Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “distributed reinforcement learning”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 181 records · Page 10

Performance-Aligned LLMs for Generating Fast HPC Code

Optimizing scientific software is a difficult task because codebases are often large and complex, and performance can depend upon several factors including the algorithm, its implementation, and hardware among others. Causes of poor performance can originate from disparate sources and be difficult to diagnose. Recent years have seen a multitude of work that use large language models (LLMs) to assist in software development tasks. However, these tools are trained to model the distribution of code as text, and are not specifically designed to understand performance aspects of code. In this work, we introduce a reinforcement learning based methodology to align the outputs of code LLMs with performance. This allows us to build upon the current code modeling capabilities of LLMs and extend them to generate better performing code. Here, we demonstrate that our fine-tuned model improves the expected speedup of generated code over base models for a set of benchmark tasks from 0.9 to 1.6 for serial code and 1.9 to 4.5 for OpenMP parallel code.

Computer science↗

Parameter estimation of large flexible aerospace structures with application to the control of the Maypole Deployable Reflector

Systems such as the Maypole deployable reflector have a distributed parameter nature. The flexible column and hoop structure and the circular antenna of 30-100 meter diameter which it supports are described by partial, rather than ordinary, differential equations. Progress completed in reduced order modelling andd controller design and digital parameter estimation and control is summarized. Topics covered include depolyment and on-orbit operation; quasi-static (steady state) operation; dynamic distributed parameter system; autoregressive moving average identification; frequency domain procedures; direct or implicit active control; adaptive observers; parameter estimation using a linear reinforcement learning factor; feedback control; and reduced order modeling for nonlinear systems.

Balas, M. J.↗

Approximating Nash Equilibrium in Day-ahead Electricity Market Bidding with Multi-agent Deep Reinforcement Learning

In this paper, a day-ahead electricity market bidding problem with multiple strategic generation company (GEN-CO) bidders is studied. The problem is formulated as a Markov game model, where GENCO bidders interact with each other todevelop their optimal day-ahead bidding strategies. Considering unobservable information in the problem, a model-free and data-driven approach, known as multi-agent deep deterministic policy gradient (MADDPG), is applied for approximating the Nash equilibrium (NE) in the above Markov game. The MADDPG algorithm has the advantage of generalization due to the automatic feature extraction ability of the deep neural networks. The algorithm is tested on an IEEE 30-bus system with three competitive GENCO bidders in both an uncongested caseand a congested case. Comparisons with a truthful bidding strategy and state-of-the-art deep reinforcement learning methods including deep Q network and deep deterministic policy gradient (DDPG) demonstrate that the applied MADDPG algorithm can find a superior bidding strategy for all the market participants with increased profit gains. In addition, the comparison with a conventional model-based method shows that the MADDPG algorithm has higher computational efficiency, which is feasible for real-world applications.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Reinforcement Learning for Intentional Islanding in Resilient Power Transmission Systems

Intentional islanding is the process of identifying and deliberately decomposing the transmission network to form self-sustained islands from an endangered network during disruptions to improve resilience and security. Most existing intentional islanding models are offline resilience decision tools and hence do not provide outage responses in a timely manner. In this paper, a reinforcement learning (RL) based model for intentional islanding is developed, which offers real-time switching control, online deployability, and adaptability to varying system conditions. The intentional islanding process is formulated as a Markov decision process, where the optimal transmission switching policy is learned using the RL approach. The control policy is learned over an environment that encompasses a Power System Simulator for Engineering (PSS/E) model of the transmission network, facilitated by an interface to the standard openAI Gym framework. The proposed RL-based methodology aims to form stable and self-sustainable islands by ensuring voltage stability while reducing the power mismatch in the formed islands. A proximal policy optimization algorithm is designed, which is suitable for controlling the on/off status of the switches with multi-layer perceptron as value and actor networks. The effectiveness of the proposed framework in the self-recovery of the grid by island formation is applied on the modified IEEE 39-bus test network and validated by dynamic simulations.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Efficient distributed continual learning for steering experiments in real-time

Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), incremental training suffers from catastrophic forgetting (i.e., new patterns are reinforced at the expense of previously acquired knowledge). Training from scratch each time new training data becomes available would result in extremely long training times and massive data accumulation. Rehearsal-based continual learning has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs to achieve high accuracy, short runtime, and scalability. It leverages a set of buffers (local to each GPU) and uses several asynchronous techniques for updating these local buffers in an embarrassingly parallel fashion, all while handling the communication overheads necessary to augment input minibatches using unbiased, global sampling. We further propose a generalization of rehearsal buffers to support both classification and generative learning tasks, as well as more advanced rehearsal strategies (notably Dark Experience Replay, leveraging knowledge distillation). We illustrate this approach with a real-life HPC streaming application from the domain of ptychographic image reconstruction. Furthermore, we run extensive experiments on up to 128 GPUs of the ThetaGPU supercomputer to compare our approach with baselines representative of training-from-scratch (the upper bound in terms of accuracy) and incremental training (the lower bound). Results show that rehearsal-based continual learning achieves a top-5 validation accuracy close to the upper bound, while simultaneously exhibiting a runtime close to the lower bound.

