Trustworthy AI for safety-critical systems

We develop theory in the areas of optimization, control, and learning inspired by real-world applications.
― Our research philosophy

In the future, data and algorithms will play an increasingly important role in solving societal-scale problems, from improving people's living conditions to modernizing the electric grid. Unlike traditional machine learning tasks, these problems involve complex physical systems that are safety-critical with humans included. Consequently, these systems require new paradigm of AI that can be trusted to do the critical work.

We design computational methods that are robust (to uncertainty and anomalies), resilient (to perturbations and faults), explainable (for human-in-the-loop decision making), and adaptable (to fast changing dynamics), with the ultimate goal of deploying in real-world safety-critical systems such as power grids and urban habitats. Along the way, we develop fundamental theories in the areas of optimization, control, and learning to solve these challenging problems.


Nonconvex optimization
Robust optimization
Time-varying optimization


Distributed control
Robust control
Adaptive control


Online learning
Robust learning
Reinforcement learning

Video Introduction

Presentation at the Engineering Faculty Organization (EFO) faculty meeting, Virginia Tech.

Domain impacts

Power system is a complex large-scale system that generates, transports and distributes electricity to millions of people. In real-time, operators need to solve fundamental problems that range from discrete optimization (e.g., unit commitment) to continuous operational problems (e.g., optimal power flow and state estimation), which are inherently nonconvex due to laws of physics, dynamically changing due to renewables, and increasingly data-driven. A 1% of improvement in the decision-making process of power systems amounts to billions of dollars in savings, while a single blackout or fault may set the economy back (especially when it causes widespread fire). My research in this domain solves problems of nonconvex and large-scale nature with optimality and cyber-physical robustness guarantees. For example, my work led to the first vulnerability map of power system state estimation to cyberattacks for the U.S. grid that has implications for regulators and utility companies to secure the next-generation grid.

Vulnerability map of the U.S. grid against cyberattacks
Vulnerability map of the U.S. grid against cyberattacks

Buildings present unique opportunities to study people and their behaviors and how human intervention can be characterized, monitored and shaped. Tools created in this domain need to address human data that are limited in quantity, uncertain, dynamic and private, and have natural extensions to other domains where humans are engaged as users or operators, such as autonomous cars, human-robot cooperation and healthcare. I build intelligent systems from sensing, inference and control, to optimal design and gamification, bridging between theory and practice. My collaborative works have received 3 best paper awards and are featured in a cover story (2019) and public media such as IEEE Spectrum (2017).

Highlighted projects

Optimization is ubiquitous in data mining applications, from the classic least-squares regression to newer problems of compressed sensing and collaborative filtering. In this line of research, we considered the case when there are a few random or adversarial "bad data," and we developed theoretical and computational foundations for adversarially robust optimization. An important application is power system state estimation, which plays a critical role in the economic and reliable operation of the grid. The problem is challenging because of nonconvexity and vulnerability to bad data, and common techniques in the literature cannot address either. To evaluate the vulnerability, we have proposed the first algorithm based on penalized semidefinite programming (SDP), which can provably solve the nonconvex strongest stealthy attack problem under reasonable assumptions [1]. Taking a step forward, we have proposed a boundary defense mechanism by building upon our previous method to detect randomly occurring bad data [2, 3]. We have developed a theoretical tool called "graphical mutual incoherence" (gMI) to study the vulnerability of the grid against cyberattacks [3]. This allows the operator to provably detect the geographical boundary of the cyberattack via a novel state estimation technique as long as the gMI conditions are satisfied for each line on the boundary. With our tool, for the first time, one can generate a single vulnerability map of the U.S. grid that addresses the number of attack scenarios much higher than the number of atoms in the observable universe. This work can potentially lead to completely new grid hardening strategies that has implications for regulators and utility companies.

