Non-stochastic Control
Various earlier works have introduce a class of controllers which can be used to control a linear dynamical systems (known or unknown) in presence of non-stochastic adversarial noise and adversarial convex cost functions. They show a sublinear policy regret against the best policy in hindsight. However, they fail to generalize to systems following non-linear dynamics, which are very prevalent in real world. In this work we would try to establish a similar framework for non-linear systems following Hamiltonian dynamics in presence of adversarial non-stochastic noise and a fixed convex cost function using Bregman Divergence in an online setting. See the initial report below which was written after a literature survey and describes the problem in detail.
![](https://www.google.com/images/icons/product/drive-32.png)
Additional research reviews (authored by me) related to the project can be found here