ICM-APT can't reproduce the performance of the described experiments.

[seolhokim]

Hi, Thanks for good quality implementation. I did pretrain.py -> finetune.py on walker_run environment.
It showed meaningful behaviors in pretraining session. but in my case It seems not different between pretrained agent and scratch agent. Actually, It doesn't work without any code modification(for example : model load directory problem).

If you want something to me for reproduce, then I can give you results and modifications.

[seolhokim]

It's worse than a scratch on a walker stand.