Yuanyuan Shi

Cheng Deng, Tianhang Zhang, Zhongmou He et al. · meta-ai, mila

Large language models (LLMs) have achieved great success in general domains of natural language processing. In this paper, we bring LLMs to the realm of geoscience with the objective of advancing research and applications in this field. To this end, we present the first-ever LLM in geoscience, K2, alongside a suite of resources developed to further promote LLM research within geoscience. For instance, we have curated the first geoscience instruction tuning dataset, GeoSignal, which aims to align LLM responses to geoscience-related user queries. Additionally, we have established the first geoscience benchmark, GeoBench, to evaluate LLMs in the context of geoscience. In this work, we experiment with a complete recipe to adapt a pre-trained general-domain LLM to the geoscience domain. Specifically, we further train the LLaMA-7B model on 5.5B tokens of geoscience text corpus, including over 1 million pieces of geoscience literature, and utilize GeoSignal's supervised data to fine-tune the model. Moreover, we share a protocol that can efficiently gather domain-specific data and construct domain-supervised data, even in situations where manpower is scarce. Meanwhile, we equip K2 with the abilities of using tools to be a naive geoscience aide. Experiments conducted on the GeoBench demonstrate the effectiveness of our approach and datasets on geoscience knowledge understanding and utilization.We open-source all the training data and K2 model checkpoints at https://github.com/davendw49/k2.

Miroslav Krstic, Luke Bhan, Yuanyuan Shi

Unlike ODEs, whose models involve system matrices and whose controllers involve vector or matrix gains, PDE models involve functions in those roles functional coefficients, dependent on the spatial variables, and gain functions dependent on space as well. The designs of gains for controllers and observers for PDEs, such as PDE backstepping, are mappings of system model functions into gain functions. These infinite dimensional nonlinear operators are given in an implicit form through PDEs, in spatial variables, which need to be solved to determine the gain function for each new functional coefficient of the PDE. The need for solving such PDEs can be eliminated by learning and approximating the said design mapping in the form of a neural operator. Learning the neural operator requires a sufficient number of prior solutions for the design PDEs, offline, as well as the training of the operator. In recent work, we developed the neural operators for PDE backstepping designs for first order hyperbolic PDEs. Here we extend this framework to the more complex class of parabolic PDEs. The key theoretical question is whether the controllers are still stabilizing, and whether the observers are still convergent, if they employ the approximate functional gains generated by the neural operator. We provide affirmative answers to these questions, namely, we prove stability in closed loop under gains produced by neural operators. We illustrate the theoretical results with numerical tests and publish our code on github. The neural operators are three orders of magnitude faster in generating gain functions than PDE solvers for such gain functions. This opens up the opportunity for the use of this neural operator methodology in adaptive control and in gain scheduling control for nonlinear PDEs.

Yuanyuan Shi, Zongyi Li, Huan Yu et al.

State estimation is important for a variety of tasks, from forecasting to substituting for unmeasured states in feedback controllers. Performing real-time state estimation for PDEs using provably and rapidly converging observers, such as those based on PDE backstepping, is computationally expensive and in many cases prohibitive. We propose a framework for accelerating PDE observer computations using learning-based approaches that are much faster while maintaining accuracy. In particular, we employ the recently-developed Fourier Neural Operator (FNO) to learn the functional mapping from the initial observer state and boundary measurements to the state estimate. By employing backstepping observer gains for previously-designed observers with particular convergence rate guarantees, we provide numerical experiments that evaluate the increased computational efficiency gained with FNO. We consider the state estimation for three benchmark PDE examples motivated by applications: first, for a reaction-diffusion (parabolic) PDE whose state is estimated with an exponential rate of convergence; second, for a parabolic PDE with exact prescribed-time estimation; and, third, for a pair of coupled first-order hyperbolic PDEs that modeling traffic flow density and velocity. The ML-accelerated observers trained on simulation data sets for these PDEs achieves up to three orders of magnitude improvement in computational speed compared to classical methods. This demonstrates the attractiveness of the ML-accelerated observers for real-time state estimation and control.

Yuanyuan Shi, Bolun Xu, Yushi Tan et al.

A vital aspect in energy storage planning and operation is to accurately model its operational cost, which mainly comes from the battery cell degradation. Battery degradation can be viewed as a complex material fatigue process that based on stress cycles. Rainflow algorithm is a popular way for cycle identification in material fatigue process, and has been extensively used in battery degradation assessment. However, the rainflow algorithm does not have a closed form, which makes the major difficulty to include it in optimization. In this paper, we prove the rainflow cycle-based cost is convex. Convexity enables the proposed degradation model to be incorporated in different battery optimization problems and guarantees the solution quality. We provide a subgradient algorithm to solve the problem. A case study on PJM regulation market demonstrates the effectiveness of the proposed degradation model in maximizing the battery operating profits as well as extending its lifetime.

Yuanyuan Shi, Bolun Xu, Baosen Zhang et al.

Energy storage in data centers has mainly been used as devices to backup generators during power outages. Recently, there has been a growing interest in using energy storage devices to actively shape power consumption in data centers to reduce their skyrocketing electricity bills. In this paper, we consider using energy storage in data centers for two applications in a joint fashion: reducing peak demand charges and enabling data centers to participate in regulation markets. We develop an optimization framework that captures the cost of electricity, degradation of energy storage devices, as well as the benefit from regulation markets. Under this frame- work, using real data Microsoft data center traces and PJM regulation signals, we show the electricity bill of a data center can be reduced by up to 20%. Furthermore, we demonstrate that the saving from joint optimization can be even larger than the sum of individually optimizing each component. We quantify the particular aspects of data center load profiles that lead to this superlinear gain. Compared to prior works that consider using energy storage devices for each single application alone, our results suggest that energy storage in data centers can have much larger impacts than previously thought possible.

