search

jc-bao / policy-adaptation-survey

This repository is for comparing the prevailing adaptive control method in both control and learning communities.
Apache License 2.0
7 stars 1 forks source link

Expert policy perform the same as vanilla policy #9

Closed jc-bao closed 1 year ago

jc-bao commented 1 year ago

Parameter range: 

        self.mass_min, self.mass_max = 0.005, 0.03
        self.delay_min, self.delay_max = 0, 0
        self.decay_min, self.decay_max = 0.0, 0.3

Performance:

image

Sensitivity analysis:

Vanilla Expert
image image
jc-bao commented 1 year ago

Try again with new parameters: 

        self.mass_min, self.mass_max = 0.1, 0.5 #0.005, 0.03
        self.delay_min, self.delay_max = 0, 0
        self.decay_min, self.decay_max = 0.0, 0.3
image
jc-bao commented 1 year ago

Problem resolved. During the evaluation, the curriculum parameter is reset to zero. And I forget to update the _set_disturb() function.

After fixing the related issue, the margin is coming back:

image