Ming Jin

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2017-228

December 15, 2017

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-228.pdf

The goal of this research is to enable optimal human-cyber-physical systems (h-CPS) by data-efficient analytics. The capacities of societal-scale infrastructures such as smart buildings and power grids are rapidly increasing, becoming physical systems capable of cyber computation that can deliver human-centric services while enhancing efficiency and resilience. Because people are central to h-CPS, the first part of this thesis is dedicated to learning about the human factors, including both human behaviors and preferences. To address the central challenge of data scarcity, we propose physics-inspired sensing by proxy and a framework of ``weak supervision'' to leverage high-level heuristics from domain knowledge. To infer human preferences, our key insight is to learn a functional abstraction that can rationalize people's behaviors. Drawing on this insight, we develop an inverse game theory framework that determines people's utility functions by observing how they interact with one another in a social game to conserve energy. We further propose deep Bayesian inverse reinforcement learning, which simultaneously learns a motivator representation to expand the capacity of modeling complex rewards and rationalizes an agent's sequence of actions to infer its long-term goals.

Enabled by this contextual awareness of the human, cyber, and physical states, we introduce methods to analyze and enhance system-level efficiency and resilience. We propose an energy retail model that enables distributed energy resource utilization and that exploits demand-side flexibility. The synergy that naturally emerges from integrated optimization of thermal and electrical energy provision substantially improves efficiency and economy. While data empowers the aforementioned h-CPS learning and control, malicious attacks can pose major security threats. The cyber resilience of power system state estimation is analyzed. The envisioning process naturally leads to a power grid resilience metric to guide ``grid hardening.'' While the methods introduced in the thesis can be applied to many h-CPS systems, this thesis focuses primarily on the implications for smart buildings and smart grid.

Advisors: Costas J. Spanos


BibTeX citation:

@phdthesis{Jin:EECS-2017-228,
    Author= {Jin, Ming},
    Title= {Data-efficient Analytics for Optimal Human-Cyber-Physical Systems},
    School= {EECS Department, University of California, Berkeley},
    Year= {2017},
    Month= {Dec},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-228.html},
    Number= {UCB/EECS-2017-228},
    Abstract= {The goal of this research is to enable optimal human-cyber-physical systems (h-CPS) by data-efficient analytics. The capacities of societal-scale infrastructures such as smart buildings and power grids are rapidly increasing, becoming physical systems capable of cyber computation that can deliver human-centric services while enhancing efficiency and resilience. Because people are central to h-CPS, the first part of this thesis is dedicated to learning about the human factors, including both human behaviors and preferences. To address the central challenge of data scarcity, we propose physics-inspired sensing by proxy and a framework of ``weak supervision'' to leverage high-level heuristics from domain knowledge. To infer human preferences, our key insight is to learn a functional abstraction that can rationalize people's behaviors. Drawing on this insight, we develop an inverse game theory framework that determines people's utility functions by observing how they interact with one another in a social game to conserve energy. We further propose deep Bayesian inverse reinforcement learning, which simultaneously learns a motivator representation to expand the capacity of modeling complex rewards and rationalizes an agent's sequence of actions to infer its long-term goals. 

Enabled by this contextual awareness of the human, cyber, and physical states, we introduce methods to analyze and enhance system-level efficiency and resilience. We propose an energy retail model that enables distributed energy resource utilization and that exploits demand-side flexibility. The synergy that naturally emerges from integrated optimization of thermal and electrical energy provision substantially improves efficiency and economy.  While data empowers the aforementioned h-CPS learning and control, malicious attacks can pose major security threats. The cyber resilience of power system state estimation is analyzed. The envisioning process naturally leads to a power grid resilience metric to guide ``grid hardening.'' While the methods introduced in the thesis can be applied to many h-CPS systems, this thesis focuses primarily on the implications for smart buildings and smart grid.},
}

EndNote citation:

%0 Thesis
%A Jin, Ming 
%T Data-efficient Analytics for Optimal Human-Cyber-Physical Systems
%I EECS Department, University of California, Berkeley
%D 2017
%8 December 15
%@ UCB/EECS-2017-228
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-228.html
%F Jin:EECS-2017-228