I’m new to RL and I’d been trying to solve the 2048 game( check it on Google ) using A3C for a while, but I couldn’t make it so far.

I tried some ideas as rewarding each step with the score points gained, changing the representation of the board. I’m currently testing a reward function that is zero in all cases except when the game is won, and -1 if the game is lose. I have the feel that this reward is uninformative, but other rewards tend to make the bot stuck.

Any ideas in how to make it work?

Thanks!