Cool HOWTO on creating an AI to play 2048

Originally published at:

1 Like

2048 is kind of garbage:

But why is Threes better? It’s better for us, for our goals. 2048 is a broken game. Something we noticed about this kind of system early on (that you’ll see hidden in the emails below). We wanted players to be able to play Threes over many months, if not years. We both beat 2048 on our first tries. We’d wager most people that have been able to score a 768 or even a 384 in Threes would be able to do the same using the fabled “corner strategy”. You probably could too! Just try tapping “up” then “right” in alternating order until you can’t move. Then press left. You may not get to a 2048, but you might just see your highest score ever.

1 Like

Threes is great, once you silence it and turn off animations. Cutesy noises get annoying real fast.

I prefer 2048 though. For me as a super-casual player, it’s more fun. It’s also open source and that’s a pretty big deal. I actually haven’t looked at Threes in a long time, so it’s possible it’s open source now too.

In what sense is this artificial intelligence?

1 Like

Just using his title. From the YouTube comments:

A great, truly “”“AI”"" based solution to this problem would be to implement a searching algorithm but when the algorithm reaches its depth, a CNN evaluates the board state for not which move is best, but how preferable the board state is. Then, apply minimax and alpha-beta pruning to minimize possible error and “win adversarially” against the rng.

1 Like



This topic was automatically closed after 5 days. New replies are no longer allowed.