Sweet Valley — Golden Gauntlet
that bot interface I made for 2048 isn’t nearly as good as how ov3y refactored the game. Maybe I’ll just gut his code out of that, write some documentation, and push that.
Although probably nobody will use it. I might also just delete my repo and assume that if someone is going to make an AI they’ll roll their own stuff.
My 2048 AI thus far. I usually see it get to at least 1024, and it’s won about 30% of games that I’ve watched.
Basically I’m doing minimax with alpha/beta pruning, but my scoring heuristic is just to run the board through an artificial neural network. (I used github user ov3y’s minimax code to save some time. Thanks!)
The network was taught via a genetic algorithm. I randomly generated 100 neural networks and each played a single game using the minimax approach. The bottom 75% of networks were purged, and the top 25% filled back in the population using genetic operations on the surviving networks (crossover, mutation). This was repeated on my crappy ARM samsung chromebook for about 6 hours while I was sleeping, so it was probably only about 40 generations of networks.
What you see here is the best network that was found in that timespan. I’m gonna rewrite this to use node or something and run it on a VPS with better specs so the learning doesn’t put me out of a computer for 8 hours :P.
It could use a little more work, like a better way to explore trees of possible moves.
Pull requests welcome.
I’m idling in #tumblrcode on irc.freenode.net and if you like programming, you should too.
Unless there is another preferable server.