Programming a computer for playing chess (1950) [pdf] | Hacker News Reader

Programming a computer for playing chess (1950) [pdf]

vision.unipv.it

117 pointsteleforce31 comments

Programming a chess program was an important challenge for the "first generation of hackers" working on the mainframe machines like TX-0 and PDP-1 in '50 and '60, as described in "Hackers: Heroes of the Computer Revolution" by Steven Levy. I highly recommend this book, I think a lot of people here can recognize their own passion and interests in the stories described there.

Also, there's interesting implementation for the ZX-81, created over 20 years later and fitting in just 1024 bytes: https://users.ox.ac.uk/~uzdm0006/scans/1kchess/

Not even 1024 bytes. There were 1024 bytes available to the computer, but some of that was required to store what was displayed on the screen. Exactly how much depended on how much of the screen was used(!).

Different times...

It's amazing how Shannon contributed to almost everything important to understanding computers at all levels. He contributed to information theory (thus communication, error correction and compression), digital circuits (Master's thesis), Artificial Intelligence (not exhaustive: this paper, Theseus), cryptography and even circuit complexity!

As well known as he is for information, sometimes I think the breadth of his contributions to the foundations of computing are not fully appreciated.

You say AI, but even more precisely: I believe that his ideas form the core basis of modern LLMs actually (their probabilistic word sequence modeling underpinning), which is indeed extremely impressive and deep.

Here's a PDF with the original publication: https://ia601600.us.archive.org/32/items/philosophical-magaz...

For more on how modern engines work (although the core idea of alpha-beta pruning [1] is still used) you can check out the chess programming wiki [2].

1: https://www.chessprogramming.org/Alpha-Beta 2: https://www.chessprogramming.org/Main_Page

There is a fantastic video about alpha-beta pruning by Sebastian Lague: https://www.youtube.com/watch?v=l-hh51ncgDI

I read "Chess Skill in Man and Machine" when I was in college.

https://www.amazon.com/Chess-machine-monographs-computer-sci...

It wasn't so much that I was interested in chess, but the book really opened my eyes to how to write computer programs. For example, a 64 bit vector was used for the board, and for each piece type for each location another 64 bit vector was used to specify the legal moves.

I know all this sounds routine today, but I had only the most rudimentary programming skills at the time, and this stuff was fascinating.

I finally gave up on playing chess because instead of thinking about my next move, I'd think about writing a program to calculate the next move. Hence I played pretty badly.

a 64-bit vector can also be used to cheat at poker

So far as I can tell, the A+B strategy described therein is exactly the principle that Stockfish -- and more or less every other chess engine -- was built on prior to AlphaZero, after which the top engines moved to NN derived evaluation criteria for what is still a pruning tree search.

AlphaZero/LC0 uses a very different kind of search, enough, which is not based on pruning to the same extent. (I guess you must count LC0 as a top engine?)

You might be thinking about how the evaluation is trained (MCTS) rather than how it’s applied in a chess game?

from https://lczero.org/dev/wiki/technical-explanation-of-leela-c...:

    Leela uses PUCT (Predictor + Upper Confidence Bound tree search). We evaluate new nodes by doing a playout: start from the root node (the current position), pick a move to explore, and repeat down the tree until we reach a game position that has not been examined yet (or a position that ends the game, called a terminal node). We expand the tree with that new position (assuming non-terminal node) and use the neural network to create a first estimate of the value for the position as well as the policy for continuing moves. In Leela, a policy for a node is a list of moves and a probability for each move. The probability specifies the odds that an automatic player that executes the policy will make that move. After this node is added to the tree, backup that new value to all nodes visited during this playout. This slowly improves the value estimation of different paths through the game tree.

    When a move is actually played on the board, the chosen move is made the new root of the tree. The old root and the other children of that root node are erased.

    This is the same search specified by the AGZ paper, PUCT (Predictor + Upper Confidence Bound tree search). Many people call this MCTS (Monte-Carlo Tree Search), because it is very similar to the search algorithm the Go programs started using in 2006. But the PUCT used in AGZ and Lc0 replaces rollouts (sampling playouts to a terminal game state) with a neural network that estimates what a rollout would do.

It is not what Stockfish does though. From the Wiki page:

Stockfish implements an advanced alpha–beta search and uses bitboards. Compared to other engines, it is characterized by its great search depth, due in part to more aggressive pruning and late move reductions.[13] As of September 2024, Stockfish 17 (4-threaded) achieved an Elo rating of 3642 +16 −16 on the CCRL 40/15 benchmark.[14]