search

tcstewar / 2015-Embodied_Benchmarks

Paper on Embodied Neuromorphic Benchmarks
GNU General Public License v2.0
0 stars 0 forks source link

L499 #16

Open celiasmith opened 8 years ago

celiasmith commented 8 years ago

https://github.com/tcstewar/2015-Embodied_Benchmarks/blob/master/paper/paper.tex#L499

It would be nice to compare to a PID controller, perhaps... but only if the PID is better than PD. It should be if K_I is chosen well. I also noticed that in the PD control you choose 2 and 1 for the gains. Probably not worth it, but things like the ZN method could be compared to: https://en.wikipedia.org/wiki/PID_controller#Ziegler.E2.80.93Nichols_method In any case, the point of the paper is partly that the gains don't have to be carefully tuned. It sort of seems something should be done for I, but it's not clear what.

tcstewar commented 8 years ago

The PID is either about the same or way worse, at least for the parameter settings I could find. The problem is the wild change in the external force at different setpoints. This is the "integral windup" problem: https://en.wikipedia.org/wiki/PID_controller#Integral_windup

And yeah, tuning of the PD controller would be good. The problem is that it could also be argued that they should be re-tuned for every randomly generated environment that it's run in... But I suppose what I could do is turn off the external force, do ZN to get some good values, and leave it there. I'll poke with that and see if it changes anything significantly....