Asynchronous data management↗

Machine Learning Augmented Predictive and Generative Model for Rupture Life in Ferritic and Austenitic Steels

The Larson-Miller parameter (LMP) offers an efficient and fast scheme to estimate the creep rupture life of alloy materials for high temperature applications. However, owing to poor generalizability and dependence on the constant C, which is typically not known a-priori, estimations using the Larson-Miller parameter often result in suboptimal performance for a wide range of materials. At best it is useful in comparing alloys of similar composition. In this work, three machine learning (ML) schemes were developed for rupture life prediction for 9-12% Cr ferritic-martensitic steels and austenitic stainless steels, i.e., a hierarchical model to parameterize LMP using the LMP constant C to compute rupture life, a hierarchical model to parameterize both C and LMP to compute rupture life, and a direct prediction of rupture life. Specifically, we show that the third scheme, using a gradient boosting algorithm, can be used to train ML models for very accurate prediction of rupture life in a variety of alloys (Pear-son Correlation Coefficient > 0.9 for 9-12% Cr and > 0.8 for austenitic stainless steels). In addition, the Shapley value was used to quantify feature importance, making the model interpretable by identifying the effect of various features on the model performance. Furthermore, a variational autoencoder-based generative model was built by conditioning on the experimental dataset to sample hypothetical synthetic candidate alloys from the learnt joint distribution not existing in both 9-12 % Cr ferritic-martensitic steel and austenitic stainless steel datasets. Finally, based on the predictive and generative model, a reinforcement learning strategy has been proposed to guide experimentalists into designing better heat resistant alloys.

Mamun, Md Osman G.↗

Adaptive Load Shedding for Grid Emergency Control via Deep Reinforcement Learning

Emergency control, typically such as under-voltage load shedding (UVLS), is broadly used to grapple with low voltage and voltage instability issues in real-world power systems under contingencies. However, existing emergency control schemes are rule-based and cannot be adaptively applied to uncertain and floating operating conditions. Here, we propose an adaptive UVLS algorithm for emergency control via deep reinforcement learning (DRL) and expert systems. We first construct dynamic components for picturing the power system operation as the environment. The transient voltage recovery criteria, which poses time-varying requirements to UVLS, is integrated into the states and reward function to advise the learning of deep neural networks. The proposed method has no tuning issue of coefficients in reward functions, and this issue was regarded as a deficiency in the existing DRL-based algorithms. Case studies illustrate that the proposed method outperforms the traditional UVLS relay in both the timeliness and efficacy for emergency control.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Q-Learning-Based Impact Assessment of Propagating Extreme Weather on Distribution Grids

The increasing number of power outage events due to extreme weather conditions is hampering us socioeconomically. Preparing in advance for the extreme weather event is critical. It can help utility operators to reduce grid damages, restore grid service quickly, allocate energy resources and repair crews strategically, and hence dramatically increase grid resilience. In this paper, we propose a method to identify the sequence of worst impact zones in the power grid caused by extreme weather events based on Q-learning (a reinforcement learning algorithm). To quantity the weather severity and its effect on the grid, we model the impact of extreme weather on the grid as a function of intensity, vulnerability, and exposure. A modified IEEE 123-node distribution feeder is presented in a mesh grid and experimented for sequences of zones identification. Finally, simulation results present the identified sequences and their associated impacts on the grid caused by extreme weather events.

distribution system↗

Q-Learning Based Impact Assessment of Propagating Extreme Weather on Distribution Grids: Preprint

Increasing number of power outage events due to extreme weather condition is hampering us socioeconomically. Preparing in advance for the extreme weather event is critical and can help utility operators to reduce grid damages, restore grid service quickly, allocate energy resources and repair crews strategically, and hence dramatically increase grid resilience. In this paper, we propose a method to identify the sequence of worst impact zones in the power grid caused by extreme weather events based on Q-learning (a reinforcement learning algorithm). To quantify weather severity and it’s effect on the grid, we model the impact of extreme weather on the grid as a function of intensity, vulnerability and exposure. A modified IEEE 123-node distribution feeder is presented in a mesh grid and experimented for sequences of zones identification. Finally, simulation results present the identified sequences and their associated impacts on the grid caused by the extreme weather events.

