Change state and reward design

Edouard360 / Halite-Python-RL

A fork of the Halite Starting Kit, aimed at providing an interface and debugging tools and for RL strategies (reinforcement learning).

6 stars 2 forks source link

Open Edouard360 opened 7 years ago

Edouard360 commented 7 years ago

Different shapes for looking around the square might be possible.

Maybe use a mask ? In tensorflow or in python
For now the game state is 27 number, and it is hard-coded, which is bad for modularity... We might have states of various sizes

leonardhussenot commented 7 years ago

We might want to implement a reward depending on production/(distance to border). Typically, if we are far from the border, loosing strength is not important (as we want it to go to the borders) and we shouldn't penalize square giving their strength. Same thing if the square is not really productive, it is not really important.
On the other side, we should reward a productive square for being productive.

Edouard360 commented 7 years ago

Solved the hard coding of the state size, creating a state object, and a strategy.json file