KeyError when training with Curriculum

Hey guys,
I'm just trying to get my agent with curriculum training up and running. I am using ML-Agents 0.12.1
In my scene I have an agent and a NavMeshAgent.

To start the training, I use the command line from the docs:
mlagents-learn config/trainer_config.yaml --curriculum=config/curricula/chaser/ --run-id=chaser-curriculum_test --train

In the trainer_config.yaml I have defined the following parameters:

ChaserLearning:

    use_curiosity: true

    summary_freq: 1000

    time_horizon: 128

    batch_size: 128

    buffer_size: 2048

    hidden_unity: 256

    beta: 1.0e-2

    max_steps: 1.0e6

and I have created a folder in \ml-agents\config\curricula\ named chaser. In it is the file ChaserLearning.json.
The file has the following values:

{

    "measure" : "reward",

    "thresholds : [-0.1, 0.7, 1, 1.5, 1.7, 2],

    "min_lesson_length" : 80,

    "signal_smoothing" : true,

    "parameters" :

    {

        "speedMultiplicator" : [0.5, 0.75, 0.9, 1.0, 1.25, 1.5, 1.75]

    }

}

The BehaviourParameter Script in the Unity Editor is named chaser-Curriculum_test01
When I start the training, I get the following error in the shell:

KeyError: 'chaser-Curriculum_test01'
I have attached a screenshot of the shell below.



Did I make a mistake somewhere, or is it because of the ML-Agents version?

thanks in advance

 

Nvm,

I found my mistake. The Behaviour Name has to match the curriculum.json and the trainer_config.yaml Code Block.

So, in this case instead of chaser-Curriculum_test01 it has to be ChaserLearning.

 

Glad you got it figured out. For what it's worth, the error handling and messaging around this got better in 0.13.0.

 

I am currently writing my bachelor thesis on machine learning using the toolkit, so i wanted to stick with one version. This makes writing a bit easier and continuous.

The documentation is also a bit unclear about the point that the behavior name also has an influence and is not only meant to find the .nn model.
maybe the parameter behavior name could be added and clarified here.
https://github.com/Unity-Technologi...Environment-Design-Agents.md#agent-properties

 

Sure, I can clarify that in the docs (and also the section on curriculum learning)

 

Updated in https://github.com/Unity-Technologies/ml-agents/pull/3380

 

Thank you, appreciate it.