closer harmony to the evaluation function of the main AB search. The stats then would intelligently sway the main searches moves rather than dictate its moves. That is the best I can do to explain the benefit of RL to an AB engine like SF.