Data-driven Risk-aware Adversarial Analysis under Uncertainty
Abstract
The scientific objective of this project is to investigate integrated models and advanced computational methodologies for decision-making problems involving adversarial activities under uncertainty. In particular, we focus on problems with limited distributional information of random parameters, due to insufficient and/or time-sensitive data. The proposed generic frameworks are fundamental to analyzing complex, multi-entity systems and control processes that involve adversary-designer dynamics. In such a context, a designer aims to minimize the cost of design plus the maximum service disruption or system damage caused by an adversaryÕs reactions under the current design. In this proposal, we will consider (i) two-stage adversarial games, (ii) multi-stage adversarial dynamics, and (iii) risk-averse games with adversarial recourse and probabilistic constraints. These diverse settings are inspired by applications in critical infrastructures, communication networks, and sensor deployment, where multiple players take into account cost and reliability (risk) in their own objectives/constraints, while facing uncertainties from both nature and adversarial activities. We propose various risk measures to quantify system reliability and service quality in different contexts for players who may have diverse risk preferences and behavior. Methods to be employed: Central to our approach is a novel coupling of distribution-free data-driven optimization and parallel computing to tackle various adversarial problems (i), (ii), and (iii) in the above. This proposal tackles two state-of-the-art challenges in the integration of data and computing in complex systems. First, there lack the use of appropriate risk measures in the existing literature to characterize diverse risk behavior of decision makers in stochastic games. Second, few studies have investigated game dynamics embedded with data dynamics and unknown distributions. In this research, we take into account uncertainty from the nature, in addition to randomness associated with an adversaryÕs motives, incentives, resource constraints and behavior. These uncertainties lead to unprecedented challenge in the conceptualization, modeling, analysis of designer-adversary dynamics; they also necessitate the use of advances in online learning and statistical inference to overcome the lack of precise information. To this end, we will adopt adversarial models that represent the worst case within a large space, and progressively refine this space using observations of the adversary. We will design effective decomposition schemes and cutting-plane algorithms for large-scale optimization. Significance of the proposed effort: Successful implementations of the research will bridge the gap between optimization, high performance computing, and data science. We anticipate the research to generate economic gains and to enhance system reliability, with broader societal impacts. The results will advance the knowledge and techniques for stochastic / robust adversarial analysis, and further facilitate the fulfillment of missions in energy, environment, and national security. The outcomes of this research will contribute to effective methodologies for handling broader stochastic problems beyond the adversarial games, where optimizers need to take into account data uncertainty with limited distributional information and multiple risk-return tradeoff preferences. This abstract is publicly releasable.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1710102
Entities
People
- Siqian Shen
Organizations
- Army Contracting Command
- United States Army
- University of Michigan