Search NASA⌕ Search

SEARCH · Search NASA

Results for “Offline 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.

14 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

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 ↗
DOE OSTI2023

Unlocking Autonomous Telescopes through Reinforcement Learning: An Offline Framework and Insights from a Case Study

Optimizing observational astronomy campaigns is becoming a complex and expensive task for next-generation telescopes, where manual planning of observations may tend to reach suboptimal results in terms of optimization.Reinforcement Learning (RL) has been well-demonstrated as a valuable approach for training autonomous systems, and it may provide the basis for self-driving telescopes capable of scanning the sky and optimizing the scheduling for astronomy campaigns.We have developed a framework for the optimization of telescope scheduling using RL techniques, based on a dataset containing data on a discrete set of sky locations that the telescope should visit, and a reward metric. We compared several RL algorithms applied to an offline simulation dataset based at the Stone Edge Observatory, considering a discrete set of sky locations to visit and using “t-effective” as a reward metric, a measure of the quality of the data.Deep Q-Networks (DQNs), belonging to the class of value-based methods, have shown remarkable success in the optimization of astronomical observations in our dataset. In the full environment, the average reward value in each state was found to be 92%±5% of the maximum possible reward, while on the test set it resulted in 87%±9% of the maximum possible reward.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2025

Efficient and assured reinforcement learning-based building HVAC control with heterogeneous expert-guided training

Abstract Building heating, ventilation, and air conditioning (HVAC) systems account for nearly half of building energy consumption and $$20\%$$ of total energy consumption in the US. Their operation is also crucial for ensuring the physical and mental health of building occupants. Compared with traditional model-based HVAC control methods, the recent model-free deep reinforcement learning (DRL) based methods have shown good performance while do not require the development of detailed and costly physical models. However, these model-free DRL approaches often suffer from long training time to reach a good performance, which is a major obstacle for their practical deployment. In this work, we present a systematic approach to accelerate online reinforcement learning for HVAC control by taking full advantage of the knowledge from domain experts in various forms . Specifically, the algorithm stages include learning expert functions from existing abstract physical models and from historical data via offline reinforcement learning, integrating the expert functions with rule-based guidelines, conducting training guided by the integrated expert function and performing policy initialization from distilled expert function. Moreover, to ensure that the learned DRL-based HVAC controller can effectively keep room temperature within the comfortable range for occupants, we design a runtime shielding framework to reduce the temperature violation rate and incorporate the learned controller into it. Experimental results demonstrate up to 8.8 X speedup in DRL training from our approach over previous methods, with low temperature violation rate.

Xu, Shichao↗
DOE OSTI2024

An innovative heterogeneous transfer learning framework to enhance the scalability of deep reinforcement learning controllers in buildings with integrated energy systems

Deep Reinforcement Learning (DRL)-based control shows enhanced performance in the management of integrated energy systems when compared with Rule-Based Controllers (RBCs), but it still lacks scalability and generalisation due to the necessity of using tailored models for the training process. Transfer Learning (TL) is a potential solution to address this limitation. However, existing TL applications in building control have been mostly tested among buildings with similar features, not addressing the need to scale up advanced control in real-world scenarios with diverse energy systems. This paper assesses the performance of an online heterogeneous TL strategy, comparing it with RBC and offline and online DRL controllers in a simulation setup using EnergyPlus and Python. The study tests the transfer in both transductive and inductive settings of a DRL policy designed to manage a chiller coupled with a Thermal Energy Storage (TES). The control policy is pre-trained on a source building and transferred to various target buildings characterised by an integrated energy system including photovoltaic and battery energy storage systems, different building envelope features, occupancy schedule and boundary conditions (e.g., weather and price signal). The TL approach incorporates model slicing, imitation learning and fine-tuning to handle diverse state spaces and reward functions between source and target buildings. Results show that the proposed methodology leads to a reduction of 10% in electricity cost and between 10% and 40% in the mean value of the daily average temperature violation rate compared to RBC and online DRL controllers. Moreover, online TL maximises self-sufficiency and self-consumption by 9% and 11% with respect to RBC. Conversely, online TL achieves worse performance compared to offline DRL in either transductive or inductive settings. However, offline Deep Reinforcement Learning (DRL) agents should be trained at least for 15 episodes to reach the same level of performance as the online TL. Therefore, the proposed online TL methodology is effective, completely model-free and it can be directly implemented in real buildings with satisfying performance.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2023