distribution system↗

RLMolLM: Reinforcement Learning-Enhanced Language Model Framework for Inverse Molecular Design

Inverse molecular design faces significant challenges due to vast chemical space and complex property requirements. While language models show promise for molecular generation, they struggle with validity, multi-property optimization, and structural constraints. This work presents RLMolLM, a reinforcement learning framework combining Proximal Policy Optimization (PPO) with genetic algorithms to address these limitations. Our approach optimizes multiple user-specified properties including quantitative estimates of drug-likeness (QED), synthetic accessibility (SA), and ADMET (absorption, distribution, metabolism, excretion, and toxicity) endpoints without requiring complete model retraining, while maintaining capability for scaffold-constrained generation where specific substructures must be preserved. We outperform state-of-the-art methods for molecular optimization, achieving best QED scores across GDB13, Moses, and Zinc datasets with up to 31% improvement over previous methods while maintaining excellent validity, uniqueness, and novelty metrics. For simultaneous multi-property optimization, our framework achieves substantial improvements in ADMET properties including 4.5-fold reduction in hERG toxicity and enhanced Caco-2 permeability compared to Moses dataset. Under structural constraints, the framework significantly improves molecular validity while preserving scaffolds and effectively optimizing properties. In conclusion, this versatile solution advances pharmaceutical and materials molecular design through effective integration of reinforcement learning and genetic algorithms with multi-property optimization and scaffold preservation.

Genetic algorithms↗

Accelerator Real-time Edge AI for Distributed Systems (READS) (Proposal)

Over the last decade, Machine Learning (ML) technologies have slowly made their way into the accelerator community. Rapid advances in recent years in deep learning, particularly reinforcement learning for control system applications and the accessibility of deep learning in embedded hardware, have generated renewed interest and spawned a number of applications. The Fermilab Accelerator Complex, shown in Fig. 1, has provided High Energy Physics (HEP) experiments with proton beams for nearly fifty years. The current focus of the laboratory is its world-class experimental program at the intensity frontier. While increasing beam intensity certainly presents its own challenges, preserving beam size while minimizing beam losses – particles lost through interactions with the beam vacuum pipe – turns out to be, in many ways, the main challenge. The accelerator is controlled via a complex system of hundreds of thousands of devices. Enabling fine tuning and real-time optimization of their parameters using ML methods and stepping beyond experience-based reasoning of human operators are key to the success of future intensity upgrades. Our objective will be to integrate ML into accelerator operations and furthermore, provide an accessible framework, which can also be used by a broad range of other accelerator systems with dynamic tuning needs.

43 PARTICLE ACCELERATORS↗

ARM-IRL: Adaptive Resilience Metric Quantification Using Inverse Reinforcement Learning

The resilience of safety-critical systems is gaining importance due to the rise in cyber and physical threats, especially within critical infrastructure. Traditional static resilience metrics may not capture dynamic system states, leading to inaccurate assessments and ineffective responses to cyber threats. This work aims to develop a data-driven, adaptive method for resilience metric learning. We propose a data-driven approach using inverse reinforcement learning (IRL) to learn a single, adaptive resilience metric. The method infers a reward function from expert control actions. Unlike previous approaches using static weights or fuzzy logic, this work applies adversarial inverse reinforcement learning (AIRL), training a generator and discriminator in parallel to learn the reward structure and derive an optimal policy. The proposed approach is evaluated on multiple scenarios: optimal communication network rerouting, power distribution network reconfiguration, and cyber–physical restoration of critical loads using the IEEE 123-bus system. The adaptive, learned resilience metric enables faster critical load restoration in comparison to conventional RL approaches.

97 MATHEMATICS AND COMPUTING↗

A synthetic building operation dataset

Abstract This paper presents a synthetic building operation dataset which includes HVAC, lighting, miscellaneous electric loads (MELs) system operating conditions, occupant counts, environmental parameters, end-use and whole-building energy consumptions at 10-minute intervals. The data is created with 1395 annual simulations using the U.S. DOE detailed medium-sized reference office building, and 30 years’ historical weather data in three typical climates including Miami, San Francisco, and Chicago. Three energy efficiency levels of the building and systems are considered. Assumptions regarding occupant movements, occupants’ diverse temperature preferences, lighting, and MELs are adopted to reflect realistic building operations. A semantic building metadata schema - BRICK, is used to store the building metadata. The dataset is saved in a 1.2 TB of compressed HDF5 file. This dataset can be used in various applications, including building energy and load shape benchmarking, energy model calibration, evaluation of occupant and weather variability and their influences on building performance, algorithm development and testing for thermal and energy load prediction, model predictive control, policy development for reinforcement learning based building controls.

