Fluther: How do chess playing computers evaluate a position?

How do chess playing computers evaluate a position?

Asked by LostInParadise (31904

) March 20th, 2019
15 responses
“Great Question” (2

)

What I have read is that the latest generation of AI programs learn how to win at chess, or any other game, simply by knowing the rules and playing millions of games against itself. No general strategies are given to the computer. There has to be more to it.

I have done Web searches on this and have not found an answer. All I found were ways of minimizing the amount of look-ahead, assuming some way of evaluating each position. But how are positions evaluated?

Follow this question

Send to a friend!

Topics: computers, chess, AI

Observing members: 0

Composing members: 0

Answers

During training, they build their own database of what move resulted in a loss/win in each given position.
Essentially, they have experience.

ragingloli (51957

)“Great Answer” (4

)

As I understand it, it a bit more than what you wrote. That’s part of it, but not all.

The computer does play games with itself, but it also looks at historical (actual) games between chess masters. It looks at placement of all pieces on the board (what is 64 factorial? – that’s how many possibilities there are), and then calculates, based on board position, the most likely outcomes of any given move.

And it does this blazingly fast.

Essentially, it looks at not only its own history of self-=played games, but also previously played historical games, and refers to those for calculating probabilities.

elbanditoroso (33145

)“Great Answer” (2

)

@elbanditoroso
If you are talking about AlphaZero and LeelaZero, then no, they have no human games in their dataset at all.
There is a reason for the “Zero” in their names.

ragingloli (51957

)“Great Answer” (4

)

@ragingloli , There must be some generalized evaluation function. Suppose, for example, each piece is initially given equal value. Then maybe that value can be updated by looking at what is around in a winning game.

LostInParadise (31904

)“Great Answer” (1

)

I think the Zeros’ goal is always a checkmate. Any value given to pieces are derived from their usefulness in achieving that goal (the Zeros are only given the rules, that means no predetermined piece values). So I would imagine the AI evaluates the probability of a checkmate.

ragingloli (51957

)“Great Answer” (0

)

How does an artificial intelligence recognise a dog? The systems that do this most effectively aren’t told anything that you might expect eg look for four legs and a tail. They are just fed billions of images and told whether their guesses are good or not.

The raw data is just pixels, the output is simply dog or not dog the secret is in the neural network that processes the data and modifies itself.

Alpha zero doesn’t assess game positions like we do and exactly why neural networks are generally so effective doesn’t seem fully understood.

flutherother (34518

)“Great Answer” (1

)

Iteratively. Pawn to king 4.

Pinguidchance (5040

)“Great Answer” (0

)

They use reinforcement learning algorithms.

ARE_you_kidding_me (20021

)“Great Answer” (1

)

Reinforcement of what? The chances of the same position appearing in another game is infinitesimal.

LostInParadise (31904

)“Great Answer” (0

)

Response moderated (Spam)

Response moderated

They are playing millions upon millions of games.
If you discard all the illegal moves, and moves that lead almost immediately to bad results, I think it becomes manageable.

ragingloli (51957

)“Great Answer” (2

)

Here is a presentation by one of the creators of AlphaZero.
https://www.youtube.com/watch?v=3N9phq_yZP0

ragingloli (51957

)“Great Answer” (2

)

Response moderated (Spam)