Content area
Abstract
Within the context of decision-making, risk is a measure of potential harm to the decision-maker’s desired outcome, arising from the uncertainties associated with the likelihood of adverse events and the severity of their impacts. It can occur in socio-cyber-physical systems (SCPSs) due to various factors, such as natural disasters, system errors, communication delays, and adversarial attacks. The analysis and design of risks are crucial to maintain the safety, functionality, adaptability, efficiency, and autonomy of SCPS. Traditional risk management approaches focus on three aspects, namely controlling the behaviors of system users to avoid risks, robustifying system design to gain resilience against risks, and mitigating the economic impacts of risks once they occur. Successes in risk mitigation and control facilitate the developments and maintain the proper operations of various types of SCPSs, including industrial control systems, smart grids, intelligent transportation systems, and public health facilities.
Risks become increasingly sophisticated in modern SCPSs. Firstly, the massive integration of subsystems and system layers in SCPSs create challenges in risk assessment. System integration enables risk contagion at different levels. Structural configurations further increase the difficulty of risk monitoring in complex systems. Secondly, due to the existence of the social layer, the basic components in SCPSs require increased rationality and federation. The incentives and objectives of rational agents need to be coherently described by their risk perceptions, risk preferences, and feasible actions under risks. The heterogeneity and connectivity of agents greatly increase the difficulties of preference elicitation and calibration. Thirdly, new types of threats exist in SCPS, including cognitive vulnerabilities of humans and populations, cascading of risk sources and risk impacts, time-evolution of adversarial endeavors, and public and private misinformation. These novel challenges call for new frameworks for holistic risk analysis and design.
This dissertation introduces the concept of holonic risk and investigates its analytic and design aspects through the lens of system sciences. Holonic risk pertains to the potential failures of a self-managing subsystem as it interacts within a larger system comprising itself and other diverse and interconnected subsystems. We use holonic risks to emphasize the multi-factorial property of risk sources, the compounding effect of risk impacts caused by structural configurations and feedbacks of systems, and the multi-scale perspectives required to understand the interdependency between individual risk behaviors at the microscopic level and systemic risk patterns at the macroscopic level. We leverage theoretical tools and techniques from diverse disciplines to assess holonic risks, including network science, game and control theory, behavioral economics, operations research, and financial mathematics.
The contributions on the analysis and design of holonic risks in this dissertation is arranged in the following three parts. In Part I, we investigate the systemic control of risk through holonic incentive design. We develop two distinct population game-theoretic frameworks to model individual and social activities over networks and the correlation between activities and individual incentives. Structural properties of the games play important roles in ensuring the existence of Nash equilibria together with their evolutionary stability. These frameworks enable the control of herd behaviors at the macroscopic perspective by incentivizing players’ behaviors under risks caused by suboptimal strategies in local tasks, misalignment of local and global objectives, and contagion of risks through the networks. The frameworks provide insights on network configurations and community formations in applications such as federated learning and pandemic control.
In Part II, we aim to shape holonic risk behaviors through crowd preference maneuvering. To this end, we lay the mathematical foundations of the preference maneuvering mechanism based on populational hierarchical games. The risk preference design framework, for the first time, regards human preferences as a maneuverable cognitive resource and presents analytical frameworks for design. With our frameworks, holonic risk behaviors can be coherently planned and controlled using information technology (IT). We have explicitly studied the impact of risk preference maneuvering in insurance contract design problems and showed its potential in mitigating the moral hazard issue, which is regarded as a barrier to the development of the first-best contracts. Due to the hierarchical structure of the framework, it can also be applied naturally to scenarios such as traffic congestion mitigation, robust meta-machine learning, and multi-robot collaboration.
We devote Part III to the mitigation of holonic risk cascading using informational cognitive herding. The informational sensitivity of intelligent agents makes information perturbations detectable and designable. Our mathematical frameworks rigorously describe the generation, propagation, processing, and influences of information in complex systems that are dynamic and partially revealed. We discover that the trade-offs between the accuracy and the dispersion of information is key to manipulate risk behaviors of intelligent agents. With the advances in IT and the connectivity provided by networking, the frameworks developed in this part adapt to numerous application domains, including defensive cyber deception, human-robot collaboration in healthcare, and information campaigns for promoting fairness and justice.
The theoretical foundation presented in this dissertation introduces a novel approach for managing and designing SCPSs by integrating the concept of holonic risks. The interconnected functional layers in SCPSs and emerging technologies enhance system operational performance through new approaches like preference maneuvering and informational herding. Our methodologies currently support efficient and intelligent semi-autonomy in SCPSs. By reducing uncertainty and mitigating its impacts, our methodologies will drive modern SCPSs toward full autonomy, universal adaptability, and evolutionary sustainability.
