AlphaZero vs Stockfish Chess Match: Game 5 

On the contrary. he's just talking without giving any insight. He could've just talked about puppies for all I care, absolutely worthless commentary. Try to at least teach viewers something, instead of commenting as if we were watching nascars or something totally braindead.

I do not have an elegantometer with me but I dare to say a computer is not more elegant than it is shy.

Um, okay... Like Kasparov. Feel better?

It could go in all the right ways, but it could go in all the wrong ways also...

Pedone Rosso Evaluation may not work the same way as Stockfish. Look at the concepts of the decisions it makes. Az seems to rely more on static positions that rely on minimal tactical moves. This is very much like Go, which it learned first. Material advantage is much less important than piece position related to the board if the calculation is strong enough. Neural networks approximate human associations, so brute force algorithm thinking is not relevant to az. Humans don't waste time calculating each potential line like Stockfish. It's a waste of time and energy and it isn't thinking, which is shown by the end of this game, Stockfish doesn't know anything about chess just about calculation that has nothing to do with chess.

according to the paper released by DeepMind, the initial inputs were only the rules of the game, and that checkmate is = 1, draw = 0 and loss = -1. There is no position evaluation algorithm, it plays to maximise the possibility of scoring 1. So since Alphazero plays what it "expects" would lead to scoring 1, it seems to come up with lines that sacrifices material in a way no engine would ever do, because they all have evaluation algorithms that heavily depends on material count.

Pedone Rosso the go version evaluates win% for each position read so it will be the same since the same algorithm can learn chess, go and shogi just by changing the rule inputs before training begins.

The problem is that with neural networks, the "evaluation function" is nothing a human can make sense of. The position is fed into it as series of numbers, and the network adds up these numbers in various ways making intermediate sums clipped by sigma function, then again adds up these intermediate results in various ways multiple times over until it presents a number or a bunch of numbers corresponding to its estimate of either the position or the next preferred move. The various weights used to sum up the input numbers is the matter that gets evolved in the learning process and for a human, it's a complete untanglible mess. I heard that AlphaGo Zero has about two million such weights. Not sure how many were there for the chess version but it may be in the same ballpark. Check this video (and the following ones) for a primer on neural networks: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-aircAruvnKk.html

I read these answers, and I got a rough idea of how the equivalent of alpha0's search function operates, which makes understanding the meaning o its equivalent of the evaluation function a complex task. Nontheless, if the process allows the program to make a decision based on mathematical operations, there has to be a possible translation. And it's of course something we need to be able to extract from this software, or the whole experiment will simply be irrelevant. Chess is the problem we fed the machine, and the evaluation of current and possible future chess positions in view of a maximization of the winning chances is the given (approximated) answer, or the software wouldn't obtain the shown results vs the Stockfish engine. I love Chess and all, but developing AIs is not something we do in order to play Chess: it's something we test using Chess. But we do it in order to build instruments that are fit to be used to solve much more relevant problems, problems that are too complex and data rich for us to solve with our "natural intelligence" alone. (Also we do it in order to understand how our own mind actually works through the creation of physical models that we can tune and measure, which is also different than winning a chess game: still a matter that requires communication of the results from the model to us.) Thanks for your answers!

Chris Isaacson if only everybody knew, but people are too lazy to actually run it through stockfish, and look at its suggested moves on lichess

Set up SF8 however you want and see how long it takes it to find A0's amazing moves. Go ahead.

Looking up a game with a sub 2200 elo player is totally irrelevant. When analyzing games, you typically only go through IM+ level games because of so much random shit going on when looking at levels below IM

@Peter sturm No I'm not? what is the purpose of your question? I'm 2100 and I would consider myself insane to refer to my own games because most openings are a mess at that level

I agreed, it weird ... maybe something personal ..

i think it'S one of those ''conditioning'' Things. Those People just don't ''know any better'' because the experiences from which they've learned their behaiviors, were such that it lead to them taking that sort of Action. I take solice in the fact that the majority of People here recognize it as, well... as nearing worthlessness to take those Actions. :D thanks for the Analysis, this vid is much appreciated

Eshan Chowdhury The haters escaped from the psychiatric clinic. Most of them have been recaptured, though sadly not all of them.

Don't forget the Sicilian! ;)

Dubious at stronger than the stockfish level... I think I’ll still keep them in my repertoire.

TPU's don't draw a lot of current.

What do you mean "protecting from trolls"?

Wasn't it 2 years ago that news came out that fighter jet dodge fight simulations with AI were already onesided, even with a partly crippled plane?

this comment aged well

Thank you so much!

I did

X2

ok...you still have to explain 80 knps beating 70,000 knps.

You should just play chess

These 10 games gave us an insight into how A0 approaches the game of chess in general(opening/middle game/endgame). I think it is not far fetched to say that we got a glimpse of what a first partial solution to chess would most probably look like.

Nope A0 does not save positions. A0 number crunches each position using the neural net while A0 is searching the tree.