Selected work

Representative Papers

Bayesian optimization for mixed variables using an adaptive dimension reduction process: applications to aircraft design

Apr 11, 2025

In multidisciplinary design optimization—particularly aircraft design—the presence of high-dimensional mixed variables (continuous, integer, and categorical) causes combinatorial explosion in the hyperparameter space of Bayesian optimization surrogate models. To address this, we propose a Partial Least Squares (PLS)-based adaptive dimensionality reduction framework. It dynamically learns variable coupling structures to significantly compress the surrogate hyperparameter space, while integrating mixed-variable encoding with an adaptive hyperparameter selection mechanism to balance modeling accuracy and tuning efficiency. Evaluated on analytical benchmarks and two real-world aircraft design cases, our method achieves >30% faster convergence and improves optimal solution quality by 12–18% compared to genetic algorithms, while reducing hyperparameter count by ~60%. Our core contribution is the first application of PLS for hyperparameter pruning in Bayesian optimization, enabling efficient and robust optimization under high-dimensional mixed-variable settings.

8 citations3 influentialRead paper

Constrained Multi-Objective Bayesian Optimization with Application to Aircraft Design

Jun 20, 2022AIAA AVIATION 2022 Forum

Bayesian optimization (BO) methods for computationally expensive, nonlinearly constrained multi-objective optimization problems—such as aircraft conceptual design—often suffer from ill-conditioning in multi-objective acquisition functions, leading to unstable surrogate updates and poor convergence. Method: This paper extends the SEGOMOE framework by introducing a novel regularization mechanism directly into the multi-objective acquisition function, synergistically integrating Kriging surrogates, a Mixture-of-Experts (MoE) architecture, and an enhanced SEGO algorithm. Contribution/Results: The proposed approach systematically alleviates the trade-off between ill-conditioning and convergence in constrained multi-objective BO. Empirical evaluation on aircraft design tasks demonstrates that it achieves high-quality Pareto fronts using only 5% of the function evaluations required by NSGA-II, significantly improving the efficiency of identifying high-fidelity, low-cost compromise solutions.

8 citationsRead paper

Development of Crop Yield Estimation Model using Soil and Environmental Parameters

Feb 10, 2021arXiv.org

Tea yield volatility—driven by soil and environmental factors—poses a significant threat to food security. To address this, we propose a pre-harvest prediction model driven by multi-source heterogeneous parameters. Leveraging ten years of monthly data from Pakistan, the model integrates soil pH with climatic variables (temperature, humidity, precipitation), agro-inputs (pesticide application), and socioeconomic factors (labor availability). Through rigorous feature engineering—including identification of key predictors and application of nonlinear transformations—we develop a novel neural network ensemble framework tailored to tea plantation contexts, coupled with multivariate regression for high-accuracy, interpretable yield forecasting. The model achieves an R² of 0.9461 and RMSE of 0.1204, substantially outperforming conventional approaches. These results demonstrate superior fitting capacity and generalizability, establishing a new paradigm for intelligent, data-driven decision-making in the global tea industry.

5 citationsRead paper

A Contextual Online Learning Theory of Brokerage

May 22, 2024arXiv.org

This paper studies the context-aware online bilateral trading problem: a broker must dynamically set transaction prices for privately informed buyers and sellers, leveraging asset- and market-related contextual features, to maximize expected revenue. We establish the first theoretical learning framework for online brokerage with contextual information, distinguishing between two realistic feedback models—full feedback (where both agents’ valuations are revealed) and binary feedback (where only the transaction outcome is observed). Under a bounded-density assumption on valuations, we propose algorithms based on linear contextual modeling and online convex optimization. In the full-feedback setting, our algorithm achieves the optimal regret bound of $O(Ld ln T)$; under binary feedback, it attains $O(sqrt{LdT ln T})$ regret, and we prove a matching lower bound of $Omega(sqrt{LdT})$. Furthermore, we show that the problem becomes statistically unlearnable without the bounded-density condition.

