In my projects, the "world" size can change (unlike Go, Chess where the board size is fixed).
Is the DoTA board size fixed?
I guess the LTSM encodes the board history as seen by the agent. But this probably slows the learning.
Some people suggested auto-encoder to compress the world, and then feed it to a regular CNN.
Any comments would be welcome.
In my projects, the "world" size can change (unlike Go, Chess where the board size is fixed).
Is the DoTA board size fixed?
I guess the LTSM encodes the board history as seen by the agent. But this probably slows the learning.
Some people suggested auto-encoder to compress the world, and then feed it to a regular CNN.
Any comments would be welcome.