ModeOpt: Mode remaining learning-based control for multimodal dynamical systems in TensorFlow/GPflow.
Disclaimer: This is unfinished research code accompanying my PhD.
ModeOpt is a package for learning and controlling unknown, or partially unknown, multimodal dynamical systems.
In particular, it is concerned with methods for learning and control that attempt to remain in a given desired dynamics
mode. For example, if some of the dynamics modes are believed to be unoperatable.
ModeOpt learns representations of multimodal dynamical systems using the Mixture of Gaussian Process Experts model from mogpe.
It then deploys multiple control strategies (trajectory optimisation algorithms) that make decisions
under the uncertainty of the learned dynamics model.
ModeOpt consists of trajectory optimisers with two main goals:
- Find trajectories between a start and end state that remain in a given dynamics mode and attempt to avoid regions of the dynamics with high epistemic uncertainty.
- Find trajectories that guide exploration of the state-control space whilst remaining in a given desired dynamics mode.