Recommendations for Crazy Eights

[umtati]

Hi,
if anyone has trained on Crazy Eights and got some recommendations for algorithms and hyperparameters I would be very thankful. I already know DQN converges fast but I don't know how good it scales compared to others. DeepCfr seems inefficient but might be better with enough optimization. IS-MCTS on high rollouts is the strongest but search is not efficient for simulating many games.

[amyssnippet]

I will recommend trying NFSP (Neural Fictitious Self-Play). It is designed specifically to solve the convergence issues DQN faces in imperfect information games without the extreme cost of tabular CFR. It's the standard deep-learning baseline in OpenSpiel. Check examples/nfsp_example.py for a starting config, but you will likely need to double the hidden layer sizes for Crazy Eights.