Yao Ji

H. Milton Stewart Postdoctoral Fellow

H. Milton Stewart School of Industrial and Systems Engineering

Georgia Institute of Technology · Atlanta, GA

Beyond research

Operations Researcher

I am an H. Milton Stewart Postdoctoral Fellow in the H. Milton Stewart School of Industrial and Systems Engineering at Georgia Tech mentored by Prof. Guanghui (George) Lan.

My current focus is on optimal parameter-free methods for stochastic optimization. I study first- and higher-order methods for convex and nonconvex problems in both deterministic and stochastic settings. I have worked on gradient minimization, variational inequalities, and reinforcement learning, including applications in healthcare.

- Branching Process
- Bellman–Harris Process
- Distributed Optimization
- Statistical Learning
- Parameter-free
- First / Higher-order
- Gradient Minimization
- Variational Inequalities
- Reinforcement Learning
- Healthcare Applications

Research Interests and Selected Works

Stochastic optimization; reinforcement learning; healthcare.

Current focus

Parameter-Free Optimization
Stochastic Optimization
Higher-Order Optimization
Variational Inequalities
Reinforcement Learning
Healthcare
Distributed Optimization

Foundation

Standard AGD Nesterov ’83 $\mathcal{O}(1/\sqrt{\varepsilon})$, known $L$

+ stochasticity

AC-SA Lan ’12 $\mathcal{O}\!\left(\tfrac{1}{\sqrt{\varepsilon}} + \tfrac{\sigma^2}{\varepsilon^2}\right)$, known $L$

+ adaptivity
− line search

+ adaptivity

AC-FGM Li & Lan ’25 $\mathcal{O}(1/\sqrt{\varepsilon})$

+ stochasticity

Stoch. AC-FGM ★ This paper $\mathcal{O}\!\left(\tfrac{1}{\sqrt{\varepsilon}} + \tfrac{\sigma^2}{\varepsilon^2}\right)$

Lower bound $\Omega(1/\sqrt{\varepsilon} + \sigma^2/\varepsilon^2)$

Selected work

Stochastic Auto-conditioned Fast Gradient Methods with Optimal Rates

Optimal iteration and sample complexities for parameter-free stochastic acceleration.

Parameter-Free Optimization
Stochastic Optimization

Yao Ji and Guanghui Lan

arXiv:2604.06525, 2026

Under review, Mathematical Programming

Research Snap, page 1 of 3 — "What is the problem?" Handwritten notes defining a variational inequality with an L-Lipschitz, monotone operator, then comparing termination criteria: weak gap, strong gap, distance to the optimal solution, and the residual. Notes that the residual works for unbounded feasible sets, and sets out the stochastic case, where the classical uniform noise model and a distance-based condition both fail for merely monotone problems — motivating a new noise model.

Research Snap, page 2 of 3 — "What are our results?" Introduces the state-dependent noise model, shows it is well defined and consistent with the strong-gap interpretation when the solution is not unique, and states the main complexity results for the monotone and strongly monotone cases, which match the lower bound up to logarithmic factors and improve on the existing literature.

Selected work

Computation of Strong Solutions to Stochastic Variational Inequalities

Optimal residual complexity for computing strong solutions to stochastic variational inequalities.

Variational Inequalities
Reinforcement Learning

Yao Ji, Guanghui Lan, and Jason Zhu

arXiv:2609.04188, 2026

Under review, Mathematical Programming

Paper — Computation of Strong Solutions to Stochastic Variational Inequalities

$\mathcal{A}$ any subroutine

AR accumulative
regularization

Gradient norm matching function residual

Algorithm AR framework for gradient minimization

Require: $S$, a strictly increasing sequence $\{\sigma_s\}_{s=0}^{S}$ with $\sigma_0 = 0$, and an initial point $x_0 \in \mathbb{R}^n$.

for $s = 1,\dots,S$ do
Compute an approximate solution $x_s$ of the proximal subproblem $$x_s \approx \operatorname*{arg\,min}_{x \in \mathbb{R}^n} \Big\{\, f_s(x) := f(x) + \textstyle\sum_{i=1}^{s} \tfrac{\sigma_i - \sigma_{i-1}}{p+\nu}\,\|x - x_{i-1}\|^{p+\nu} \Big\}$$ where $p + \nu \geq 2$, by running a subroutine $\mathcal{A}$ initialized at $x_{s-1}$ for $N_s$ iterations.
end for
Output: $x_S$

Selected work

High-Order Accumulative Regularization for Gradient Minimization in Convex Programming

Optimal gradient complexity for finding points with small gradients.

Parameter-Free Optimization
Higher-Order Optimization

Yao Ji and Guanghui Lan

arXiv:2511.03723, 2025

R&R 2nd round review, SIAM Journal on Optimization

Learning and Decision-Making

The SCA 2026 conference poster. Sections cover the background and motivation for cardiac surgery-associated acute kidney injury, the dataset of 108,258 adult cardiac surgery admissions, a reinforcement learning framework formulating perioperative management as sequential decision-making, a transformer model architecture, and results on kernel performance, AKI prediction, and complication prediction, closing with discussion and future work.

