I wonder what would be the impact of using Counterfactual Regret Minimization instead of training a neural network based on hands played by real players?
Whys is using CFR better than training based on real data?
It's not necesserily better but with CFR you can learn beyond what humans have learned, but on the other hand you dont learn their usual mistakes to more easily exploit them.
Also in this approach you need CRM since at every point you are checking what would've happened if you picked something else, which is just impossible with a fixed dataset.
Whys is using CFR better than training based on real data?