Sahin Lale, Yuanyuan Shi, Guannan Qu et al.

Learning a dynamical system requires stabilizing the unknown dynamics to avoid state blow-ups. However, current reinforcement learning (RL) methods lack stabilization guarantees, which limits their applicability for the control of safety-critical systems. We propose a model-based RL framework with formal stability guarantees, Krasovskii Constrained RL (KCRL), that adopts Krasovskii's family of Lyapunov functions as a stability constraint. The proposed method learns the system dynamics up to a confidence interval using feature representation, e.g. Random Fourier Features. It then solves a constrained policy optimization problem with a stability constraint based on Krasovskii's method using a primal-dual approach to recover a stabilizing policy. We show that KCRL is guaranteed to learn a stabilizing policy in a finite number of interactions with the underlying unknown system. We also derive the sample complexity upper bound for stabilization of unknown nonlinear dynamical systems via the KCRL framework.

Yuanyuan Shi, Bolun Xu, Yushi Tan et al.

We study the optimal control of battery energy storage under a general "pay-for-performance" setup such as providing frequency regulation and renewable integration. In these settings, batteries need to carefully balance the trade-off between following the instruction signals and their degradation costs in real-time. Existing battery control strategies either do not consider the uncertainty of future signals, or cannot accurately account for battery cycle aging mechanism during operation. In this work, we take a different approach to the optimal battery control problem. Instead of attacking the complexity of battery degradation function or the lack of future information one at a time, we address these two challenges together in a joint fashion. In particular, we present an electrochemically accurate and trackable battery degradation model called the rainflow cycle-based model. We prove the degradation cost is convex. Then we propose an online control policy with a simple threshold structure and show it achieve near-optimal performance with respect to an offline controller that has complete future information. We explicitly characterize the optimality gap and show it is independent to the duration of operation. Simulation results with both synthetic and real regulation traces are conducted to illustrate the theoretical results.

Chi Zhang, Yuanyuan Shi, Yize Chen

Recent advancements in reinforcement learning algorithms have opened doors for researchers to operate and optimize building energy management systems autonomously. However, the lack of an easily configurable building dynamical model and energy management task simulation and evaluation platform has arguably slowed the progress in developing advanced and dedicated reinforcement learning (RL) and control algorithms for building operation tasks. Here we propose "BEAR", a physics-principled Building Environment for Control And Reinforcement Learning. The platform allows researchers to benchmark both model-based and model-free controllers using a broad collection of standard building models in Python without co-simulation using external building simulators. In this paper, we discuss the design of this platform and compare it with other existing building simulation frameworks. We demonstrate the compatibility and performance of BEAR with different controllers, including both model predictive control (MPC) and several state-of-the-art RL methods with two case studies.

Ningkun Zheng, Xiaoxiang Liu, Bolun Xu et al.

This paper proposes a novel energy storage price arbitrage algorithm combining supervised learning with dynamic programming. The proposed approach uses a neural network to directly predicts the opportunity cost at different energy storage state-of-charge levels, and then input the predicted opportunity cost into a model-based arbitrage control algorithm for optimal decisions. We generate the historical optimal opportunity value function using price data and a dynamic programming algorithm, then use it as the ground truth and historical price as predictors to train the opportunity value function prediction model. Our method achieves 65% to 90% profit compared to perfect foresight in case studies using different energy storage models and price data from New York State, which significantly outperforms existing model-based and learning-based methods. While guaranteeing high profitability, the algorithm is also light-weighted and can be trained and implemented with minimal computational cost. Our results also show that the learned prediction model has excellent transferability. The prediction model trained using price data from one region also provides good arbitrage results when tested over other regions.

Jie Feng, Wenqi Cui, Jorge Cortés et al.

The increasing integration of renewable energy resources into power grids has led to time-varying system inertia and consequent degradation in frequency dynamics. A promising solution to alleviate performance degradation is using power electronics interfaced energy resources, such as renewable generators and battery energy storage for primary frequency control, by adjusting their power output set-points in response to frequency deviations. However, designing a frequency controller under time-varying inertia is challenging. Specifically, the stability or optimality of controllers designed for time-invariant systems can be compromised once applied to a time-varying system. We model the frequency dynamics under time-varying inertia as a nonlinear switching system, where the frequency dynamics under each mode are described by the nonlinear swing equations and different modes represent different inertia levels. We identify a key controller structure, named Neural Proportional-Integral (Neural-PI) controller, that guarantees exponential input-to-state stability for each mode. To further improve performance, we present an online event-triggered switching algorithm to select the most suitable controller from a set of Neural-PI controllers, each optimized for specific inertia levels. Simulations on the IEEE 39-bus system validate the effectiveness of the proposed online switching control method with stability guarantees and optimized performance for frequency control under time-varying inertia.

Yujia Huang, Ivan Dario Jimenez Rodriguez, Huan Zhang et al.

Forward invariance is a long-studied property in control theory that is used to certify that a dynamical system stays within some pre-specified set of states for all time, and also admits robustness guarantees (e.g., the certificate holds under perturbations). We propose a general framework for training and provably certifying robust forward invariance in Neural ODEs. We apply this framework to provide certified safety in robust continuous control. To our knowledge, this is the first instance of training Neural ODE policies with such non-vacuous certified guarantees. In addition, we explore the generality of our framework by using it to certify adversarial robustness for image classification.

Zhouhan Lin, Cheng Deng, Le Zhou et al.

