VMAS-1.4.0
Differentiable VMAS
That's right, VMAS is now fully differentiable!
How do I use it?
Just set grad_enabled=True
at environement construction time and have any input that requires gradients. This can be actions or scenario parameters. VMAS will keep track of the computation graph on that tensor over time.
What does it mean?
It means that you can differentiate any VMAS output, enabling differentaition of the transition dynamics, reward functions, and observation functions.
Why is it useful?
You can now optimize parameters in VMAS scenarios (e.g., parameters of the various scenario functions or simply initial state values) using losses computed on rewards or observations. It further allows you to backpropagate through time (simulation steps).
What's Changed
- [Feature] Differentiable VMAS by @matteobettini in #80
- [Refactor] DIff drive dynamics by @matteobettini in #81
Full Changelog: VMAS-1.3.4...VMAS-1.4.0