Correlated Trajectory Uncertainty for Adaptive Sequential Decision Making

One of the great challenges with decision making tasks on real world systems is the fact that data is sparse and acquiring additional data is expensive. In these cases, it is often crucial to make a model of the environment to assist in making decisions. At the same time, limited data means that learned models are erroneous, making it just as important to equip the model with good predictive uncertainties. In the context of learning sequential decision making policies, these uncertainties can prove useful for informing which data to collect for the greatest improvement in policy performance \citep{mehta2021experimental, mehta2022exploration} or informing the policy about unsure regions of state and action space to avoid during test time \citep{yu2020mopo}. Additionally, assuming that realistic samples of the environment can be drawn, an adaptable policy can be trained that attempts to make optimal decisions for any given possible instance of the environment \citep{ghosh2022offline, chen2021offline}. In this work, we examine the so-called ``probabilistic neural network'' (PNN) model that is ubiquitous in model-based reinforcement learning (MBRL) works. We argue that while PNN models may have good marginal uncertainties, they form a distribution of non-smooth transition functions. Not only are these samples unrealistic and may hamper adaptability, but we also assert that this leads to poor uncertainty estimates when predicting multiple step trajectory estimates. To address this issue, we propose a simple sampling method that can be implemented on top of pre-existing models.We evaluate our sampling technique on a number of environments, including a realistic nuclear fusion task, and find that, not only do smooth transition function samples produce more calibrated uncertainties, but they also lead to better downstream performance for an adaptive policy.

Offline Reinforcement Learning↗
DOE OSTI2025

Safe and Robust Binary Classification and Fault Detection Using Reinforcement Learning

In this paper, we propose a learning-based method utilizing the Soft Actor-Critic (SAC) algorithm to train a binary Support Vector Machine (SVM) classifier. This classifier is designed to identify valid input spaces in high-dimensional, highly constrained systems while minimizing the total runtime of offline simulations. The simulations adapt their runtime based on the likelihood that a given training input will be informative to the classifier. Furthermore, we introduce a method for using the trained SAC model to predict whether a desired system input is likely to violate constraints, along with a technique to adjust the input as necessary. Additionally, we explore the potential of this model to detect faults or adversarial attacks within the system. The effectiveness of our approach is demonstrated through various simulations of challenging classification problems and a constrained quadrotor model.

Netter, Josh [Georgia Institute of Technology, Atl↗
DOE OSTI2024

Learning Robust Marking Policies for Adaptive Mesh Refinement

Here in this work, we revisit the marking decisions made in the standard adaptive finite element method (AFEM). Experience shows that a naïve marking policy leads to inefficient use of computational resources for adaptive mesh refinement (AMR). Consequently, using AMR in practice often involves ad-hoc or time-consuming offline parameter tuning to set appropriate parameters for the marking subroutine. To address these practical concerns, we recast AMR as a Markov decision process in which refinement parameters can be selected on-the-fly at run time, without the need for pre-tuning by expert users. In this new paradigm, the refinement parameters are also chosen adaptively via a marking policy that can be optimized using methods from reinforcement learning. We use the Poisson equation to demonstrate our techniques on h- and hp-refinement benchmark problems, and our experiments suggest that superior marking policies remain undiscovered for many classical AFEM applications. Furthermore, an unexpected observation from this work is that marking policies trained on one family of PDEs are sometimes robust enough to perform well on problems far outside the training family. For illustration, we show that a simple hp-refinement policy trained on 2D domains with only a single re-entrant corner can be deployed on far more complicated 2D domains, and even 3D domains, without significant performance loss. For reproduction and broader adoption, we accompany this work with an open-source implementation of our methods.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Reinforcement Learning-Based Approach for EMT Automation of Large-Scale PV Plants

In the pursuit of efficient and precise modeling of large-scale power systems, particularly utility-scale photovoltaic (PV) plants, Electromagnetic Transient (EMT) simulations play a crucial role. As utility-scale PV plants increase in size and complexity, traditional computational methods become inadequate, necessitating more advanced techniques. This paper highlights the progressive efforts made to accelerate EMT simulations. A novel continuous reinforcement learning (RL) strategy is explored to automate the differentiation and categorization of stiff and non-stiff differential algebraic equations (DAEs). The use of stiff and non-stiff integration methods applied to relevant parts of the DAEs assists with the speed-up of the simulations. The paper details the data acquisition, development and offline training of the RL model, leading to its validation that demonstrates a high precision in optimizing simulation methods. The proposed RL promises to significantly enhance the efficacy of EMT simulations, offering a robust framework for the future of power system analysis.

