Search NASA⌕ Search

SEARCH · Search NASA

Results for “Deep 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 307 records · Page 17

Designing reinforcement learning algorithms for building HVAC control: From experimental observation to simulation comparisons

Advanced supervisory-level control with reinforcement learning (RL) is regarded as a promising solution for HVAC systems to minimize energy consumption while maintaining thermal comfort and indoor air quality. However, most RL applications were conducted in the simulation environment rather than real-world HVAC systems. This paper developed a value-based RL controller termed Deep Q-Network (DQN) for a typical central HVAC system and evaluated its performance in a building test facility. By comparing DQN with a rule-based controller, the study not only demonstrated the cases where DQN could properly maintain indoor comfort but also discussed possible reasons why DQN failed in some other situations. Recognizing the limitations of value-based RL algorithms from the experimental tests, a simulation study was conducted to compare DQN with an alternative RL approach, an actor–critic algorithm termed Deep Deterministic Policy Gradient (DDPG). In scenarios with a relatively large action space, DDPG outperformed DQN by requiring fewer computational resources and achieving better thermal comfort, lower energy consumption, and more stable control actions. The findings suggest that the ability of DDPG to handle continuous control variables more effectively allows for faster convergence in training and more precise control in practice, which enhances the overall efficiency and reliability of the HVAC system.

Guo, Fangzhou↗

Adaptive Reinforcement Learning (ARL) Control of a Multi-port Resonant Converter in UAV Systems

This study presents an adaptive reinforcement learning (ARL) control framework for a multi-port resonant converter used in hybrid unmanned aerial vehicle (UAV) power systems. The converter integrates high-frequency half-bridge input ports connected to a rectified engine–generator set and a battery energy storage system, along with a semi-bridgeless active rectifier supplying the propulsion load. A deep RL agent is trained to dynamically regulate inter-port phase-shift commands in real time based on flight conditions and load power demand. The ARL controller autonomously identifies phase-shift combinations that maximize conversion efficiency while maintaining stable and coordinated power flow, even under rapidly varying operating scenarios. This data-driven approach eliminates the need for explicit system modeling or extensive manual tuning and enables coordinated control among multiple power ports without inter-port communication. Experimental results validate that the ARL based strategy achieves reliable power sharing and consistently high-efficiency operation across diverse UAV operating conditions.

Asa, Erdem [ORNL] (ORCID:0000000190884812)↗

Speeding up and reducing memory usage for scientific machine learning via mixed precision

Scientific machine learning (SciML) has emerged as a versatile approach to address complex computational science and engineering problems. Within this field, physics-informed neural networks (PINNs) and deep operator networks (DeepONets) stand out as the leading techniques for solving partial differential equations by incorporating both physical equations and experimental data. However, training PINNs and DeepONets require significant computational resources, including long computational times and large amounts of memory. In search of computational efficiency, training neural networks using half precision (float16) rather than the conventional single (float32) or double (float64) precision has gained substantial interest, given the inherent benefits of reduced computational time and memory consumed. However, we find that float16 cannot be applied to SciML methods, because of gradient divergence at the start of training, weight updates going to zero, and the inability to converge to a local minima. To overcome these limitations, we explore mixed precision, which is an approach that combines the float16 and float32 numerical formats to reduce memory usage and increase computational speed. Our experiments showcase that mixed precision training not only substantially decreases training times and memory demands but also maintains model accuracy. Here, we also reinforce our empirical observations with a theoretical analysis. The research has broad implications for SciML in various computational applications.

97 MATHEMATICS AND COMPUTING↗

Do optimization methods in deep learning applications matter?

With advances in deep learning, exponential data growth and increasing model complexity, developing efficient optimization methods are attracting much research attention. Several implementations favor the use of Conjugate Gradient (CG) and Stochastic Gradient Descent (SGD) as being practical and elegant solutions to achieve quick convergence, however, these optimization processes also present many limitations in learning across deep learning applications. Recent research is exploring higher-order optimization functions as better approaches, but these present very complex computational challenges for practical use. Comparing first and higher-order optimization functions, in this paper, our experiments reveal that Levemberg-Marquardt (LM) significantly supersedes optimal convergence but suffers from very large processing time increasing the training complexity of both, classification and reinforcement learning problems. Our experiments compare off-the-shelf optimization functions(CG, SGD, LM and L-BFGS) in standard CIFAR, MNIST, CartPole and FlappyBird experiments.The paper presents arguments on which optimization functions to use and further, which functions would benefit from parallelization efforts to improve pretraining time and learning rate convergence.

