When I start the mlagents-learn command then the training process always exits after 500 000 steps. I tried to load a config file which contains bigger max_steps value but it ignored the config file. If I set a config file in argument then after the training process this config file will be copied to the results (yes the copied version also has bigger max_steps value). After starting the mlagents-learn it seems a default configuration is always written out to the console which has only 500 000 steps. I don’t know where does it come from but I cannot overwrite this value.
Okay, I’ve figured out. Those settings are default and I had to use absolute path for the config file in the argument. I also changed the behavior name in the config to the one defined in Unity editor.