Selected presentation

Transformer-Based Patient Response Modeling Across All Perioperative Phases for Cardiac Surgery-Associated Acute Kidney Injury Management

Transformer-based modeling of patient response across all perioperative phases, built to support sequential clinical decision-making for cardiac surgery-associated acute kidney injury.

Healthcare

Yao Ji, Yan Li, Guanghui Lan, Zehua Dong, Shuoling Li, Ilker Guven, Xiaoyu Chen, Lihui Bai, and Jiapeng Huang

Society of Cardiovascular Anesthesiologists Annual Meeting — poster, 2026

Poster — Transformer-Based Patient Response Modeling Across All Perioperative Phases for Cardiac Surgery-Associated Acute Kidney Injury Management

Two motivating settings for distributed optimization. Left: a swarm of UAVs linked by local communication over a mesh network. Right: several hospitals and a university jointly analysing gene expression data over local links. Data stay stored and processed locally at each agent, and the agents work together to solve a common problem.

Selected work

Distributed Sparse Regression via Penalization

Statistical and computational guarantees for sparse linear regression over a network of agents, formulated as local LASSO losses plus a quadratic penalty on the consensus constraint.

Distributed Optimization

Yao Ji, Gesualdo Scutari, Ying Sun, and Harsha Honnappa

Journal of Machine Learning Research, 24(272):1–62, 2023

Paper — Distributed Sparse Regression via Penalization

Test mean squared error against the number of communications. ATC-DGD with step size 0.06 drops sharply within a few thousand communications and settles on the centralized mean squared error. CTA-DGD with the same step size stalls at the local mean squared error, and CTA-DGD with much smaller step sizes, 0.0001 and 0.00005, decreases only slowly and is still far above the centralized level after 45,000 communications.

Selected work

Distributed (ATC) Gradient Descent for High Dimension Sparse Regression

First statistical study of adapt-then-combine distributed gradient descent, showing linear convergence to centralized statistical precision with communication growing only logarithmically in the ambient dimension.

Distributed Optimization

Yao Ji, Gesualdo Scutari, Ying Sun, and Harsha Honnappa

IEEE Transactions on Information Theory, 69(8):5253–5276, 2023

Teaching

Teaching Experience

Georgia Institute of Technology, ISyE

Instructor

Fall 2026

Purdue University, Industrial Engineering

Teaching Assistant

IE 335: Operations Research Spring and Fall 2023
IE 330: Probability and Statistics in Engineering Fall 2022
IE 590: Introduction to Optimization Algorithms (graduate-level) Fall 2022

Beijing Normal University, School of Mathematical Sciences

Co-Lecturer

Large Deviation Theory Spring 2018
Brownian Motion Fall 2017
Random Walk in Random Environment Spring 2017
Galton–Watson Branching Process Fall 2016

Teaching Assistant Outstanding Teaching Assistant Award

Measure Theory I Fall 2017 and Fall 2018
Measure Theory II Spring 2017 and Spring 2018
Stochastic Calculus for Finance (graduate-level) Fall 2016

Outreach

Instructor Pre-College Reinforcement Learning Course for High School Students, Georgia Tech Summer 2026
Judge STEM Spark: INFORMS K–12 Poster Competition 2026
Judge Georgia Tech 2026 CRIDC Poster Competition 2026

Yao Ji

Operations Researcher

Recent News

Co-organizing a Session on Stochastic Optimization for RL at INFORMS 2026

Presenting in the INFORMS 2026 Job Market Showcase

Co-organizing a Session on Stochastic Optimization at INFORMS 2026

Attending ColOpt 2026

Talk at MOPTA 2026

Co-organizing Two Sessions at MOPTA 2026

Poster at Southern California OR/OM Day

Seminar at Clemson University

Field Visit to Gwinnett County Wastewater Facility

Poster Presentation at SCA 2026 in Nashville, Tennessee

Chairing a Session at INFORMS Optimization Society Conference 2026

See you at INFORMS 2025

H. Milton Stewart Postdoctoral Fellow at Georgia Tech

Research Interests and Selected Works

Current focus

Foundation

Stochastic Auto-conditioned Fast Gradient Methods with Optimal Rates

Computation of Strong Solutions to Stochastic Variational Inequalities

High-Order Accumulative Regularization for Gradient Minimization in Convex Programming

Learning and Decision-Making

Transformer-Based Patient Response Modeling Across All Perioperative Phases for Cardiac Surgery-Associated Acute Kidney Injury Management

Distributed Sparse Regression via Penalization

Distributed (ATC) Gradient Descent for High Dimension Sparse Regression

Teaching

Teaching Experience

Georgia Institute of Technology, ISyE

Purdue University, Industrial Engineering

Beijing Normal University, School of Mathematical Sciences

Outreach