It's unclear from the docs how to set up a curriculum with multiple parameters

The docs show an example for only one environment parameter and it's not obvious how to work with multiple at a time. Are you supposed to set up multiple parameters, each with their own curriculum/lessons?

The goal would be to change more than one parameter at a time in a given lesson.

 

I assume the doc you're referring to is this one.

Yes your assumption is correct, the

environment_parameters:

section can contain multiple parameters each with their own custom curriculum.

Something like this -

Code (CSharp):

1. behaviors:
2.   BehaviorY:
3.     # < Same as above >
4.  
5. # Add this section
6. environment_parameters:
7.   my_environment_parameter:
8.     curriculum:
9.       - name: MyFirstLesson # The '-' is important as this is a list
10.         completion_criteria:
11.           measure: progress
12.           behavior: my_behavior
13.           signal_smoothing: true
14.           min_lesson_length: 100
15.           threshold: 0.2
16.         value: 0.0
17.       - name: MySecondLesson # This is the start of the second lesson
18.         completion_criteria:
19.           measure: progress
20.           behavior: my_behavior
21.           signal_smoothing: true
22.           min_lesson_length: 100
23.           threshold: 0.6
24.           require_reset: true
25.         value:
26.           sampler_type: uniform
27.           sampler_parameters:
28.             min_value: 4.0
29.             max_value: 7.0
30.       - name: MyLastLesson
31.         value: 8.0
32.  
33.   second_parameter:
34.     sampler_type: uniform
35.     sampler_parameters:
36.         min_value: 0.5
37.         max_value: 10
38.  
39.   third:
40.     curriculum:
41.       - name: Lesson1 # The '-' is important as this is a list
42.         completion_criteria:
43.           measure: progress
44.           behavior: my_behavior
45.           signal_smoothing: true
46.           min_lesson_length: 100
47.           threshold: 0.2
48.         value: 0.0
49.       - name: Lesson2 # This is the start of the second lesson
50.         completion_criteria:
51.           measure: progress
52.           behavior: my_behavior
53.           signal_smoothing: true
54.           min_lesson_length: 100
55.           threshold: 0.6
56.           require_reset: true
57.         value:
58.           sampler_type: uniform
59.           sampler_parameters:
60.             min_value: 4.0
61.             max_value: 7.0
62.       - name: FinalLesson
63.         value: 8.0

 

Ah ok, so since in my particular case I want to have multiple parameters effectively part of the same lesson, I'll just duplicate the lesson settings.

Thank you much for the help!!

 

How does the "threshold, signal_smoothing, min_lesson_length" work for multiple parameters? Which one gets used if the values are different and Lesson 0 has these duplicated several times (for different parameters) with different values?

 

All of them get used. Each parameter has its own curriculum that is run independently of other parameters. In my example those values are copy-pasted and are the same but in a real implementation they could be set to a different value that makes the most sense for that particular environment parameter.

 

Thanks for the clarification. I do agree that the docs could be improved. After reading them I had made the faulty assumptions that the lessons for all the parameters advanced in lockstep. When one parameter advanced to the next lesson then they all would.

 

Hi TreyK-47, thanks for the comment.

Is there documentation for ML-Agents in the official docs yet? When I search all the docs are on the github site. I don't see a preview link there. What is the best way to submit feedback on those docs. Submit an issue?