Xia, Qianxue↗
DOE OSTI2025

Risk-Aware Reinforcement Learning Framework for User-Centric O-RAN

The evolution of Open Radio Access Networks (O-RAN) presents an opportunity to enhance network performance by enabling dynamic orchestration of configuration and optimization parameters (COPs) through online learning methods. However, leveraging this potential requires overcoming the limitations of traditional cell-centric RAN architectures, which lack the necessary flexibility. On the other hand, despite their recent popularity, the practical deployment of online learning frameworks, such as Deep Reinforcement Learning (DRL)-based COP optimization solutions, remains limited due to their risk of deteriorating network performance during the exploration phase. In this article, we propose and analyze a novel risk-aware DRL framework for user-centric RAN (UC-RAN), which offers both the architectural flexibility and COP optimization to exploit this flexibility. We investigate and identify UC-RAN COPs that can be optimized via a soft actor-critic algorithm implementable as an O-RAN application (rApp) to jointly maximize latency satisfaction, reliability satisfaction, area spectral efficiency, and energy efficiency. We use the offline learning on UC-RAN to reliably accelerate DRL training, thus minimizing the risk of DRL deteriorating cellular network performance. Results show that our proposed solution approaches near-optimal performance in just a few hundred iterations with a decrease in risk score by a factor of ten.

6G and beyond↗
DOE OSTI2023

Beyond PID Controllers: PPO with Neuralized PID Policy for Proton Beam Intensity Control in Mu2e

We introduce a novel Proximal Policy Optimization (PPO) algorithm aimed at addressing the challenge of maintaining a uniform proton beam intensity delivery in the Muon to Electron Conversion Experiment (Mu2e) at Fermi National Accelerator Laboratory (Fermilab). Our primary objective is to regulate the spill process to ensure a consistent intensity profile, with the ultimate goal of creating an automated controller capable of providing real-time feedback and calibration of the Spill Regulation System (SRS) parameters on a millisecond timescale. We treat the Mu2e accelerator system as a Markov Decision Process suitable for Reinforcement Learning (RL), utilizing PPO to reduce bias and enhance training stability. A key innovation in our approach is the integration of a neuralized Proportional-Integral-Derivative (PID) controller into the policy function, resulting in a significant improvement in the Spill Duty Factor (SDF) by 13.6%, surpassing the performance of the current PID controller baseline by an additional 1.6%. This paper presents the preliminary offline results based on a differentiable simulator of the Mu2e accelerator. It paves the groundwork for real-time implementations and applications, representing a crucial step towards automated proton beam intensity control for the Mu2e experiment.

43 PARTICLE ACCELERATORS↗
DOE OSTI2025

Neural network approaches for parameterized optimal control

Here, we consider numerical approaches for deterministic, finite-dimensional optimal control problems whose dynamics depend on unknown or uncertain parameters. We seek to amortize the solution over a set of relevant parameters in an offline stage to enable rapid decision-making and be able to react to changes in the parameter in the online stage. To tackle the curse of dimensionality arising when the state and/or parameter are high-dimensional, we represent the policy using neural networks. We compare two training paradigms: First, our model-based approach leverages the dynamics and definition of the objective function to learn the value function of the parameterized optimal control problem and obtain the policy using a feedback form. Second, we use actor-critic reinforcement learning to approximate the policy in a data-driven way. Using an example involving a two-dimensional convection-diffusion equation, which features high-dimensional state and parameter spaces, we investigate the accuracy and efficiency of both training paradigms. While both paradigms lead to a reasonable approximation of the policy, the model-based approach is more accurate and considerably reduces the number of PDE solves.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2026

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↗
DOE OSTI2024

Learning Constrained Parametric Differentiable Predictive Control Policies With Guarantees

We present differentiable predictive control (DPC), a method for offline learning of constrained neural control policies for nonlinear dynamical systems with performance guarantees. We show that the sensitivities of the parametric optimal control problem can be used to obtain direct policy gradients. Specifically, we employ automatic differentiation (AD) to efficiently compute the sensitivities of the model predictive control (MPC) objective function and constraints penalties. To guarantee safety upon deployment, we derive probabilistic guarantees on closed-loop stability and constraint satisfaction based on indicator functions and Hoeffding’s inequality. We empirically demonstrate that the proposed method can learn neural control policies for various parametric optimal control tasks. In particular, we show that the proposed DPC method can stabilize systems with unstable dynamics, track time-varying references, and satisfy nonlinear state and input constraints. Our DPC method has practical time savings compared to alternative approaches for fast and memory-efficient controller design. Specifically, DPC does not depend on a supervisory controller as opposed to approximate MPC based on imitation learning. We demonstrate that, without losing performance, DPC is scalable with greatly reduced demands on memory and computation compared to implicit and explicit MPC while being more sample efficient than model-free reinforcement learning (RL) algorithms.

97 MATHEMATICS AND COMPUTING↗