Ozyildirim, Buse Melis↗

FIRE: A Failure-Adaptive RL Framework for Edge Computing Migrations

In edge computing, users' service profiles are migrated between edge servers due to user mobility. Reinforcement Learning (RL) frameworks have been proposed to do so, often trained on simulated data. However, existing RL frameworks overlook occasional server failures, which although rare, impact latency-sensitive applications like AR/VR and real- time obstacle detection. These rare failures, being not adequately represented in historical training data, pose a challenge for data-driven RL algorithms. We introduce FIRE, a framework that adapts to rare events by training a RL policy in an edge computing digital twin environment. We propose FIRE-ImRE, an importance sampling-based Q-learning algorithm, which samples rare events proportionally to their impact on the value function. FIRE considers delay, migration, failure, and backup placement costs across individual and shared service profiles. We prove FIRE-ImRE's boundedness and convergence to optimality. Next, we introduce novel deep Q-learning (FIRE-ImDQL) and actor critic (FIRE-ImACRE) versions of our algorithm to enhance scalability. Here, we extend our framework to accommodate users with varying risk tolerances of rare failure events. Through trace-driven experiments, we show that FIRE reduces edge computing costs compared to vanilla RL and the greedy baseline in the event of failures.

Edge computing↗

Reinforcement Learning-Based Oscillation Dampening: Scaling Up Single-Agent Reinforcement Learning Algorithms to a 100-Autonomous-Vehicle Highway Field Operational Test

In this article, we explore the technical details of the reinforcement learning (RL) algorithms that were deployed in the largest field test of automated vehicles designed to smooth traffic flow in history as of 2023, uncovering the challenges and breakthroughs that come with developing RL controllers for automated vehicles. We delve into the fundamental concepts behind RL algorithms and their application in the context of self-driving cars, discussing the developmental process from simulation to deployment in detail, from designing simulators to reward function shaping. We present the results in both simulation and deployment, discussing the flow-smoothing benefits of the RL controller. From understanding the basics of Markov decision processes to exploring advanced techniques such as deep RL, our article offers a comprehensive overview and deep dive of the theoretical foundations and practical implementations driving this rapidly evolving field. We also showcase real-world case studies and alternative research projects that highlight the impact of RL controllers in revolutionizing autonomous driving. From tackling complex urban environments to dealing with unpredictable traffic scenarios, these intelligent controllers are pushing the boundaries of what automated vehicles can achieve. Furthermore, we examine the safety considerations and hardware-focused technical details surrounding deployment of RL controllers into automated vehicles. As these algorithms learn and evolve through interactions with the environment, ensuring their behavior aligns with safety standards becomes crucial. Here, we explore the methodologies and frameworks being developed to address these challenges, emphasizing the importance of building reliable control systems for automated vehicles.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗

Double Deep Q-Networks for Optimizing Electricity Cost of a Water Heater

Electric water heaters represent 14% of the electricity consumption in the residential buildings and the cost associated with domestic water heating account for a good portion of the household expenses in the United States. In this context, intelligent control of water heaters gained a lot of research attention. In recent years, a significant number of intelligent water heater controllers, with various methods and intended uses, have been proposed. However, existing studies are mostly model-based approaches that require an accurate modelling of the water heater. Towards addressing this research gap, this paper presents a model-free reinforcement learning-based controller for a day-ahead price market. The controller aims to minimize the cost of domestic water heating while maintaining the user comfort. The results showed that the developed controller can help save energy cost while maintaining the temperatures within the desired comfort band.

