A Python AI for two-player board games

zulko.github.io

57 points·rubinelli··7 comments

Oh nice. Reminds me of a hackathon I did a while back where I tried to implement a connect 4 AI bot using a genetic programming approach.

Here's the code: https://github.com/dangoldin/connect4bot

And here's a writeup: http://dangoldin.com/2013/11/30/genetic-programming-connect-...

This looks like a really tidy framework for this class of general games (http://en.wikipedia.org/wiki/General_game_playing). There's actually an annual competition for this type of thing.

In this competition game rules are described using a language called GDL (http://en.wikipedia.org/wiki/Game_Description_Language). Broadly there are two approaches. Super fast Monte-Carlo tree searches and algos which, from rules, infer game-state heuristics which are used in more traditional depth-limited search.

there was a Coursera course on GGP this fall:

General Game Playing: https://www.coursera.org/course/ggp

they were using java/javascript for their players. And there were other issues making their framework a little bulky

Very nice, the examples are very clear and terse.

I started on a similar project in C which is unique in that the game logic is written more-or-less procedurally and the library takes care of backtracking the game state to iterate through the tree nodes:

https://github.com/sehugg/Starthinker

Why didn't you go for monte carlo tree search? Wouldn't that have been more robust to handle different games?

Negamax is very good for a wide variety of relatively simple games. It will utterly fail for Go, but I've written negamax AIs that dominate any casual player at chess, connect 4, reversi, gomoku, etc. They become yet stronger if you add a quiescence search.

Finally, we can play WarGames in Python http://zulko.github.io/easyAI/examples/games.html

Reply on news.ycombinator.com