Alpha Zero vs Stockfish

[OngBonga]

Why does a computer evaluate the same position differently?

It's actually quite a profound question when you think about it. A basic understanding of what happened is required... AZ's rook was under attack from SF's knight, so AZ plays Rf6. SF moves the knight to attack the rook on f6, so AZ moves the rook back to f7. SF in turn moves his knight back so the rook is once again under attack, repeating the position from before. Now AZ plays Re7.

So why didn't it play Re7 before?

Best I can assume is quite scary... AZ is testing SF to evaluate its skill level. It's actively probing for weaknesses in SF's methods.

Noone told it to do that, it taught itself to do that simply by being programmed to want to win at chess.

[Savy]

Best I can assume is quite scary... AZ is testing SF to evaluate its skill level. It's actively probing for weaknesses in SF's methods.

It isn't that at all it's because it's learning as it plays. It played a move that it evaluated as best then after having gained more information when the same position happened again it thought another move was better.

[OngBonga]

It isn't that at all it's because it's learning as it plays. It played a move that it evaluated as best then after having gained more information when the same position happened again it thought another move was better.

But the information it gained was a single move ahead, it can anticipate this move and already know it's not winning.

[Poopadoop]

It might have expected AZ to do something else the first time, or it may have just realised its first response was not optimal. Don't think it was trying to probe AZ for weaknesses. That's theory of mind shit and not something computers are capable of afaik.

[Savy]

But the information it gained was a single move ahead, it can anticipate this move and already know it's not winning.

The reason it's different is because it's constantly learning, it evaluated a move then it happened and could see further so when the move was repeated it then saw a better option. There is also the fact that it might realise that if it plays move A and the other player doesn't play the exact perfect move it's more winning and if it does it's repeated therefore playing move A first before trying the next best option of B is a better play.

There are loads of reasons and literally none of it is trying to exploit how the player plays.

Stockfish wasn't allowed to use opening books, which is an important part of the story you're missing in your OP.

How much difference do you think this would make? I've only really seen that a lot of people in the chess community are pretty blown away by it.

It'd be interesting to see how quickly AZ caps out. You'd think the next step would be beating something like stockfish with an opening book.

[spoonitnow]

How much difference do you think this would make? I've only really seen that a lot of people in the chess community are pretty blown away by it.

No clue. It's just an important piece of the story that should be mentioned because traditional engines don't play the opening relatively well without them. The results are still pretty interesting and will definitely change the game for engine competition.