Search NASA⌕ Search

SEARCH · Search NASA

Results for “multiagent 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 19 records

Interpreting Primal-Dual Algorithms for Constrained Multiagent Reinforcement Learning

Constrained multiagent reinforcement learning (C-MARL) is gaining importance as MARL algorithms find new applications in real-world systems ranging from energy systems to drone swarms. Most C-MARL algorithms use a primal-dual approach to enforce constraints through a penalty function added to the reward. In this paper, we study the structural effects of this penalty term on the MARL problem. First, we show that the standard practice of using the constraint function as the penalty leads to a weak notion of safety. However, by making simple modifications to the penalty term, we can enforce meaningful probabilistic (chance and conditional value at risk) constraints. Second, we quantify the effect of the penalty term on the value function, uncovering an improved value estimation procedure. We use these insights to propose a constrained multiagent advantage actor critic (C-MAA2C) algorithm. Simulations in a simple constrained multiagent environment affirm that our reinterpretation of the primal-dual method in terms of probabilistic constraints is effective, and that our proposed value estimate accelerates convergence to a safe joint policy.

chance constraints↗

Interpreting Primal-Dual Algorithms for Constrained Multiagent Reinforcement Learning: Preprint

We study multiagent reinforcement learning (MARL) with constraints. This setting is gaining importance as MARL algorithms find new applications in real-world systems ranging from power grids to drone swarms. Most constrained MARL (C-MARL) algorithms use a primal-dual approach to enforce constraints through a penalty function added to the reward. In this paper, we study the structural effects of the primal-dual approach on the constraints and value function. First, we show that using the constraint evaluation as the penalty leads to a weak notion of safety, but by making simple modifications to the penalty function, we can enforce meaningful probabilistic safety constraints. Second, we show that the penalty term changes the value function in a way that is easy to model, and demonstrate the consequences of not doing so. We conclude with simulations in a simple constrained multiagent environment to back up the theoretical results.

data-driven control↗

Large-Eddy Simulation of Flow Over Boeing Gaussian Bump Using Multiagent Reinforcement Learning Wall Model: Preprint

We develop a wall model for large-eddy simulation (LES) that takes into account various pressure-gradient effects using multi-agent reinforcement learning. The model is trained using low-Reynolds-number flow over periodic hills with agents distributed on the wall at various computational grid points. It utilizes a wall eddy-viscosity formulation as the boundary condition to apply the modeled wall shear stress. Each agent receives states based on local instantaneous flow quantities at an off-wall location, computes a reward based on the estimated wall-shear stress, and provides an action to update the wall eddy viscosity at each time step. The trained wall model is validated in wall-modeled LES of flow over periodic hills at higher Reynolds numbers, and the results show the effectiveness of the model on flow with pressure gradients. The analysis of the trained model indicates that the model is capable of distinguishing between the various pressure gradient regimes present in the flow. To further assess the robustness of the developed wall model, simulations of flow over the Boeing Gaussian bump are conducted at a Reynolds number of 2 x 10^6, based on the free-stream velocity and the bump width. The results of mean skin friction and pressure on the bump surface, as well as the velocity statistics of the flow field, are compared to those obtained from equilibrium wall model (EQWM) simulations and published experimental data sets. The developed wall model is found to successfully capture the acceleration and deceleration of the turbulent boundary layer on the bump surface, providing better predictions of skin friction near the bump peak and exhibiting comparable performance to the EQWM with respect to the wall pressure and velocity field. We also conclude that the subgrid-scale model is crucial to the accurate prediction of the flow field, in particular the prediction of separation.

boundary layer↗

A Cooperative Multi-Agent Deep Reinforcement Learning Framework for Real-Time Residential Load Scheduling

