Browsing by Subject "data-driven"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Data-Driven Analysis and Insight of Human Motion
    (2020-01) Sohre, Nicholas
    Motion is a central element of the human experience. Artificial Intelligence (AI) and robotics technologies continue to transform society, but work is needed to enable solutions that engage with our motion-driven reality. Critical to an understanding human motion is the ability to model and accurately simulate virtual humans. To that end, my thesis provides data-driven analysis and insight for human motion. I identify two key aspects of realistic human motion simulations: being both \textit{natural} in appearance while covering the rich \textit{variety} of motions exhibited by humans. I describe how motion data can be leveraged to both simulate realistic motion, as well as validate simulation realism through a combination of data-driven analysis and user study approaches. Computational methods for human motion are largely studied in the context of computer graphics and virtual character animation. Drawing from and expanding on work in this field, my work applies data-driven methods for simulating humans in several settings: that of facial motion, local crowd simulation, and global navigation. The methods and analysis in this dissertation present contributions to the fields of AI, robotics, and computer graphics in supporting my thesis that data-driven methods can be used to create and validate realistic simulations of human motion. In the first part of my thesis, I study the simulation of realistic human smiles by conducting a large user study to connect observer reactions to computer animated faces. The result is a rich dataset providing value beyond that of this thesis to interdisciplinary research. I use the data to train a generative model with a new machine learning heuristic (PVL) that I develop, which tunes the trade-offs in creating a variety of happy smiles. I validate the realism of the PVL results with a follow up user study. The second part of my thesis studies the simulation of realistic human navigation. I perform a data-driven evaluation of the impact of collision avoidance on user experiences in virtual reality (VR), validating its importance for enabling the feeling of presence. I leverage motion data of shoppers to drive new insights for human navigation decisions, discovering an entropy law governing item retrieval patterns. Finally, I present a deep-learning technique (SPNets) for simulating realistic human navigation behaviors in indoor settings trained on optimal paths. The resulting agents exhibit several human-like behaviors, such as intelligent backtracking, narrowing down goal locations, and environment familiarity. I validate the realism of SPNet simulations using paths from a user study on the same navigation tasks.
  • Thumbnail Image
    Item
    Data-driven Distributionally Robust Stochastic Optimization via Wasserstein Distance with Applications to Portfolio Risk Management and Inventory Control
    (2020-12) Singh, Derek
    The central theme of this dissertation is stochastic optimization under distributional ambiguity. One canthink of this as a two player game between a decision maker, who tries to minimize some loss or maximize some reward, and an adversarial agent that chooses the worst case, or least favorable, distribution (to the decision maker) from some ambiguity set. The Wasserstein distance metric is used to specify the ambiguity set which is known as a Wasserstein ball of some finite radius d. At the center of this ball, is the empirical distribution, which serves as a proxy for the true underlying distribution. In that sense, this line of research has been called data-driven robust optimization in the academic literature. The primal problem is infinite dimensional since the Wasserstein ball contains all finite and discrete distributions within distance d of the empirical distribution. As such, it would appear more difficult to solve the stochastic optimization problem in this setting. This research makes use of (recent) Lagrangian duality results in distributional robustness and (classical) moments duality results to formulate and solve the simpler finite dimensional dual problem. Different problem formulations are considered, both with and without moment constraints on the ambiguity set. Some interesting practical applications of these results include single stage and multistage problems in portfolio risk management and inventory control. We also investigate the notion of time consistency between the static and dynamic (multi-period) problem formulations. Time consistency is a desirable property in that the decision maker knows that the optimal policy determined at time zero will not change as realizations of the data process and corresponding system state are observed. In particular, this dissertation considers optimal decision making for portfolio problems in counterparty credit risk, funding risk, statistical arbitrage, option exercise, asset purchasing/selling, and quantification of certain profit and risk metrics. In addition, we consider the classical newsvendor model (both with and without moment constraints) in the single period and multi-period settings. We conclude with some commentary on our findings throughout this work and provide some suggestions for further research.
  • Thumbnail Image
    Item
    Real-Time System Identification and Control of Engine System Using Least Squares Learning and Simplex Tessellation
    (2022-12) Tranquillo, Holden
    To aid in engine control for achieving the stable combustion of varying cetane level fuels, a computationally efficient algorithm for the online learning of an engine model based on real-time input and output measurements is developed. Innovations in engine technology has led to the feasibility of robust, multi-fuel engine systems capable of operating on unknown or non-ideal fuel types. To attain such performance, advanced control strategies must be implemented in order to achieve stable engine combustion using such fuels. The method developed in this work, based on piecewise-linear modeling via discrete nodes and recursive linear least squares is first derived for the one-dimensional system of injection timing and combustion phasing. The learning model is then used for adaptive feedforward and feedback control of the SISO system in simulation using a gaussian process model as a virtual engine. The algorithm is then extended to the two-input/two-output system of injection timing and fuel mass and their effect on combustion phasing and indicated mean effective pressure (IMEP). Data generated using computational fluid mechanics is used to supplement experimental data in the development of the 2D model. The theory of barycentric and affine coordinates is introduced and applied to the concept of piecewise planes to approximate nonlinear surfaces. The learning model is utilized in an adaptive MIMO feedforward algorithm to control the engine to a desired combustion phasing and IMEP. Additionally, a decoupled integral feedback control scheme is presented and shown effective in simulation. A generalization of the learning algorithm for higher dimensions is made in order to model higher order systems. Specifically, simplex tessellation and barycentric coordinates as regressor coefficients are shown to generalize node locating and updating in arbitrary dimensions. The generalized learning algorithm is applied to a synthetic three-input data set in order show feasibility of the model for higher order nonlinear systems. The algorithm developed in this work is a unique, generalized, data-driven model capable of the real-time learning and control of multi-dimensional systems. The computational efficiency and generalization of the method allows for the real-time system identification of engine systems that are operating in unknown or untested environments. Existing engine models lack the efficiency to perform at the operating times seen in internal combustion engines. Implemented in a physical engine, the developed algorithm could be used for adaptive modeling of the system when undergoing a fuel or environmental change, which in turn can be used to aid in adaptive control of the engine. In commercial application, the real-time learning model could be used to decrease or eliminate the traditional in-house testing of engines required for lookup table generation, which would in turn decrease the time and cost in getting the engine to final application.