Large language models (LLMs) have achieved huge success for their general knowledge and ability to solve a wide spectrum of tasks in natural language processing (NLP). Due to their impressive abilities, LLMs have shed light on potential inter-discipline applications to foster scientific discoveries of a specific domain by using artificial intelligence (AI for science, AI4S). In the meantime, utilizing NLP techniques in geoscience research and practice is wide and convoluted, contributing from knowledge extraction and document classification to question answering and knowledge discovery. In this work, we take the initial step to leverage LLM for science, through a rather straightforward approach. We try to specialize an LLM into geoscience, by further pre-training the model with a vast amount of texts in geoscience, as well as supervised fine-tuning (SFT) the resulting model with our custom collected instruction tuning dataset. These efforts result in a model GeoGalactica consisting of 30 billion parameters. To our best knowledge, it is the largest language model for the geoscience domain. More specifically, GeoGalactica is from further pre-training of Galactica. We train GeoGalactica over a geoscience-related text corpus containing 65 billion tokens, preserving as the largest geoscience-specific text corpus. Then we fine-tune the model with 1 million pairs of instruction-tuning data consisting of questions that demand professional geoscience knowledge to answer. In this technical report, we will illustrate in detail all aspects of GeoGalactica, including data collection, data cleaning, base model selection, pre-training, SFT, and evaluation. We open-source our data curation tools and the checkpoints of GeoGalactica during the first 3/4 of pre-training.

Luke Bhan, Yuanyuan Shi, Miroslav Krstic

Neural operator approximations of the gain kernels in PDE backstepping has emerged as a viable method for implementing controllers in real time. With such an approach, one approximates the gain kernel, which maps the plant coefficient into the solution of a PDE, with a neural operator. It is in adaptive control that the benefit of the neural operator is realized, as the kernel PDE solution needs to be computed online, for every updated estimate of the plant coefficient. We extend the neural operator methodology from adaptive control of a hyperbolic PDE to adaptive control of a benchmark parabolic PDE (a reaction-diffusion equation with a spatially-varying and unknown reaction coefficient). We prove global stability and asymptotic regulation of the plant state for a Lyapunov design of parameter adaptation. The key technical challenge of the result is handling the 2D nature of the gain kernels and proving that the target system with two distinct sources of perturbation terms, due to the parameter estimation error and due to the neural approximation error, is Lyapunov stable. To verify our theoretical result, we present simulations achieving calculation speedups up to 45x relative to the traditional finite difference solvers for every timestep in the simulation trajectory.

Maxence Lamarque, Luke Bhan, Yuanyuan Shi et al.

To stabilize PDEs, feedback controllers require gain kernel functions, which are themselves governed by PDEs. Furthermore, these gain-kernel PDEs depend on the PDE plants' functional coefficients. The functional coefficients in PDE plants are often unknown. This requires an adaptive approach to PDE control, i.e., an estimation of the plant coefficients conducted concurrently with control, where a separate PDE for the gain kernel must be solved at each timestep upon the update in the plant coefficient function estimate. Solving a PDE at each timestep is computationally expensive and a barrier to the implementation of real-time adaptive control of PDEs. Recently, results in neural operator (NO) approximations of functional mappings have been introduced into PDE control, for replacing the computation of the gain kernel with a neural network that is trained, once offline, and reused in real-time for rapid solution of the PDEs. In this paper, we present the first result on applying NOs in adaptive PDE control, presented for a benchmark 1-D hyperbolic PDE with recirculation. We establish global stabilization via Lyapunov analysis, in the plant and parameter error states, and also present an alternative approach, via passive identifiers, which avoids the strong assumptions on kernel differentiability. We then present numerical simulations demonstrating stability and observe speedups up to three orders of magnitude, highlighting the real-time efficacy of neural operators in adaptive control. Our code (Github) is made publicly available for future researchers.

Yuan Zhuang, Yuexin Bian, Sihong He et al.

Scaling critic capacity is a promising direction for enhancing off-policy reinforcement learning (RL). However, larger critics are prone to overfitting and unstable in replay-buffer-based bootstrap training. This paper leverages Low-Rank Adaptation (LoRA) as a structural-sparsity regularizer for off-policy critics. Our approach freezes randomly initialized base matrices and solely optimizes low-rank adapters, thereby constraining critic updates to a low-dimensional subspace. Built on top of SimbaV2, we further develop a LoRA formulation, compatible with SimbaV2, that preserves its hyperspherical normalization geometry under frozen-backbone training. We evaluate our method with SAC and FastTD3 on DeepMind Control locomotion and IsaacLab robotics benchmarks. LoRA consistently achieves lower critic loss during training and stronger policy performance. Extensive experiments demonstrate that adaptive low-rank updates provide a simple, scalable, and effective structural regularization for critic learning in off-policy RL.

Luke Bhan, Yuexin Bian, Miroslav Krstic et al.

Over the last decade, data-driven methods have surged in popularity, emerging as valuable tools for control theory. As such, neural network approximations of control feedback laws, system dynamics, and even Lyapunov functions have attracted growing attention. With the ascent of learning based control, the need for accurate, fast, and easy-to-use benchmarks has increased. In this work, we present the first learning-based environment for boundary control of PDEs. In our benchmark, we introduce three foundational PDE problems - a 1D transport PDE, a 1D reaction-diffusion PDE, and a 2D Navier-Stokes PDE - whose solvers are bundled in an user-friendly reinforcement learning gym. With this gym, we then present the first set of model-free, reinforcement learning algorithms for solving this series of benchmark problems, achieving stability, although at a higher cost compared to model-based PDE backstepping. With the set of benchmark environments and detailed examples, this work significantly lowers the barrier to entry for learning-based PDE control - a topic largely unexplored by the data-driven control community. The entire benchmark is available on Github along with detailed documentation and the presented reinforcement learning models are open sourced.

Xiao Liang, Chenxi Liu, Zhi Ma et al.