Amasyali, Kadir↗

Microsecond-latency feedback at a particle accelerator by online reinforcement learning on hardware

The commissioning and operation of future large-scale scientific experiments will challenge current tuning and control methods. Reinforcement learning (RL) algorithms are a promising solution due to their ability to dynamically adapt to changing environments and consider delayed consequences. In many real-world applications, RL policies must produce actions in real time, often within microseconds to milliseconds, imposing significant constraints on system latency and computational overhead that conventional machine learning libraries are not designed to handle. To control phenomena in real time at these timescales, RL needs to be deployed on-the-edge, namely on dedicated hardware located near the system it controls, without relying on a host CPU or cloud-based inference. In this work we present the design and deployment of an experience accumulator system in a particle accelerator. In this system, deep-RL algorithms run using hardware acceleration and act within a few microseconds, enabling the use of RL for control of phenomena like beam instabilities. The training uses the collected data offline to reduce the number of operations carried out on the acceleration hardware. The proposed architecture was tested in real experimental conditions at the Karlsruhe research accelerator, a synchrotron light source, where the system was used to control artificially induced horizontal betatron oscillations in real-time, with a control loop period of just 2.7 μs. The results showed a performance comparable to the commercial feedback system available at the accelerator, demonstrating the viability and potential of this approach. Due to the self-learning and reconfiguration capability of this implementation, a seamless application to other control problems is possible. Applications range from particle accelerators to large-scale research and industrial facilities.

FPGA↗

Self-Driving Telescopes: Autonomous Scheduling of Astronomical Observation Campaigns with Offline Reinforcement Learning

Modern astronomical experiments are designed to achieve multiple scientific goals, from studies of galaxy evolution to cosmic acceleration. These goals require data of many different classes of night-sky objects, each of which has a particular set of observational needs. These observational needs are typically in strong competition with one another. This poses a challenging multi-objective optimization problem that remains unsolved. The effectiveness of Reinforcement Learning (RL) as a valuable paradigm for training autonomous systems has been well-demonstrated, and it may provide the basis for self-driving telescopes capable of optimizing the scheduling for astronomy campaigns. Simulated datasets containing examples of interactions between a telescope and a discrete set of sky locations on the celestial sphere can be used to train an RL model to sequentially gather data from these several locations to maximize a cumulative reward as a measure of the quality of the data gathered. We use simulated data to test and compare multiple implementations of a Deep Q-Network (DQN) for the task of optimizing the schedule of observations from the Stone Edge Observatory (SEO). We combine multiple improvements on the DQN and adjustments to the dataset, showing that DQNs can achieve an average reward of 87%+-6% of the maximum achievable reward in each state on the test set. This is the first comparison of offline RL algorithms for a particular astronomical challenge and the first open-source framework for performing such a comparison and assessment task.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Evaluation of Advanced Signal Processing Techniques to Improve Detection and Identification of Embedded Defects

