#stockfish

New ship - Stockfish

I just watched this video and it got me thinking on chess AI.

For those unfamiliar, chess AI will evaluate possible future positions of the current board, then evaluate which of them are most desirable, and then it’ll play the moves to get it there.

If a future branch for example has a lot of good but also a couple really bad boards, it’ll go with a safer branch instead. It always plays the move that maximizes it’s future state. Once it finds a forced checkmate sequence, of course it will follow that branch until it’s opponent is defeated.

~~~

However, back when chess AI was in it’s infancy and just did this, humans could still beat it for a while. The positions in the future were evaluated on if there was a forced checkmate sequence as well as material points. Different chess pieces are worth different so called material points, and by evaluating future boards on material point advantage, the AI could take all it’s opponents options and eventually win that way. But humans still sometimes won. Why?

Well, they understood something that the AI didn’t, but they couldn’t formulate what exactly that was.

So the solution: Make an AI that emulates humans.

~~~

The second generation of chess AI then proceeded to use a combination of the brute force “look for a forced checkmate sequence and material advantage“ approach, but mixed it with emulating actual human played games too, emulating the winning side, of course.

This made the AI much better at openings and endgames. And the mid-game was already it’s strong suit, no human is better at tactics than an AI that can calculate into the future that far.

After that advancement, AI completely outpaced humans. No human alive can beat such an AI anymore, and those are used to evaluate positions of human games. The most prevalent and powerful of that type of AI is called StockFish, and it has gone through countless iterations, each one improving further and further away from the human skill-cap.

But a problem remains: The developers have no real idea what it is we humans understand better, that enabled us to beat the first generation of chess AI.

They knew it was some “positional advantage“ but had no way of quantifying that, no way of making the AI understand that. Except by letting it emulate humans, who do understand and utilize that concept.

~~~

And then this AI from the video i linked came on to the scene, AlphaZero is it’s name. And from how it plays, i think it’s creators have finally found what humans did differently that allowed them to best 1st generation AI. I believe it simply changed the way it evaluates/rates how good future boards are!

And the measure it rates them by is the amount of possible legal moves the AI will have at that future board vs. the amount of possible moves their opponent will have.

The extreme case is a forced checkmate sequence: The enemy has no moves except those that lead to the king being unable to move, as well as attacked.

A checkmate.

So first of all, if AlphaZero evaluates future positions like that, it doesn’t have to make an exception for forced checkmate sequences. Those just happen to be favored naturally. No if-case needed. And also no more evaluation based on material points.

This enables the AI to apply pressure on it’s opponent, because pressure means less possibilities for them, more for the AI.

Many things GothamChess said in the video point to this being the case:

it likes long bishop diagonals

it likes it’s own king to be mobile

it doesn’t care about sacrificing pawns, and will gladly do so

it closes down positions, but usually on the enemy side

The first two points are simple to understand if it evaluates future boards on how mobile it’s pieces are supposed to be.

The third also makes sense, as pawns are just in the way of it’s own piece moves. A pawn can only move one square, a piece in way more ways. So it’ll gladly sacrifice a pawn just to make it’s own pieces able to move more freely.

The fourth point can be understood as minimizing the enemy movement options.

This way of evaluation allows AlphaZero to incorporate things like trapping pieces and developing moves into it’s natural behavior. Developing moves were previously relegated to emulating human openings, but that was because the material point evaluation didn’t account for them. They don’t have to be handled like that anymore, possible move evaluation naturally leads to the chess AI playing developing moves.