Title:  On the Modeling of Dynamic-Systems using Sequence based Deep-Neural-Network

Committee: 

Dr. Pradalier, Advisor 

Dr. Locquet, Co-Advisor  

Dr. Bloch, Chair

Dr. Kira

Abstract: The objective of the proposed research is the adaptation and development of sequence-based deep-neural-network applied to the modeling of dynamic systems. More specifically, we will focus our study on 3 sub-problems: the modeling of time-series, the modeling and control of multi-input multi-output systems, and the analysis of time-dependent image-series. These 3 sub-problems will be explored through the modeling of crops, the modeling and control of robots, and finally, the modeling of trees' internal structures. To solve these problems we build on state of the art neural-networks 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 neural-network are powerful tools capable of modeling the water consumption of crops while observing only a portion of what is currently required by the reference methods. In robotics, we show that prioritization techniques can be used to better learn robot dynamic models. Additionally, we use model-based RL to solve a sensor-based modeling and control problem. In doing so, we demonstrate that RL-based controllers can be used for sensor-based control on real robots without being trained on this robot. Finally, we built methods capable of modeling systems from sequences of images. With our method, we showed that one could infer the internal structure of trees from the trees' outer shape through the processing image sequences. What is further planned is the extension of our work on model-based RL. More specifically, we are working on implementing a physics guided variant of DREAMER. And we are diversifying the tasks and robots we apply dreamer to. We will also expand our work on trees and apply it to six new coniferous trees species. And finally, we are preparing this summer irrigation experiments with new models deep-neural-networks models that can estimate their uncertainty.