Selected publications:
[1] M. Jin, J. Lavaei, and K. Johansson, "Power grid AC-based state estimation: vulnerability analysis against cyber attacks." IEEE TAC (2018)
[2] M. Jin, I. Molybog, R. Mohammadi, and J. Lavaei, "Scalable and robust state estimation from abundant but untrusted data." IEEE TSG (2019)
[3] M. Jin, J. Lavaei, S. Sojoudi, and R. Baldick, "Boundary Defense against Cyber Threat for Power System Operation." IEEE TIFS (2020) (supplementary)

Graph-structured data are ubiquitous in physical networks such as power grids and transportation networks. Graph convolutional networks are powerful generalizations of convolutional architectures from regular grid structures (e.g., images, sound, text) to graph-structured data. However, this type of method is vulnerable to adversarial attacks on topological information, which is discrete in nature. Existing architectures can be generally summarized as designing spectral functions for the graph Laplacian operator. Nevertheless, for adversarial robustness, a key property is "spectral separation," which is not enjoyed by graph Laplacian. Our work goes beyond classical spectral theory and involves robust spectral theory, which develops a new graph convolutional operator that provably possesses spectral separation and can be incorporated in data-driven methods for end-to-end learning on graphs [1]. Furthermore, we have proposed a robust training paradigm called "graph augmentation," which generates a sequence of higher-ordered graphs from the original graph that spans a range of spectral and spatial behaviors to facilitate learning a transferrable representation. Through extensive experiments, we have shown that our method can simultaneously improve accuracies in both benign and adversarial settings against an array of strong attackers.

Selected publication:
[1] M. Jin, H. Chang, W. Zhu, and S. Sojoudi, "Power up! Robust Graph Convolutional Network based on Graph Powering," AAAI (2021)

By interacting with the environment and learning from experiences, reinforcement learning (RL) has surpassed human-level performance in Atari games and playing chess. However, RL is far from ready to be deployed in safety-critical systems due to the lack of performance guarantees under complicated and uncertain dynamics. It is imperative to certify the robustness of RL-controlled systems against natural or even adversarial perturbations. This poses fundamental challenges since (1) both the world dynamics and the neural-network policy are highly nonlinear and (2) the policy continuously changes during online learning. We have proposed a computational method to certify a "safety set" of policies based on conic optimization and integral quadratic constraints (IQC), so that the RL policy would be adversarially robust as long as it stays within this safety set [1, 2]. For example, we have demonstrated our results in a multi-agent RL setting to solve distributed control problems for a power system, which are NP-hard in general. With existing RL algorithms, the system becomes unstable at some point, which eventually leads to catastrophic failure; however, by keeping the policies within the safety set, the agents continue improving without destabilizing the grid. This work contributes to the field of safe RL from a control-theoretic perspective.

Selected publications:
[1] F. Gu, H. Yin, L. El Ghaoui, M. Arcak, P. Seiler, and M. Jin, "Recurrent Neural Network Controllers Synthesis with Stability Guarantees for Partially Observed Systems." AAAI (2022)
[2] H. Yin, P. Seiler, M. Jin, and M. Arcak, "Imitation Learning with Stability and Safety Guarantees." IEEE Control Systems Letters (2021)
[3] M. Jin and J. Lavaei, "Control-theoretic analysis of smoothness for stability-certified reinforcement learning." IEEE CDC (2018)
[4] M. Jin and J. Lavaei, "Stability-certified reinforcement learning: a control-theoretic perspective." IEEE Access (2020)

Human behaviors may appear complicated. However, as the Nobel Laureate Daniel Kahneman states in his book "Thinking: Fast and Slow," our behavior is fundamentally determined by two different mechanisms -- one subconscious and fast (system 1), and the other deliberate and slow (system 2). This naturally provides a framework to characterize and quantify human uncertainty. In this line of research, we have developed and implemented algorithms in the built environment to extract behavioral information about people (system 1), such as location and activities [1-5]. These physical footprints reveal face-to-face interactions as strong predictors of social dynamics and organizational productivity compared to electronic footprints such as emails and social media data. By observing the variation in physical footprints, we aim to understand people's conscious choices, such as preferences (system 2). The challenge is to address the uncertainty of behaviors. We have developed inverse reinforcement learning based on a deep Bayesian network to learn complex human preference functions. The advantage of this method over existing ones is that it is data-efficient and can model complex decision-making behaviors and deal with uncertainty in a principled way, as demonstrated in learning driving behaviors [6].

