Closed farzadab closed 5 years ago
The torque limit experiments show that the curriculum is working great for low torque limits but seems to be bad for higher starting limits; bug maybe? Yes! The num_step
was changed between runs and apparently it wasn't even doing curriculum!
Figure 1: Walker2D final reward (test) with different torque limits (x-axis).
Figure 1: Walker2D curriculum final reward (test) with 1.6x starting torque limit and different final torque limits (x-axis).
Same plots but nicer and put side-by-side: Figure 3: Final reward achieved with different torque limits (x-axis).
Figure 4: Final progress with different torque limits (x-axis).
Re: what is the best value to start from? Can't really say, there's no definitive answer here:
Figures: Final progress/reward with different starting torque limits (x-axis) for the target of 0.6x limit.
[x] scatterplot of final torque limit vs final reward grid after curriculum