Last week I finished training an agent which can play a simplified version of 2048. This week came time to train an agent for the original version. I tried both fine-tuning the previously trained agent and training an agent from scratch. While the pre-trained agent performed better initially, they both seemed to converge to about the same strategy. Therefore, I decided to continue with the pre-trained agent.
The original version of 2048 is much harder than the simplified version. The performance increases much slower. I'll keep training until the performance appears to stabilize. Since there is not way to play 2048 perfectly (see articles below), I'll see how may games out of 100 the agent can win at the end.
Is the game 2048 always solveable?
I did see the agent win a few games so far during training. It is following a strategy similar to the one described here:
2048 Game Strategy - How to Always Win at 2048