Selected publications:
[1] M. Jin, R. Jia, and C. Spanos, "Virtual occupancy sensing: Using smart meters to indicate your presence." IEEE TMC (2017) (supplementary | conference version at ACM BuildSys (2014) | IEEE Spectrum featured article)
[2] M. Jin, N. Bekiaris-Liberis, K. Weekly, C. Spanos, and A. Bayen, "Occupancy detection via environmental sensing" IEEE TASE (2017) ( code and data | synopsis | conference version at UBICOMM (2015) | Featured article) (Best Paper Award)
[3] M. Jin, H. Zou, K. Weekly, R. Jia, A. Bayen, and C. Spanos, "Environmental sensing by wearable device for indoor activity and location estimation." IEEE IECON (2014) (poster)
[4]W. Gu, M. Jin, C. Spanos, and L. Zhang, "MetroEye: Towards a fine-grained passenger tracking under the ground." MobiQuitous (2016) (poster at Ubicomp 2016) (Best Paper Runner-up)
[5] H. Zou, M. Jin, H. Jiang, L. Xie, and C. Spanos, "WinIPS: An WiFi-based non-intrusive indoor positioning system enabling online radio map construction." IEEE TWC (2017)
[6] M. Jin, A. Damianou, P. Abbeel, and C. Spanos, "Inverse reinforcement learning via deep Gaussian Process." UAI (2017) (code | supplementary)

In many cases (e.g., learning about people, assessing the risk of a cyber intrusion), it could be very costly and sometimes prohibitive to collect data. A question arises as to how one can build models and learn representations that generalize to the open world with limited data? We have explored two paradigms to tackle this challenge. The first paradigm is weak supervision, where we have leveraged domain knowledge to initial labels with noisy labels and learned a classifier while accounting for the label noise [1]. The second paradigm is "sensing by proxy," which builds a physics-inspired control-theoretic model [2]. We have demonstrated their superior performance over the existing methods without collecting any ground truth labels [1, 2], and observed that the models were also transferrable to data with different distributions [1].

Selected publications:
[1] M. Jin, R. Jia, and C. Spanos, "Virtual occupancy sensing: Using smart meters to indicate your presence." IEEE TMC (2017) (supplementary | conference version at ACM BuildSys (2014) | IEEE Spectrum featured article)
[2] M. Jin, N. Bekiaris-Liberis, K. Weekly, C. Spanos, and A. Bayen, "Occupancy detection via environmental sensing" IEEE TASE (2017) ( code and data | synopsis | conference version at UBICOMM (2015) | featured article) (Best Paper Award)

Unlike physical systems, people are complex social species that tend to resist "control." A question arises as to how human uncertainty can be shaped to benefit social welfare? In a community-based microgrid, we have demonstrated the possibility of shaping demand curves with economic incentives [1, 2], and shown that with the proper design of mechanisms, it is possible to reduce consumer costs and increase operator profits simultaneously [2]. However, economic incentives could be ineffective in many other situations. We have explored a gamification approach to nudging people by appealing to their "inner child" [3-5]. The social game also provides an easy method for collecting data about people's preferences with high granularity, which are otherwise difficult to obtain through traditional surveys. We have proposed an inverse optimization framework to estimate people's utility functions [3]. This framework has been further extended to a robust inverse optimization framework [4] to account for mixed strategies [5] and coalitions within users \cite{konstantakopoulos2017leveraging}. By closing the loop of estimation and nudging, we have been able to raise people's environmental awareness and save energy.

Selected publications:
[1] M. Jin, W. Feng, P. Liu, C. Marnay, and C. Spanos, "MOD-DR: Microgrid optimal dispatch with demand response." Applied Energy (2017)
[2] M. Jin, W. Feng, C. Marnay, and C. Spanos, "Microgrid to enable optimal distributed energy retail and end-user demand response." Applied Energy (2018) (poster on toolset MODEST)
[3] L. Ratliff, M. Jin, I. Konstantakopoulos, C. Spanos, and S. Sastry, "Social game for building energy efficiency: Incentive design." Allerton (2014)
[4] I. Konstantakopoulos, L. Ratliff, M. Jin, and C. Spanos, "Leveraging correlations in utility learning." IEEE TCST (2017) (slides)
[5] I. Konstantakopoulos, L. Ratliff, M. Jin, C. Spanos, and S. Sastry, "Inverse modeling of non-cooperative agents via mixture of utilities." IEEE CDC (2016)
[6] I. Konstantakopoulos, L. Ratliff, M. Jin, and C. Spanos, "Leveraging correlations in utility learning." IEEE ACC (2017) (slides)