There is another question that asks "Has anyone ever written a chess program that does have insights of its own" that question is about a chess only game AI. I am asking having to learn the rules of the game, pieces, board size, etc. The program should be able to learn any board game, checkers, go, etc. Computers programmed by humans are better at chess than humans. Any computer programmed by itself (via some self learning algorithm, where even the rules have to be learned, i.e. number of pieces, board dimensions, piece move rules, game objective, etc) playing any good at all? Let's ignore that the self learning algorithm would have been devised by a human...

If you know of any such generic game AI engines, please post references to them, at least their name. Thanks!

  • 1
    The name of the research field is "General Game Playing". Jun 1, 2015 at 18:52

3 Answers 3



This is the blog of a guy who tries to do what you describe: Create an artificial intelligence that learns to play chess purely by trial and error. I don't think he ever got very far. He also believes that the old testament contains coded information on how to create an AI …

It is surely possible to create an AI that learns to play chess from scratch, but I doubt it would ever be any good (Though this General Game Playing stuff doesn't learn the rules, the deepmind company atari AI does.). The reason is that chess is a tactical game and calculating tactics requires extremely optimised code. That's nothing you just stumble into after playing a few million games.

  • The example paper only illustrate a framework. Unfortunately, it fails to address how exactly the learning can be done. We still don't know the inputs and outputs, there is no workflows or protocol.
    – SmallChess
    Jun 1, 2015 at 14:13
  • 1
    The link looks like it talks about how to simulate a brain for playing chess, rather than learning a new game of chess at run time.
    – SmallChess
    Jun 1, 2015 at 14:20
  • Well, you have to read a bit more, there is no representation of the board programmed into this brain, the input is purely what the 3x3 "eyes" can see and it is supposed to learn from "changes" in his view. The first paper is just an example, there is a lot of General Game Playing stuff out there. And the Deepmind paper should have a lot of technical details. Jun 1, 2015 at 15:02
  • Unfortunately, the atari AI doesn't learn the rules, it just learns a set of actions that follow them. While it could successfully play many games, it would probably not be very good at chess. Jun 2, 2015 at 16:03
  • Isn't that ultimately the same? But I agree that the set of actions that follow the rules are vastly more complicated in chess than in any atari game. Jun 3, 2015 at 7:54

Zillions of Games is a General Game Playing system that plays chess. It was perhaps designed with chess specifically in mind, but in order for the program to play a game, you must first load a zdf file which specifies how the pieces move and what the goal is. It can also play checkers, go, shogi, etc.

The engine Nebiyu running in the GUI Winboard Alien Edition can play checkers, chess, go, reversi, Amazons and a few other games. On Android, there is a program called XO Demo which plays many board games. On iOS, there is an app called Social Games.

The Association for Advancement of Artificial Intelligence holds a General Game Playing competition every year at its conference. The programs are designed to play almost any game, and are not told in advance which games will be played. Each of the programs could play chess if someone wrote up a rule set. I don't think the GGP competitions have ever included chess as one of the games however.

In the context of chess engines, the word "learning" refers to the engine adapting its opening book (book leaning) or evaluation (automated tuning) based on its results. Many engines learn in this sense, eg. Crafty, Deep Thought, Exchess, Gaviota, Junior, MChess, Neurochess, Romichess and Tao.


The 2016 conferece paper "DeepChess: End-to-End Deep Neural Network for Automatic Learning in Chess" by David, Netanyahu and Wolf reports such a learning method.

Quote from the abstract of the paper (emphasis is mine):

We present an end-to-end learning method for chess, relying on deep neural networks. Without any a priori knowledge, in particular without any knowledge regarding the rules of chess, a deep neural network is trained using a combination of unsupervised pretraining and supervised training. The unsupervised training extracts high level features from a given position, and the supervised training learns to compare two chess positions and select the more favorable one. The training relies entirely on datasets of several million chess games, and no further domain specific knowledge is incorporated.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.