Medical Vision-Language Models (MedVLMs) show immense promise in clinical applicability. However, their reliability is hindered by hallucinations, where models often fail to derive answers from visual evidence, instead relying on learned textual priors. Existing mitigation strategies for MedVLMs have distinct limitations: training-based methods rely on costly expert annotations, limiting scalability, while training-free interventions like contrastive decoding, though data-efficient, apply a global, untargeted correction whose effects in complex real-world clinical settings can be unreliable. To address these challenges, we introduce Anatomical Region-Guided Contrastive Decoding (ARCD), a plug-and-play strategy that mitigates hallucinations by providing targeted, region-specific guidance. Our module leverages an anatomical mask to direct a three-tiered contrastive decoding process. By dynamically re-weighting at the token, attention, and logits levels, it verifiably steers the model's focus onto specified regions, reinforcing anatomical understanding and suppressing factually incorrect outputs. Extensive experiments across diverse datasets, including chest X-ray, CT, brain MRI, and ocular ultrasound, demonstrate our method's effectiveness in improving regional understanding, reducing hallucinations, and enhancing overall diagnostic accuracy.

Luke Bhan, Miroslav Krstic, Yuanyuan Shi

Due to simplicity and strong stability guarantees, predictor feedback methods have stood as a popular approach for time delay systems since the 1950s. For time-varying delays, however, implementation requires computing a prediction horizon defined by the inverse of the delay function, which is rarely available in closed form and must be approximated. In this work, we formulate the inverse delay mapping as an operator learning problem and study predictor feedback under approximation of the prediction horizon. We propose two approaches: (i) a numerical method based on time integration of an equivalent ODE, and (ii) a data-driven method using neural operators to learn the inverse mapping. We show that both approaches achieve arbitrary approximation accuracy over compact sets, with complementary trade-offs in computational cost and scalability. Building on these approximations, we then develop an output-feedback predictor design for systems with delays in both the input and the measurement. We prove that the resulting closed-loop system is globally exponentially stable when the prediction horizon is approximated with sufficiently small error. Lastly, numerical experiments validate the proposed methods and illustrate their trade-offs between accuracy and computational efficiency.

Luke Bhan, Peter Quawas, Miroslav Krstic et al.

Modern control systems frequently operate under input delays and sampled state measurements. A common delay-compensation strategy is predictor feedback; however, practical implementations require solving an implicit ODE online, resulting in intractable computational cost. Moreover, predictor formulations typically assume continuously available state measurements, whereas in practice measurements may be sampled, irregular, or temporarily missing due to hardware faults. In this work, we develop two neural-operator predictor-feedback designs for nonlinear systems with delayed inputs and sampled measurements. In the first design, we introduce a sampling-horizon prediction operator that maps the current measurement and input history to the predicted state trajectory over the next sampling interval. In the second design, the neural operator approximates only the delay-compensating predictor, which is then composed with the closed-loop flow between measurements. The first approach requires uniform sampling but yields residual bounds that scale directly with the operator approximation error. In contrast, the second accommodates non-uniform, but bounded sampling schedules at the cost of amplified approximation error, revealing a practical tradeoff between sampling flexibility and approximation sensitivity for the control engineer. For both schemes, we establish semi-global practical stability with explicit neural operator error-dependent bounds. Numerical experiments on a 6-link nonlinear robotic manipulator demonstrate accurate tracking and substantial computational speedup of 25$\times$ over a baseline approach.

Yuexin Bian, Jie Feng, Tao Wang et al.

On-policy deep reinforcement learning remains a dominant paradigm for continuous control, yet standard implementations rely on Gaussian actors and relatively shallow MLP policies, often leading to brittle optimization when gradients are noisy and policy updates must be conservative. In this paper, we revisit policy representation as a first-class design choice for on-policy optimization. We study discretized categorical actors that represent each action dimension with a distribution over bins, yielding a policy objective that resembles a cross-entropy loss. Building on architectural advances from supervised learning, we further propose regularized actor networks, while keeping critic design fixed. Our results show that simply replacing the standard actor network with our discretized regularized actor yields consistent gains and achieve the state-of-the-art performance across diverse continuous-control benchmarks.

Yixuan Tang, Yuanyuan Shi, Yiqun Sun et al.

Access to diverse perspectives is essential for understanding real-world events, yet most news retrieval systems prioritize textual relevance, leading to redundant results and limited viewpoint exposure. We propose NEWSCOPE, a two-stage framework for diverse news retrieval that enhances event coverage by explicitly modeling semantic variation at the sentence level. The first stage retrieves topically relevant content using dense retrieval, while the second stage applies sentence-level clustering and diversity-aware re-ranking to surface complementary information. To evaluate retrieval diversity, we introduce three interpretable metrics, namely Average Pairwise Distance, Positive Cluster Coverage, and Information Density Ratio, and construct two paragraph-level benchmarks: LocalNews and DSGlobal. Experiments show that NEWSCOPE consistently outperforms strong baselines, achieving significantly higher diversity without compromising relevance. Our results demonstrate the effectiveness of fine-grained, interpretable modeling in mitigating redundancy and promoting comprehensive event understanding. The data and code are available at https://github.com/tangyixuan/NEWSCOPE.

Jie Feng, Haohan Zou, Yuanyuan Shi

We propose CoNSAL (Combining Neural networks and Symbolic regression for Analytical Lyapunov function) to construct analytical Lyapunov functions for nonlinear dynamic systems. This framework contains a neural Lyapunov function and a symbolic regression component, where symbolic regression is applied to distill the neural network to precise analytical forms. Our approach utilizes symbolic regression not only as a tool for translation but also as a means to uncover counterexamples. This procedure terminates when no counterexamples are found in the analytical formulation. Compared with previous results, CoNSAL directly produces an analytical form of the Lyapunov function with improved interpretability in both the learning process and the final results. We apply CoNSAL to 2-D inverted pendulum, path following, Van Der Pol Oscillator, 3-D trig dynamics, 4-D rotating wheel pendulum, 6-D 3-bus power system, and demonstrate that our algorithm successfully finds their valid Lyapunov functions. Code examples are available at https://github.com/HaohanZou/CoNSAL.

