i am trying to implement a transformer based agent for a tile placing board game, and i was thinking about this kind of architecture:
- Src input is embedded board, with an empty block (for example 10x10 with a 3x3 block missing, so 3*3=9 tiles to be placed)
- Target tokens are the tiles to be placed.
- The transformers somehow outputs something that can be interpreted as a permutation of these tokens to fill the empty block.
Is there an architecture that is made / proefficient in making those kinds of predictions (not just tokens but a permutation of the target tokens)? Or do you see a better way to approach this type of situation ? (i am pretty new to this)
