*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: On the Modeling of Dynamic-Systems using Sequence-based Deep Neural-networks
Committee:
Dr. Cedric Pradalier, ECE, Chair, Advisor
Dr. Alexandre Locquet, ECE, Co-Advisor
Dr. Matthieu Bloch, ECE
Dr. Fumin Zhang, ECE
Dr. Matthieu Geist, Google
Dr. Zsolt Kira, IC
Abstract: The objective of this thesis is the adaptation and development of sequence-based Neural-Networks (NNs) applied to the modeling of dynamic systems. More specifically, we will focus our study on 2 sub-problems: the modeling of time-series, the modeling and control of Multi-Inputs Multi-Outputs (MIMO) systems. These 2 sub-problems will be explored through the modeling of crops, and the modeling and control of robots. To solve these problems, we build on NNs and training schemes allowing our models to out-perform the state-of-the-art results in their respective fields. In the irrigation field, we show that NNs are powerful tools capable of modeling the water consumption of crops while observing only a portion of what is currently required by reference methods. We further demonstrate the potential of NNs by inferring irrigation recommendations in real-time. In robotics, we show that prioritization techniques can be used to learn better robot dynamic models. We apply the models learned using these methods inside an \gls{mpc} controller, further demonstrating their benefits. Additionally, we leverage Dreamer, an Model-Based Reinforcement Learning (MBRL) agent, to solve visuomotor tasks. We demonstrate that MBRL controllers can be used for sensor-based control on real robots without being trained on real systems. Adding to this result, we developed a physics-guided variant of Dreamer. This variation of the original algorithm is more flexible and designed for mobile robots. This novel framework enables reusing previously learned dynamics and transferring environment knowledge to other robots. Furthermore, using this new model, we train agents to reach various goals without interacting with the system.This increases the reusability of the learned models and makes for a highly data-efficient learning scheme. Moreover, this allows for efficient dynamics randomization, creating robust agents that transfer well to unseen dynamics.
