UC Berkeley

This dissertation thesis explores Human-Centric Cyber-Physical Systems by simultaneously considering users' behavior/preference and their interaction as strategic agents. We envision smart-building systems in which humans take control, interact, and improve the environment they live in. People's interaction in a cyber-physical system is a core mechanism of the implementation of smart building technology. Adoption of human-centric building services and amenities also leads to improvements in the operational efficiency of cyber-physical systems that are used to control building energy usage. However, human preference in regard to living conditions is usually unknown and heterogeneous in its manifestation as control inputs to a building. Furthermore, the occupants of a building typically lack the independent motivation necessary to contribute to and play a key role in the control of smart building infrastructure. We focus on the development of a generalized gamification abstraction towards enabling strategic interactions among non-cooperative agents in Human-Centric Cyber-Physical Systems. The proposed framework enables a humans-in-the-loop strategy using an interface to allow building managers to interact with occupants. This interface is designed for occupants' engagement---integration while it supports learning occupants' preferences over shared or scarce resources in addition to understanding how preferences change as a function of external stimuli such as physical control, time or incentives. Our gamification framework can be used in the design of incentive mechanisms that realign agents' preferences with those of the planner which often represent system-level performance criteria through fair compensation.

In our first approach, we model user interaction as a continuous game between non-cooperative players. Game theoretic analysis often relies on the assumption that the utility function of each agent is known a priori; however, this assumption usually does not hold in many real-world applications. We propose a parametric utility learning framework leveraging inverse optimization techniques and explore vulnerability from adversarial attacks in utility learning and present potential security risks. A generalized robust framework of the proposed learning method is introduced by employing constrained feasible generalized least squares estimations with heteroskedastic inference. We further develop the theoretical formulation of a new parametric utility learning method that uses a probabilistic interpretation---i.e. a mixture of utilities---of agent utility functions that allows us to account for variations in agents' parameters over time. Furthermore, towards the reduction of the complexity of the advanced learning methods we propose a new method of data-driven modeling of human decision-making by accounting for possible correlations between players and form coalitions between agents.

Advancements in cyber-physical systems lead to the collection of more and more data as a result of users' interactions with cyber-physical systems' sensing/actuation platforms. This enables new ways to improve infrastructure systems and lead to smart-building energy efficiency. Towards modeling users in their engagement and integration in a Human-Centric Cyber-Physical System, we characterize their interaction as a sequential discrete game between non-cooperative players. We propose the design and implementation of a large-scale network gamification application with the goal of improving the energy efficiency of a building through the utilization of cutting-edge Internet of Things (IoT) sensors and cyber-physical systems sensing/actuation platforms. Then, by observing human decision-makers and their decision strategies in their operation of building systems, we can apply inverse learning techniques in order to estimate their utility functions. We propose a benchmark utility learning framework that employs robust estimations for classical discrete choice models provided with high dimensional imbalanced data. To improve forecasting performance, we extend the benchmark utility learning scheme by leveraging Deep Learning end-to-end training with Deep bi-directional Recurrent Neural Networks. Most importantly, we use conventional deep variational auto-encoders and recurrent network based adaptation of variational auto-encoders as an approach to create nonlinear manifolds (encoders) that can be used as a generative model of agents' decision-making process.

A series of experimental trials was conducted to generate real-world data, which was then used as the main source of data for our approaches. We apply the proposed methods to data from social game experiments designed to encourage energy efficient behavior among smart building occupants in the Nanyang Technological University (NTU) residential housing and the Center for Research in Energy Systems Transformation (CREST) on the UC Berkeley campus. This differentiates our work from a large portion of other works in the same field that use simulations in lieu of experimental methods.