The iLQG nominal rollout, defined here, is inefficient. We should do this with a threadpool that rolls out with a parallel "line search" on the feedback term, that includes a 0 coefficient (no feedback, just open-loop). For this, the Action method should also get a feedback_scale parameter (defaulting to 1), and this line should use mju_addToScl.
This is a very simple change that I suspect will improve iLQG significantly.
thowell
commented
1 year ago
The iLQG nominal rollout, defined here, is inefficient. We should do this with a threadpool that rolls out with a parallel "line search" on the feedback term, that includes a 0 coefficient (no feedback, just open-loop). For this, the
Actionmethod should also get afeedback_scaleparameter (defaulting to 1), and this line should usemju_addToScl.This is a very simple change that I suspect will improve iLQG significantly.