Information Constrained Control and Prediction of Invariances in Dynamical Systems

Dynamic identification of a target is a basic capability that guided weapons need for robust control to reduce the miss distance. This could be achieved by using a physical model of the target (Peled-Eitan et al., 2013). The actual physical model of the target is unknown or too complicated than actually required for the efficient control. A complex dynamical system requires lots of details to be modeled and controlled.

We plan to validate and demonstrate of the developed algorithms and theoretical results in two level of simulations: simplified and systematic simulation. Where the simplified simulation enables to build the concept and demonstrate the theoretic results. While the systematic simulation enables to simulate the real-world complexity and constraints. It is composed of the real system with full degree-of-freedom and a model of the physics environment. This simulation enables to learn the real physics environment according to emulations and to demonstrate the engagement testing without any experiments.
At the Hebrew University, we have 4 clusters of GPU’s for training large machine learning models, as well as top graduate students. In addition, we have access to Intel, Google, and NVIDIA computational facilities for machine learning in general and Neural Networks in particular.
The University of Hertfordshire offers one cluster shared between Computer Science and Astrophysics for large-scale machine-learning and astronomical analysis; we are equipped with several robotic laboratories, including 3D printing facilities, and various kinds of robotic devices, including humanoid robots; this includes robots such the humanoid Darwin or iCub robots, or Baxter equipped with arm grippers which offer the possibility for additional, hardware-based, test scenarios, beyond the core scenarios of the proposal.