By the end of 1996, 109 Nuclear Power Plants were operating in the United States, producing 22% of the Nation’s electricity [1]. At present, more than two thirds of these power plants are more than 40 years old. The purpose of the U.S. Department of Energy Office of Nuclear Energy’s Light Water Reactor Sustainability (LWRS) Program is to develop technologies and other solutions that can improve the reliability, sustain the safety, and extend the operating lifetimes of nuclear power plants (NPPs) beyond 60 years [2]. The most important safety structures in an NPP are constructed of concrete. The structures generally do not allow for destructive evaluation and access is limited to one side of the concrete element. Therefore, there is a need for techniques and technologies that can assess the internal health of complex, reinforced concrete structures nondestructively. Previously, we documented the challenges associated with Non-Destructive Evaluation (NDE) of thick, reinforced concrete sections and prioritized conceptual designs of specimens that could be fabricated to represent NPP concrete structures [3]. Consequently, a 7 feet tall, by 7 feet wide, by 3 feet and 4-inch-thick concrete specimen was constructed with 2.257-inch-and 1-inch-diameter rebar every 6 to 12 inches. In addition, defects were embedded the specimen to assess the performance of existing and future NDE techniques. The defects were designed to give a mix of realistic and controlled defects for assessment of the necessary measures needed to overcome the challenges with more heavily reinforced concrete structures. Information on the embedded defects is documented in [4]. We also documented the superiority of Frequency Banded Decomposition (FBD) Synthetic Aperture Focusing Technique (SAFT) over conventional SAFT when probing defects under deep concrete cover. Improvements include seeing an intensity corresponding to a defect that is either not visible at all in regular, full frequency content SAFT, or an improvement in contrast over conventional SAFT reconstructed images. This report documents our efforts in four fronts: 1) Comparative study between traditional SAFT and FBD SAFT for concrete specimen with and without Alkali-Silica Reaction (ASR) damage, 2) improvement of our Model-Based Iterative Reconstruction (MBIR) for thick reinforced concrete [5], 3) development of a universal framework for sharing, reconstruction, and visualization of ultrasound NDE datasets, and 4) application of machine learning techniques for automated detection of ASR inside concrete. Our comparative study between FBD and traditional SAFT reconstruction images shows a clear difference between images of ASR and non-ASR specimens. In particular, the left first harmonic shows an increased contrast and sensitivity to ASR damage. For MBIR, we show the superiority of model-based techniques over delay and sum techniques such as SAFT. Improvements include elimination of artifacts caused by direct arrival signals, and increased contrast and Signal to Noise Ratio. For the universal framework, we document a format for data storage based on the HDF5 file format, and also propose a modular Graphic User Interface (GUI) for easy customization of data conversion, reconstruction, and visualization routines. Finally, two techniques for ASR automated detection are presented. The first technique is based on an analysis of the frequency content using Hilbert Transform Indicator (HTI) and the second technique employees Artificial Neural Network (ANN) techniques for training and classification of ultrasound data as ASR or non-ASR damaged classes. The ANN technique shows great potential with classification accuracy above 95%. These approaches are extensible to the detection of additional reinforced, thick concrete defects and damage.

11 NUCLEAR FUEL CYCLE AND FUEL MATERIALS↗

ReLVis: Visual Analytics for Situational Awareness During Reinforcement Learning Experimentation

Reinforcement learning (RL) is a branch of machine learning where an agent learns to maximize reward through trial and error. RL is challenging and data/compute intensive, and RL practitioners can easily become overwhelmed and make poor modeling decisions. Our contribution is a visual analytics tool designed to help data scientists maintain situation awareness during RL experimentation. Our tool allows the user to under- stand which hyper-parameter values lead to better or worse outcomes, what behaviors are associated with high and low reward, and how behaviors evolve and progress over training time. We evaluated our tool by applying state of the art deep RL models to six different RL tasks. We present three tasks as use cases that show how our tool leads to insightful findings and improves situation awareness.

Saldanha, Emily G.↗

"Reinforcement Learning Based Oscillation Dampening: Scaling up Single-Agent RL algorithms to a 100 AV highway field operational test"

In this article, we explore the technical details of the reinforcement learning (RL) algorithms that were deployed in the largest field test of automated vehicles designed to smooth traffic flow in history as of 2023, uncovering the challenges and breakthroughs that come with developing RL controllers for automated vehicles. We delve into the fundamental concepts behind RL algorithms and their application in the context of self-driving cars, discussing the developmental process from simulation to deployment in detail, from designing simulators to reward function shaping. We present the results in both simulation and deployment, discussing the flow-smoothing benefits of the RL controller. From understanding the basics of Markov decision processes to exploring advanced techniques such as deep RL, our article offers a comprehensive overview and deep dive of the theoretical foundations and practical implementations driving this rapidly evolving field. We also showcase real-world case studies and alternative research projects that highlight the impact of RL controllers in revolutionizing autonomous driving. From tackling complex urban environments to dealing with unpredictable traffic scenarios, these intelligent controllers are pushing the boundaries of what automated vehicles can achieve. Furthermore, we examine the safety considerations and hardware-focused technical details surrounding deployment of RL controllers into automated vehicles. As these algorithms learn and evolve through interactions with the environment, ensuring their behavior aligns with safety standards becomes crucial. We explore the methodologies and frameworks being developed to address these challenges, emphasizing the importance of building reliable control systems for automated vehicles.

