The readme is well-done and makes it easy to follow the code in its different sections. I really liked the idea of unbalancing the reward to address the first player advantage. I noticed the same problem in my simulations but wasn't able to figure out a solution, this could be a smart one. Definitively worth a try!
One small suggestion, you could take advantage of the highly symmetric structure of the game to aggregate the learned features on "different" boards that are actually the same
The readme is well-done and makes it easy to follow the code in its different sections. I really liked the idea of unbalancing the reward to address the first player advantage. I noticed the same problem in my simulations but wasn't able to figure out a solution, this could be a smart one. Definitively worth a try! One small suggestion, you could take advantage of the highly symmetric structure of the game to aggregate the learned features on "different" boards that are actually the same