Haohan Zou, Jie Feng, Hao Zhao et al.

Despite advances in learning-based methods, finding valid Lyapunov functions for nonlinear dynamical systems remains challenging. Current neural network approaches face two main issues: challenges in scalable verification and limited interpretability. To address these, we propose an end-to-end framework using transformers to construct analytical Lyapunov functions (local), which simplifies formal verification, enhances interpretability, and provides valuable insights for control engineers. Our framework consists of a transformer-based trainer that generates candidate Lyapunov functions and a falsifier that verifies candidate expressions and refines the model via risk-seeking policy gradient. Unlike Alfarano et al. (2024), which utilizes pre-training and seeks global Lyapunov functions for low-dimensional systems, our model is trained from scratch via reinforcement learning (RL) and succeeds in finding local Lyapunov functions for high-dimensional and non-polynomial systems. Given the analytical nature of the candidates, we employ efficient optimization methods for falsification during training and formal verification tools for the final verification. We demonstrate the efficiency of our approach on a range of nonlinear dynamical systems with up to ten dimensions and show that it can discover Lyapunov functions not previously identified in the control literature. Full implementation is available on \href{https://github.com/JieFeng-cse/Analytical-Lyapunov-Function-Discovery}{Github}

Kevin Huang, Sahin Lale, Ugo Rosolia et al.

Current state-of-the-art model-based reinforcement learning algorithms use trajectory sampling methods, such as the Cross-Entropy Method (CEM), for planning in continuous control settings. These zeroth-order optimizers require sampling a large number of trajectory rollouts to select an optimal action, which scales poorly for large prediction horizons or high dimensional action spaces. First-order methods that use the gradients of the rewards with respect to the actions as an update can mitigate this issue, but suffer from local optima due to the non-convex optimization landscape. To overcome these issues and achieve the best of both worlds, we propose a novel planner, Cross-Entropy Method with Gradient Descent (CEM-GD), that combines first-order methods with CEM. At the beginning of execution, CEM-GD uses CEM to sample a significant amount of trajectory rollouts to explore the optimization landscape and avoid poor local minima. It then uses the top trajectories as initialization for gradient descent and applies gradient updates to each of these trajectories to find the optimal action sequence. At each subsequent time step, however, CEM-GD samples much fewer trajectories from CEM before applying gradient updates. We show that as the dimensionality of the planning problem increases, CEM-GD maintains desirable performance with a constant small number of samples by using the gradient information, while avoiding local optima using initially well-sampled trajectories. Furthermore, CEM-GD achieves better performance than CEM on a variety of continuous control benchmarks in MuJoCo with 100x fewer samples per time step, resulting in around 25% less computation time and 10% less memory usage. The implementation of CEM-GD is available at $\href{https://github.com/KevinHuang8/CEM-GD}{\text{https://github.com/KevinHuang8/CEM-GD}}$.

Sunki Hong, Jisoo Lee, Yuanyuan Shi

Selecting the right deep learning model for power grid forecasting is challenging, as performance heavily depends on the data available to the operator. This paper presents a comprehensive benchmark of five modern neural architectures: two state space models (PowerMamba, S-Mamba), two Transformers (iTransformer, PatchTST), and a traditional LSTM. We evaluate these models on hourly electricity demand across six diverse US power grids for forecast windows between 24 and 168 hours. To ensure a fair comparison, we adapt each model with specialized temporal processing and a modular layer that cleanly integrates weather covariates. Our results reveal that there is no single best model for all situations. When forecasting using only historical load, PatchTST and the state space models provide the highest accuracy. However, when explicit weather data is added to the inputs, the rankings reverse: iTransformer improves its accuracy three times more efficiently than PatchTST. By controlling for model size, we confirm that this advantage stems from the architecture's inherent ability to mix information across different variables. Extending our evaluation to solar generation, wind power, and wholesale prices further demonstrates that model rankings depend on the forecast task: PatchTST excels on highly rhythmic signals like solar, while state space models are better suited for the chaotic fluctuations of wind and price. Ultimately, this benchmark provides grid operators with actionable guidelines for selecting the optimal forecasting architecture based on their specific data environments.

Yihong Guo, Yixuan Wang, Yuanyuan Shi et al.

Training a policy in a source domain for deployment in the target domain under a dynamics shift can be challenging, often resulting in performance degradation. Previous work tackles this challenge by training on the source domain with modified rewards derived by matching distributions between the source and the target optimal trajectories. However, pure modified rewards only ensure the behavior of the learned policy in the source domain resembles trajectories produced by the target optimal policies, which does not guarantee optimal performance when the learned policy is actually deployed to the target domain. In this work, we propose to utilize imitation learning to transfer the policy learned from the reward modification to the target domain so that the new policy can generate the same trajectories in the target domain. Our approach, Domain Adaptation and Reward Augmented Imitation Learning (DARAIL), utilizes the reward modification for domain adaptation and follows the general framework of generative adversarial imitation learning from observation (GAIfO) by applying a reward augmented estimator for the policy optimization step. Theoretically, we present an error bound for our method under a mild assumption regarding the dynamics shift to justify the motivation of our method. Empirically, our method outperforms the pure modified reward method without imitation learning and also outperforms other baselines in benchmark off-dynamics environments.

Luke Bhan, Peijia Qin, Miroslav Krstic et al.

