Suggestion: Training outputs

[Torrenal]

I'm assuming that when training the AI, it had only 4 outputs (north,south,east,west). Apologies if that's a faulty assumption.

For training the network, you might consider training it for a few additional outputs - not because you need them to play the game, but because by needing to provide them, the network will need an additional awareness of the game mechanics.
4 outputs (N/S/E/W) for when a move in that direction is possible, 0 if not possible.
4 outputs (N/S/E/W) for when a move in that direction will merge tiles.
1 output for a complexity score of the remaining tiles, after whatever mergers happen for the requested move.

The 'human' analog of training for these outputs would be 'learning the rules of the game' - It'll know when tiles merge and it'll know when moves are invalid, and it may internalize some of the logic for that in its decision process for moves.

[tjwei]

The assumption is correct, it has four actions. Thanks for the suggestions. Torrenal <[email protected]> 於 2019年5月28日 週二 上午11:49寫道：

…

I'm assuming that when training the AI, it had only 4 outputs (north,south,east,west). Apologies if that's a faulty assumption. For training the network, you might consider training it for a few additional outputs - not because you need them to play the game, but because by needing to provide them, the network will need an additional awareness of the game mechanics. 4 outputs (N/S/E/W) for when a move in that direction is possible, 0 if not possible. 4 outputs (N/S/E/W) for when a move in that direction will merge tiles. 1 output for a complexity score of the remaining tiles, after whatever mergers happen for the requested move. The 'human' analog of training for these outputs would be 'learning the rules of the game' - It'll know when tiles merge and it'll know when moves are invalid, and it may internalize some of the logic for that in its decision process for moves. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#8?email_source=notifications&email_token=AAH4CSV5W7G5J2JVCT5WPR3PXSTULA5CNFSM4HP7QS52YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4GWD7XFQ>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAH4CSVAQXZBUN5YM577LVLPXSTULANCNFSM4HP7QS5Q> .