How to train exclusively with BC?

How do I exclusively train an agent with Behavioral Cloning? I don't see any configuration file or explanation of this. I can't really tell if PPO is running in parallel.

As an additional question, why can't I run BC without running the game? If I have a slow game but hours of recorded demonstrations to train on I don't see why the game has to run. Is there a way to disable this?

 

You enable behavior cloning by adding the following section to your Brain config:

Code (csharp):

1.     behavioral_cloning:
2.         demo_path: path to file.demo
3.         strength: 1
4.         steps: 200000

Check out

mlagents-learn --help

. It allows you to set time scale factor (x20 by default) and disable graphics. The answer why is basically because Unity acts as environment simulator and will produce states and rewards for the training agent.

 

Thanks. Let me try to clarify.

1. The documentation does not clearly state whether BC happens before (in the specified steps) or simultaneously with the PPO/SAC trainer. I want to be sure that it is only BC running.

2. If only BC is running, it wouldn't need an environment. Only PPO, SAC, and GAIL need an environment, not BC.

 

But it is not possible that only the BCwould work!
After all, you define only a set of rewards for a set of behaviors, but not the neural network itself in the Demo.
I do not think that the bot can repeat the BC. Only simulate for experience.

 

I found out that this was a feature once:
if trainer_type == "offline_bc":
raise UnityTrainerException(
"The offline_bc trainer has been removed. To train with demonstrations, "
"please use a PPO or SAC trainer with the GAIL Reward Signal and/or the "
"Behavioral Cloning feature enabled."
)
Sadly, it has been removed.