Predictor feedback designs are critical for delay-compensating controllers in nonlinear systems. However, these designs are limited in practical applications as predictors cannot be directly implemented, but require numerical approximation schemes, which become computationally prohibitive when system dynamics are expensive to compute. To address this challenge, we recast the predictor design as an operator learning problem, and learn the predictor mapping via a neural operator. We prove the existence of an arbitrarily accurate neural operator approximation of the predictor operator. Under the approximated predictor, we achieve semiglobal practical stability of the closed-loop nonlinear delay system. The estimate is semiglobal in a unique sense - one can enlarge the set of initial states as desired, though this increases the difficulty of training a neural operator, which appears practically in the stability estimate. Furthermore, our analysis holds for any black-box predictor satisfying the universal approximation error bound. We demonstrate the approach by controlling a 5-link robotic manipulator with different neural operator models, achieving significant speedups compared to classic predictor feedback schemes while maintaining closed-loop stability.

Yufan Zhang, Honglin Wen, Yuexin Bian et al.

Large penetration of renewable energy sources (RESs) brings huge uncertainty into the electricity markets. The current deterministic clearing approach in the day-ahead (DA) market, where RESs participate based on expected production, has been criticized for causing a lack of coordination between the DA and real-time (RT) markets, leading to high overall operating costs. Previous works indicate that improving day-ahead RES entering quantities can significantly mitigate the drawbacks of deterministic clearing. In this work, we propose using a trained forecasting model, referred to as value-oriented forecasting, to determine RES Improved Entering Quantities (RIEQ) more efficiently during the operational phase. Unlike traditional models that minimize statistical forecasting errors, our approach trains model parameters to minimize the expected overall operating costs across both DA and RT markets. We derive the exact form of the loss function used for training, which becomes piecewise linear when market clearing is modeled by linear programs. Additionally, we provide the analytical gradient of the loss function with respect to the forecast, enabling an efficient training strategy. Numerical studies demonstrate that our forecasts significantly reduce overall operating costs for deterministic market clearing compared to conventional forecasts based on expected RES production.

Luke Bhan, Miroslav Krstic, Yuanyuan Shi

In this work, we propose a rigorous method for implementing predictor feedback controllers in nonlinear systems with unknown and arbitrarily long actuator delays. To address the analytically intractable nature of the predictor, we approximate it using a learned neural operator mapping. This mapping is trained once, offline, and then deployed online, leveraging the fast inference capabilities of neural networks. We provide a theoretical stability analysis based on the universal approximation theorem of neural operators and the transport partial differential equation (PDE) representation of the delay. We then prove, via a Lyapunov-Krasovskii functional, semi-global practical convergence of the dynamical system dependent on the approximation error of the predictor and delay bounds. Finally, we validate our theoretical results using a biological activator/repressor system, demonstrating speedups of 15 times compared to traditional numerical methods.

Luke Bhan, Yuanyuan Shi, Iasson Karafyllis et al.

Observers for PDEs are themselves PDEs. Therefore, producing real time estimates with such observers is computationally burdensome. For both finite-dimensional and ODE systems, moving-horizon estimators (MHE) are operators whose output is the state estimate, while their inputs are the initial state estimate at the beginning of the horizon as well as the measured output and input signals over the moving time horizon. In this paper we introduce MHEs for PDEs which remove the need for a numerical solution of an observer PDE in real time. We accomplish this using the PDE backstepping method which, for certain classes of both hyperbolic and parabolic PDEs, produces moving-horizon state estimates explicitly. Precisely, to explicitly produce the state estimates, we employ a backstepping transformation of a hard-to-solve observer PDE into a target observer PDE, which is explicitly solvable. The MHEs we propose are not new observer designs but simply the explicit MHE realizations, over a moving horizon of arbitrary length, of the existing backstepping observers. Our PDE MHEs lack the optimality of the MHEs that arose as duals of MPC, but they are given explicitly, even for PDEs. In the paper we provide explicit formulae for MHEs for both hyperbolic and parabolic PDEs, as well as simulation results that illustrate theoretically guaranteed convergence of the MHEs.

Jianghao Lin, Yuanyuan Shi, Xin Peng et al.

Large language models (LLMs) are increasingly demonstrating strong capabilities as autonomous agents, with function calling serving as a core mechanism for interaction with the environment. Meanwhile, inference scaling has become a cutting-edge technique to enhance LLM performance by allocating more computational resources during the inference process. However, current research on inference scaling primarily focuses on unstructured output generation tasks, leaving its application in structured outputs, like function calling, largely underexplored. To bridge this gap, we propose an inference scaling framework that combines fine-grained beam search with a process reward model, ToolPRM, which scores the internal steps of each single function call. To train ToolPRM, we construct the first fine-grained intra-call process supervision dataset, automatically annotated with function-masking techniques to provide step-level rewards for structured tool-use reasoning. Extensive experiments demonstrate that ToolPRM beats the coarse-grained and outcome reward models in terms of predictive accuracy, indicating its stronger capability in supervising the function calling inference process. Inference scaling technique equipped with ToolPRM also significantly improves the backbone model performance across various function calling tasks and benchmarks. More importantly, we reveal a key principle for applying inference scaling techniques to structured outputs: "explore more but retain less" due to the unrecoverability characteristics of structured function calling generation.

Luke Bhan, Miroslav Krstic, Yuanyuan Shi

This work establishes the first rigorous stability guarantees for approximate predictors in delay-adaptive control of nonlinear systems, addressing a key challenge in practical implementations where exact predictors are unavailable. We analyze two scenarios: (i) when the actuated input is directly measurable, and (ii) when it is estimated online. For the measurable input case, we prove semi-global practical asymptotic stability with an explicit bound proportional to the approximation error $ε$. For the unmeasured input case, we demonstrate local practical asymptotic stability, with the region of attraction explicitly dependent on both the initial delay estimate and the predictor approximation error. To bridge theory and practice, we show that neural operators-a flexible class of neural network-based approximators-can achieve arbitrarily small approximation errors, thus satisfying the conditions of our stability theorems. Numerical experiments on two nonlinear benchmark systems-a biological protein activator/repressor model and a micro-organism growth Chemostat model-validate our theoretical results. In particular, our numerical simulations confirm stability under approximate predictors, highlight the strong generalization capabilities of neural operators, and demonstrate a substantial computational speedup of up to 15x compared to a baseline fixed-point method.