Jang, Kathy↗

Language model-accelerated deep symbolic optimization

Symbolic optimization methods have been used to solve varied challenging and relevant problems such as symbolic regression and neural architecture search. However, the current state of the art typically learns each problem from scratch and is unable to leverage pre-existing knowledge and datasets that are available for many applications. Here, inspired by the similarity between sequence representations learned in natural language processing and the formulation of symbolic optimization as a discrete sequence optimization problem, we propose language model-accelerated deep symbolic optimization (LA-DSO), a method that leverages language models to learn symbolic optimization solutions more efficiently. We demonstrate LA-DSO in two tasks: symbolic regression, which allows us to perform extensive experimentation due to its low computation requirements, and computational antibody optimization, which shows that our proposal accelerates learning in challenging real-world problems.

97 MATHEMATICS AND COMPUTING↗

Modeling inter‐reader variability in clinical target volume delineation for soft tissue sarcomas using diffusion model

Abstract Background Accurate delineation of the clinical target volume (CTV) is essential in the radiotherapy treatment of soft tissue sarcomas. However, this process is subject to inter‐reader variability due to the need for clinical assessment of risk and extent of potential microscopic spread. This can lead to inconsistencies in treatment planning, potentially impacting treatment outcomes. Most existing automatic CTV delineation methods do not account for this variability and can only generate a single CTV for each case. Purpose This study aims to develop a deep learning‐based technique to generate multiple CTV contours for each case, simulating the inter‐reader variability in the clinical practice. Methods We employed a publicly available dataset consisting of fluorodeoxyglucose positron emission tomography (FDG‐PET), x‐ray computed tomography (CT), and pre‐contrast T1‐weighted magnetic resonance imaging (MRI) scans from 51 patients with soft tissue sarcoma, along with an independent validation set containing five additional patients. An experienced reader drew a contour of the gross tumor volume (GTV) for each patient based on multi‐modality images. Subsequently, two additional readers, together with the first one, were responsible for contouring three CTVs in total based on the GTV. We developed a diffusion model‐based deep learning method that is capable of generating arbitrary number of different and plausible CTVs to mimic the inter‐reader variability in CTV delineation. The proposed model incorporates a separate encoder to extract features from the GTV masks, leveraging the critical role of GTV information in accurate CTV delineation. Results The proposed diffusion model demonstrated superior performance with the highest Dice Index (0.902 compared to values below 0.881 for state‐of‐the‐art models) and the best generalized energy distance (GED) (0.209 compared to values exceeding 0.221 for state‐of‐the‐art models). It also achieved the second‐highest recall and precision metrics among the compared ambiguous image segmentation models. Results from both datasets exhibited consistent trends, reinforcing the reliability of our findings. Additionally, ablation studies exploring different model structures and input configurations highlighted the significance of incorporating prior GTV information for accurate CTV delineation. Conclusions The proposed diffusion model successfully generates multiple plausible CTV contours for soft tissue sarcomas, effectively capturing inter‐reader variability in CTV delineation.

Dong, Yafei [Yale Biomedical Imaging Institute Yal↗

Reinforcement Learning Approach to Cybersecurity in Space (RELACSS)

Securing satellite groundstations against cyber-attacks is vital to national security missions. However, these cyber threats are constantly evolving. As vulnerabilities are discovered and patched, new vulnerabilities are discovered and exploited. In order to automate the process of discovering existing vulnerabilities and the means to exploit them, a reinforcement learning framework is presented in this report. We demonstrate that this framework can learn to successfully navigate an unknown network and detect nodes of interest despite the presence of a moving target defense. The agent then exfiltrates a file of interest from the node as quickly as possible. This framework also incorporates a defensive software agent that learns to impede the attacking agents progress. This setup allows for the agents to work against each other and improve their abilities. We anticipate that this capability will help uncover unforeseen vulnerabilities and the means to mitigate them. The modular nature of the framework enables users to swap out learning algorithms and modify the reward functions in order to adapt the learning tasks to various use cases and environments. Several algorithms, viz., tabular Q learning, deep Q networks, proximal policy optimization, advantage actor-critic, generative adversarial imitation learning, are explored for the agents and the results highlighted. The agent learns to solve the tasks in a light-weight abstract environment. Once the agent learns to perform sufficiently well, it can be deployed in a minimega virtual machine environment (or a real network) with wrappers that map abstract actions to software commands. The agent also uses a local representation of the actions called a ‘slot-mechanism’. This allows the agent to learn in a certain network and generalize it to different networks. The defensive agent learns to predict the actions taken by an offensive agent and uses that information to anticipate the threat. This information can then either be used to raise an alarm or to take actions to thwart the attack. We believe that with the appropriate reward design, a representative environment, and action set, this framework can be generalized to tackle other cybersecurity tasks. By sufficiently training these agents, we can anticipate vulnerabilities leading to robust future designs. We can also deploy automated defensive agents that can help secure satellite groundstation and their vital national security missions.

