When outcome_type = “binary”, reinforcement learning often slows down and does not complete.

[MatsuuraKentaro]

Though iteration numbers that begin to slow down are probabilistic, it occurs after some progress in learning as follows:

This may be an RLlib issue. The current solution is to (1) use checkpoints saved in the middle of training, (2) use the resume_learning function, or (3) try several runs. We would like to wait for RLlib update.

[MatsuuraKentaro]

This occurs on Windows 11 native, but not on Ubuntu 22.04 (WSL).