Yuan Zhuang, Yi Shen, Yuexin Bian et al.

Recent studies have shown that combining parameter-efficient fine-tuning (PEFT) with mixture-of-experts (MoE) is an effective strategy for adapting large language models (LLMs) to the downstream tasks. However, most existing approaches rely on conventional TopK routing, which requires careful hyperparameter tuning and assigns a fixed number of experts to each token. In this work, we propose LD-MoLE, a Learnable Dynamic routing mechanism for Mixture of LoRA Experts that enables adaptive, token-dependent, and layer-wise expert allocation. Our method replaces the non-differentiable TopK selection with a differentiable routing function and a closed-form solution. Moreover, our design allows the model to adaptively determine the number of experts to activate for each token at different layers. In addition, we introduce an analytical sparsity control objective to regularize the number of activated experts. Extensive experiments on the Qwen3-1.7B and Llama-3.2-3B models show that LD-MoLE achieves the highest average scores compared to state-of-the-art baselines, across a diverse set of benchmarks. Our method not only achieves superior performance, but also demonstrates the ability to learn token-dependent and layer-wise expert allocation.

Filip Bajraktari, Luke Bhan, Miroslav Krstic et al.

In this work, we present the first stability results for approximate predictors in multi-input non-linear systems with distinct actuation delays. We show that if the predictor approximation satisfies a uniform (in time) error bound, semi-global practical stability is correspondingly achieved. For such approximators, the required uniform error bound depends on the desired region of attraction and the number of control inputs in the system. The result is achieved through transforming the delay into a transport PDE and conducting analysis on the coupled ODE-PDE cascade. To highlight the viability of such error bounds, we demonstrate our results on a class of approximators - neural operators - showcasing sufficiency for satisfying such a universal bound both theoretically and in simulation on a mobile robot experiment.

Yuexin Bian, Oliver Schmidt, Yuanyuan Shi

Energy-efficient ventilation control plays a vital role in reducing building energy consumption while ensuring occupant health and comfort. While Computational Fluid Dynamics (CFD) simulations provide detailed and physically accurate representation of indoor airflow, their high computational cost limits their use in real-time building control. In this work, we present a neural operator learning framework that combines the physical accuracy of CFD with the computational efficiency of machine learning to enable building ventilation control with the high-fidelity fluid dynamics models. Our method jointly optimizes the airflow supply rates and vent angles to reduce energy use and adhere to air quality constraints. We train an ensemble of neural operator transformer models to learn the mapping from building control actions to airflow fields using high-resolution CFD data. This learned neural operator is then embedded in an optimization-based control framework for building ventilation control. Experimental results show that our approach achieves significant energy savings compared to maximum airflow rate control, rule-based control, as well as data-driven control methods using spatially averaged CO2 prediction and deep learning based reduced order model, while consistently maintaining safe indoor air quality. These results highlight the practicality and scalability of our method in maintaining energy efficiency and indoor air quality in real-world buildings.

Jeffrey Ma, Alistair Letcher, Florian Schäfer et al.

Many economic games and machine learning approaches can be cast as competitive optimization problems where multiple agents are minimizing their respective objective function, which depends on all agents' actions. While gradient descent is a reliable basic workhorse for single-agent optimization, it often leads to oscillation in competitive optimization. In this work we propose polymatrix competitive gradient descent (PCGD) as a method for solving general sum competitive optimization involving arbitrary numbers of agents. The updates of our method are obtained as the Nash equilibria of a local polymatrix approximation with a quadratic regularization, and can be computed efficiently by solving a linear system of equations. We prove local convergence of PCGD to stable fixed points for $n$-player general-sum games, and show that it does not require adapting the step size to the strength of the player-interactions. We use PCGD to optimize policies in multi-agent reinforcement learning and demonstrate its advantages in Snake, Markov soccer and an electricity market game. Agents trained by PCGD outperform agents trained with simultaneous gradient descent, symplectic gradient adjustment, and extragradient in Snake and Markov soccer games and on the electricity market game, PCGD trains faster than both simultaneous gradient descent and the extragradient method.

Yujia Huang, Huan Zhang, Yuanyuan Shi et al.

Certified robustness is a desirable property for deep neural networks in safety-critical applications, and popular training algorithms can certify robustness of a neural network by computing a global bound on its Lipschitz constant. However, such a bound is often loose: it tends to over-regularize the neural network and degrade its natural accuracy. A tighter Lipschitz bound may provide a better tradeoff between natural and certified accuracy, but is generally hard to compute exactly due to non-convexity of the network. In this work, we propose an efficient and trainable \emph{local} Lipschitz upper bound by considering the interactions between activation functions (e.g. ReLU) and weight matrices. Specifically, when computing the induced norm of a weight matrix, we eliminate the corresponding rows and columns where the activation function is guaranteed to be a constant in the neighborhood of each given data point, which provides a provably tighter bound than the global Lipschitz constant of the neural network. Our method can be used as a plug-in module to tighten the Lipschitz bound in many certifiable training algorithms. Furthermore, we propose to clip activation functions (e.g., ReLU and MaxMin) with a learnable upper threshold and a sparsity loss to assist the network to achieve an even tighter local Lipschitz bound. Experimentally, we show that our method consistently outperforms state-of-the-art methods in both clean and certified accuracy on MNIST, CIFAR-10 and TinyImageNet datasets with various network architectures.

Alexander Pan, Yongkyun Lee, Huan Zhang et al.