45 MILITARY TECHNOLOGY, WEAPONRY, AND NATIONAL DEF↗

Exploration with Scalable Gaussian Process Reinforcement Learning

Exploration is a challenging problem in reinforcement learning (RL), especially in environments with sparse rewards. Quantifying and utilizing the parametric uncertainty has been shown to be paramount for successful exploration [Osband et al., 2018]. Bayesian, or approximately Bayesian, methods present a principled means of estimating the parametric uncertainty in RL problems. Gaussian processes, nonparametric Bayesian models, are often impractical due to poor scalability and computational bottlenecks. We introduce a scalable Gaussian process RL (GPRL) method which directly induces sparsity in the covariance matrix to facilitate faster computation. This is a departure from previous GPRL methods which instead rely on data reduction and subsampling. We compare various covariance-based exploration techniques (Thompson sampling, upper confidence bound, and probabilistic maximum variance) which leverage our scalable GP framework in sparse reward environments. Finally, we show favorable comparison against the bootstrapped deep Q-Network.

97 MATHEMATICS AND COMPUTING↗

Solving high-dimensional partial integral differential equations: The finite expression method

Partial integro-differential equations (PIDEs) have broad applications in the sciences, from electro-magnetism to options pricing. Here, in this paper, we introduce a new finite expression method (FEX) to solve PIDEs. This approach builds upon the original FEX and its inherent advantages with new advances: 1) A novel method of parameter grouping is proposed to reduce the number of coefficients in high-dimensional function approximation; 2) A Taylor series approximation method is implemented to significantly improve the computational efficiency and accuracy of the evaluation of the integral terms of PIDEs. The new FEX based method, denoted FEX-PG to indicate the addition of the parameter grouping (PG) step to the algorithm, provides both high accuracy and interpretable numerical solutions, with the outcome being an explicit equation that facilitates intuitive understanding of the underlying solution structures. These features are often absent in traditional methods, such as finite element methods (FEM) and finite difference methods, as well as in deep learning-based approaches. To benchmark our method against recent advances, we apply the new FEX-PG to solve benchmark PIDEs in the literature. In high-dimensional settings, FEX-PG exhibits strong and robust performance, achieving relative errors on the order of single precision machine epsilon, significantly outperforming existing approaches based on neural networks.

Combinatorial optimization↗

Finite-Time Analysis of Whittle Index based Q-Learning for Restless Multi-Armed Bandits with Neural Network Function Approximation

Whittle index policy is a heuristic to the intractable restless multi-armed bandits (RMAB) problem. Although it is provably asymptotically optimal, finding Whittle indices remains difficult. In this paper, we present Neural-Q-Whittle, a Whittle index based Q-learning algorithm for RMAB with neural network function approximation, which is an example of nonlinear two-timescale stochastic approximation with Q-function values updated on a faster timescale and Whittle indices on a slower timescale. Despite the empirical success of deep Q-learning, the non-asymptotic convergence rate of Neural-Q-Whittle, which couples neural networks with two-timescale Q-learning largely remains unclear. This paper provides a finite-time analysis of Neural-Q-Whittle, where data are generated from a Markov chain, and Q-function is approximated by a ReLU neural network. Our analysis leverages a Lyapunov drift approach to capture the evolution of two coupled parameters, and the nonlinearity in value function approximation further requires us to characterize the approximation error. Combing these provide Neural-Q-Whittle with convergence rate, where is the number of iterations.

reinforcement learning, structured learning, conve↗