24 POWER TRANSMISSION AND DISTRIBUTION↗

The Heuristic Dynamic Programming Approach in Boost Converters

In this study, a heuristic dynamic programming controller is proposed to control a boost converter. Conventional controllers such as proportional-integral-derivative (PID) or proportional-integral (PI) are designed based on the linearized small-signal model near the operating point. Therefore, the performance of the controller during the start-up, the load change, or the input voltage variation is not optimal since the system model changes by varying the operating point. The heuristic dynamic programming controller optimally controls the boost converter by following the approximate dynamic programming. The advantage of the HDP is that the neural network-based characteristic of the proposed controller enables boost converters to easily cope with large disturbances. An HDP with a well-trained critic and action networks can perform as an optimal controller for the boost converter. To compare the effectiveness of the traditional PI-based and the HDP boost converter, the simulation results are provided.

24 POWER TRANSMISSION AND DISTRIBUTION↗

Product Distribution Theory for Control of Multi-Agent Systems

Product Distribution (PD) theory is a new framework for controlling Multi-Agent Systems (MAS's). First we review one motivation of PD theory, as the information-theoretic extension of conventional full-rationality game theory to the case of bounded rational agents. In this extension the equilibrium of the game is the optimizer of a Lagrangian of the (probability distribution of) the joint stare of the agents. Accordingly we can consider a team game in which the shared utility is a performance measure of the behavior of the MAS. For such a scenario the game is at equilibrium - the Lagrangian is optimized - when the joint distribution of the agents optimizes the system's expected performance. One common way to find that equilibrium is to have each agent run a reinforcement learning algorithm. Here we investigate the alternative of exploiting PD theory to run gradient descent on the Lagrangian. We present computer experiments validating some of the predictions of PD theory for how best to do that gradient descent. We also demonstrate how PD theory can improve performance even when we are not allowed to rerun the MAS from different initial conditions, a requirement implicit in some previous work.

Lee, Chia Fan↗

Adaptive critic design-based reinforcement learning approach in controlling virtual inertia-based grid-connected inverters

In this report, an adaptive critic design (ACD) approach is proposed to control the phase and voltage of a grid-connected virtual synchronous generator (VSG). The penetration of fast responding inertia-less power converters significantly affect the stability of the power system, especially weak systems such as micro grids. The concept of virtual inertia addresses this concern by virtually emulating the behavior of a synchronous generator. However, the conventional VSG is designed based on two conditions: (i) fixed operating point and (ii) inductive grid connections. The performance of VSGs in low-voltage semi-resistive microgrids is far from optimal. To overcome the aforementioned concerns, a heuristic dynamic programing (HDP) approach is proposed to optimally control grid-connected VSGs. The neural-network-based inherence of the HDP enables the proposed technique to adapt to any impedance angle. The HDP controller includes two subnetworks: (i) the action network that controls the system optimally and (ii) the critic network, which evaluates the effectiveness of the action network. The simulation and experimental results are provided to evaluate the effectiveness of the proposed technique. As shown, the HDP-based approach illustrates a better performance in comparison with the conventional PI-based VSG in various operating conditions.

24 POWER TRANSMISSION AND DISTRIBUTION↗

An Intelligent Distributed Ledger Construction Algorithm for IoT

Blockchain is the next generation of secure data management that creates near-immutable decentralized storage. Secure cryptography created a niche for blockchain to provide alternatives to well-known security compromises. However, design bottlenecks with traditional blockchain data structures scale poorly with increased network usage and are extremely computation-intensive. This made the technology difficult to combine with limited devices, like those in Internet of Things networks. In protocols like IOTA, replacement of blockchain's linked-list queue processing with a lightweight dynamic ledger showed remarkable throughput performance increase. However, current stochastic algorithms for ledger construction suffer distinct trade-offs between efficiency and security. This work proposed a machine-learning approach with a multi-arm bandit that resolved these issues and was designed for auditing on limited devices. This algorithm was tested in a reinforcement-learning environment simulating the IOTA ledger's construction with a decision tree. This study showed through regret analysis and experimentation that this approach was secure against impulse manipulation attacks while remaining energy-efficient. Although the IOTA protocol was a pioneer for lightweight distributed ledgers, it is expected that future blockchain protocols will adopt techniques similar to those presented in this work.

multi-arm bandit↗