Due to the proliferation of renewable energy and its intrinsic intermittency and stochasticity, current power systems face severe operational challenges. Data-driven decision-making algorithms from reinforcement learning (RL) offer a solution towards efficiently operating a clean energy system. Although RL algorithms achieve promising performance compared to model-based control models, there has been limited investigation of RL robustness in safety-critical physical systems. In this work, we first show that several competition-winning, state-of-the-art RL agents proposed for power system control are vulnerable to adversarial attacks. Specifically, we use an adversary Markov Decision Process to learn an attack policy, and demonstrate the potency of our attack by successfully attacking multiple winning agents from the Learning To Run a Power Network (L2RPN) challenge, under both white-box and black-box attack settings. We then propose to use adversarial training to increase the robustness of RL agent against attacks and avoid infeasible operational decisions. To the best of our knowledge, our work is the first to highlight the fragility of grid control RL algorithms, and contribute an effective defense scheme towards improving their robustness and security.

Yuanyuan Shi, Bolun Xu

This paper proposes a novel end-to-end deep learning framework that simultaneously identifies demand baselines and the incentive-based agent demand response model, from the net demand measurements and incentive signals. This learning framework is modularized as two modules: 1) the decision making process of a demand response participant is represented as a differentiable optimization layer, which takes the incentive signal as input and predicts user's response; 2) the baseline demand forecast is represented as a standard neural network model, which takes relevant features and predicts user's baseline demand. These two intermediate predictions are integrated, to form the net demand forecast. We then propose a gradient-descent approach that backpropagates the net demand forecast errors to update the weights of the agent model and the weights of baseline demand forecast, jointly. We demonstrate the effectiveness of our approach through computation experiments with synthetic demand response traces and a large-scale real world demand response dataset. Our results show that the approach accurately identifies the demand response model, even without any prior knowledge about the baseline demand.

Guannan Qu, Yuanyuan Shi, Sahin Lale et al.

Linear time-varying (LTV) systems are widely used for modeling real-world dynamical systems due to their generality and simplicity. Providing stability guarantees for LTV systems is one of the central problems in control theory. However, existing approaches that guarantee stability typically lead to significantly sub-optimal cumulative control cost in online settings where only current or short-term system information is available. In this work, we propose an efficient online control algorithm, COvariance Constrained Online Linear Quadratic (COCO-LQ) control, that guarantees input-to-state stability for a large class of LTV systems while also minimizing the control cost. The proposed method incorporates a state covariance constraint into the semi-definite programming (SDP) formulation of the LQ optimal controller. We empirically demonstrate the performance of COCO-LQ in both synthetic experiments and a power system frequency control example.

Yuanyuan Shi, Baosen Zhang

In this work, we study the interaction of strategic agents in continuous action Cournot games with limited information feedback. Cournot game is the essential market model for many socio-economic systems where agents learn and compete without the full knowledge of the system or each other. We consider the dynamics of the policy gradient algorithm, which is a widely adopted continuous control reinforcement learning algorithm, in concave Cournot games. We prove the convergence of policy gradient dynamics to the Nash equilibrium when the price function is linear or the number of agents is two. This is the first result (to the best of our knowledge) on the convergence property of learning algorithms with continuous action spaces that do not fall in the no-regret class.

Liyuan Zheng, Yuanyuan Shi, Lillian J. Ratliff et al.

This paper focuses on finding reinforcement learning policies for control systems with hard state and action constraints. Despite its success in many domains, reinforcement learning is challenging to apply to problems with hard constraints, especially if both the state variables and actions are constrained. Previous works seeking to ensure constraint satisfaction, or safety, have focused on adding a projection step to a learned policy. Yet, this approach requires solving an optimization problem at every policy execution step, which can lead to significant computational costs. To tackle this problem, this paper proposes a new approach, termed Vertex Networks (VNs), with guarantees on safety during exploration and on learned control policies by incorporating the safety constraints into the policy network architecture. Leveraging the geometric property that all points within a convex set can be represented as the convex combination of its vertices, the proposed algorithm first learns the convex combination weights and then uses these weights along with the pre-calculated vertices to output an action. The output action is guaranteed to be safe by construction. Numerical examples illustrate that the proposed VN algorithm outperforms vanilla reinforcement learning in a variety of benchmark control tasks.

Daniel J. Mankowitz, Nir Levine, Rae Jeong et al.

We provide a framework for incorporating robustness -- to perturbations in the transition dynamics which we refer to as model misspecification -- into continuous control Reinforcement Learning (RL) algorithms. We specifically focus on incorporating robustness into a state-of-the-art continuous control RL algorithm called Maximum a-posteriori Policy Optimization (MPO). We achieve this by learning a policy that optimizes for a worst case expected return objective and derive a corresponding robust entropy-regularized Bellman contraction operator. In addition, we introduce a less conservative, soft-robust, entropy-regularized objective with a corresponding Bellman operator. We show that both, robust and soft-robust policies, outperform their non-robust counterparts in nine Mujoco domains with environment perturbations. In addition, we show improved robust performance on a high-dimensional, simulated, dexterous robotic hand. Finally, we present multiple investigative experiments that provide a deeper insight into the robustness framework. This includes an adaptation to another continuous control RL algorithm as well as learning the uncertainty set from offline data. Performance videos can be found online at https://sites.google.com/view/robust-rl.

Yuanyuan Shi, Bolun Xu, Di Wang et al.

We consider using a battery storage system simultaneously for peak shaving and frequency regulation through a joint optimization framework which captures battery degradation, operational constraints and uncertainties in customer load and regulation signals. Under this framework, using real data we show the electricity bill of users can be reduced by up to 15\%. Furthermore, we demonstrate that the saving from joint optimization is often larger than the sum of the optimal savings when the battery is used for the two individual applications. A simple threshold real-time algorithm is proposed and achieves this super-linear gain. Compared to prior works that focused on using battery storage systems for single applications, our results suggest that batteries can achieve much larger economic benefits than previously thought if they jointly provide multiple services.