2 citationsRead paper

Precision Where It Matters: A Novel Spike Aware Mixed-Precision Quantization Strategy for LLaMA-based Language Models

Apr 30, 2025

To address the substantial performance degradation caused by quantization in deploying large language models (e.g., LLaMA), this paper proposes an activation-spike-aware mixed-precision quantization method. We first identify that activation spikes in LLaMA architectures are highly concentrated in specific projection layers—a previously unobserved phenomenon—and accordingly design an architecture-customized quantization strategy: spike-prone projection layers are preserved in high-precision formats (FP16/FP8), while other modules undergo low-bit quantization (INT4/INT8). Our approach integrates layer-wise spike localization with per-tensor calibration. Extensive experiments on LLaMA2, LLaMA3, and Mistral demonstrate that the method significantly reduces perplexity and improves zero-shot accuracy. Notably, under 8-bit per-tensor quantization, it outperforms existing state-of-the-art general-purpose quantization methods across multiple benchmarks.

1 citationsRead paper

Recent publications

Latest Papers

Strong aggregation of the Markov chains associated with matching models based on the automorphism group of their compatibility graphs

Sep 15, 2026

研究通过图的自同构群分析匹配模型中的马尔可夫链强聚合性，扩展了贪婪和非贪婪匹配规则下的强聚合条件。

0 citationsRead paper

Stable Transformer-Actor-Critic Model Predictive Control: A Contraction Analysis Approach

Jun 18, 2026

This work addresses the challenge of ensuring closed-loop stability in Transformer-based Actor-Critic model predictive control (MPC) by proposing a novel Transformer-Actor-Critic MPC architecture. For the first time, incremental input-to-state stability (δISS) is integrated with Riemannian contraction theory to rigorously analyze the coupled dynamics of neural networks and physical systems, establishing that Transformers satisfy global δISS. The derived theoretical bounds are then incorporated as regularization terms during training to learn control policies with verifiable robustness. Evaluated on nonlinear 3D drone target-reaching and obstacle-avoidance tasks involving complex nonconvex constraints, the proposed method effectively solves these challenging control problems while demonstrating guaranteed closed-loop stability and robustness.

0 citationsRead paper

Goal-Oriented Lower-Tail Calibration of Gaussian Processes for Bayesian Optimization

May 19, 2026

This work addresses the inaccuracy of lower-tail predictions in Gaussian processes (GPs) under Bayesian optimization, which arises from kernel and hyperparameter choices and undermines acquisition functions such as expected improvement. Focusing on the reliability of lower-tail predictions in noiseless settings, the paper introduces a target-oriented tail calibration framework featuring two novel concepts: spatial occurrence calibration and threshold μ-calibration. This framework establishes theoretical guarantees for predictive reliability in low-threshold regions and yields a post-processing method, termed tcGP, to enhance calibration quality. Empirical evaluations on standard benchmarks demonstrate that tcGP significantly outperforms both standard and globally calibrated GPs, improving lower-tail prediction accuracy, boosting Bayesian optimization performance, and ensuring that calibrated sampling points remain dense across the design space.

0 citationsRead paper

Geometry-Aware Discretization Error of Diffusion Models

May 08, 2026

This work addresses the dominant role of discretization error in reverse-time sampling of diffusion models under a fixed inference budget, a factor overlooked by existing non-asymptotic analyses that are often loose and ignore data structure. By deriving first-order asymptotic expansions for both the weak error of the Euler–Maruyama scheme and the Fréchet discretization error, the paper establishes—for the first time—an explicit connection between this error and intrinsic data geometry, such as the covariance spectrum, as well as the diffusion schedule. Under the exact score assumption, the integration of numerical analysis for stochastic differential equations with asymptotic theory yields a computable, geometry-aware optimization objective in the Gaussian setting. This formulation demonstrates strong predictive performance across diverse image generation and posterior sampling tasks, offering both theoretical grounding and practical tools for geometry-informed schedule design.

0 citationsRead paper

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

Apr 24, 2026

This study addresses the challenge of scaling EEG foundation models across heterogeneous electrode layouts by establishing the first systematic benchmark for channel adaptation that spans diverse model architectures, downstream tasks, and training paradigms. It evaluates four adaptation strategies: Conv1d projection, spherical spline interpolation, source-space decomposition, and Riemannian recentering. The findings reveal that the optimal adaptation method is highly architecture-dependent; while flexible architectures can intrinsically adapt during fine-tuning, they still require external adaptation when the encoder is frozen. Notably, negative transfer is observed between probing and supervised fine-tuning. Experiments demonstrate that CBraMod, with only 5M parameters, outperforms a general-purpose model 31 times larger in parameter count on four out of five datasets, underscoring the efficacy of compact, EEG-specific architectures.

0 citationsRead paper