Internet-of-Things (IoT) enabled monitoring and control capabilities are enabling increasing numbers of household users with controllable loads to actively participate in smart grid energy management. Realizing an efficient real-time energy management system that takes advantage of these developments requires novel techniques for managing the increased complexity of the control action space in resolving multiple challenges such as the uncertainty in energy prices and renewable energy output along with the need to satisy physical grid constraints such as transformer capacity. Addressing these challenges, we develop a multi-household energy management framework for residential units connected to the same transformer and containing DERs such as PV, ESS and controllable loads. The goal of our framework is to schedule controllable household appliances and ESS such that the cost of procuring electricity from the utility over a horizon is minimized while physical grid constraints are satisfied at each scheduling step. Traditional energy management frameworks either perform global optimization to satisfy grid constraints but suffer from high computational complexity (for example Integer Program, Mixed Integer Programming frameworks and centralized reinforcement learning) or perform decentralized real-time energy management without satisfying global grid constraints (for example multi-agent reinforcement learning with no cooperation). In contrast, we propose a cooperative multiagent reinforcement learning (MARL) framework that i) operates in real-time, and ii) performs explicit collaboration to satisfy global grid constraints. The novelty in our framework is two fold. Firstly, our framework trains multiple independent learners (IL) for each household in parallel using historical data and performs real-time inferencing of control actions using the most recent system state. Secondly, our framework contains a low complexity knapsack based cooperation agent which combines the outputs of ILs to minimize cost while satisfying grid constraints. Simulation results show that our cooperative MARL approach achieves significant cost improvement over centralized reinforcement learning and day-ahead planning baselines. Moreover, our approach strictly satisfies physical constraints with no apriori knowledge of system dynamics while the baseline approaches have occasional violations. We also measure the training and inference time by ranging the number of households from 1 to 25. Results show that our cooperative MARL approach scales best among various approaches.

14 SOLAR ENERGY↗

A Multiagent Deep Reinforcement Learning-Enabled Dual-Branch Damping Controller for Multimode Oscillation

Here, this study develops a multiagent deep reinforcement learning (MADRL)-enabled framework for the decentralized cooperative control of a novel dual-branch (DB) damping controller for both low-frequency oscillation (LFO) and ultralow-frequency oscillation (ULFO). It has two branches, each of which consists of a proportional resonance (PR) and a second-order polynomial that is designed to handle target oscillation modes. To improve the robustness of the controller to system uncertainties, MADRL is developed, where multiagents are centrally trained to obtain the coordinated adaptive control policy while being executed in a decentralized manner to provide the optimal parameter setting for each controller with only local states. Comparisons with the IEEE 10-machine 39-bus system demonstrate that the proposed method achieves better robustness to uncertainties, lower communication delay, and single-point failure, as well as damping control performances for both LFO and ULFO.

97 MATHEMATICS AND COMPUTING↗

Multi-Agent Reinforcement Learning for Distribution System Critical Load Restoration

Grid resilience has become a critical topic recently because of the increasing occurrence of extreme events and the growing integration of intermittent renewable energy sources. To build a resilient distribution system, this paper develops a multiagent reinforcement learning-based (MARL) method to coordinate distribution energy resources (DERs) dispatch, load pickup, and network reconfiguration for load restoration after a system outage. With the help of two types of control agents, namely critical load restoration (CLR) and coordination (COR) agents, system loads can be restored efficiently, given available resources. The effectiveness and superiority of the proposed algorithm are demonstrated through simulations and comparative studies on a real distribution feeder in Western Colorado.

distribution system↗

CLEANing the Reward: Counterfactual Actions to Remove Exploratory Action Noise in Multiagent Learning

Learning in multiagent systems can be slow because agents must learn both how to behave in a complex environment and how to account for the actions of other agents. The inability of an agent to distinguish between the true environmental dynamics and those caused by the stochastic exploratory actions of other agents creates noise in each agent's reward signal. This learning noise can have unforeseen and often undesirable effects on the resultant system performance. We define such noise as exploratory action noise, demonstrate the critical impact it can have on the learning process in multiagent settings, and introduce a reward structure to effectively remove such noise from each agent's reward signal. In particular, we introduce Coordinated Learning without Exploratory Action Noise (CLEAN) rewards and empirically demonstrate their benefits

Reinforcement Learning↗

Decentralized Voltage Control of Large-Scale Distribution System with PVs Based on MADRL

This paper proposes a model-free decentralized control framework for the voltage regulation of large-scale distribution systems through the coordinated control of PV inverters. This is achieved by developing a novel interaction mechanism between the surrogate model and the centralized training and decentralized execution multiagent deep reinforcement learning framework. Specifically, the sparse Gaussian processes regression method is first utilized to develop the surrogate model of the original distribution system for reward calculation during the training stage, where each agent represents a sub-region in the centralized fashion for coordination strategy learning. After that, the learned control rules are used to inform controllers within each sub-region for real-time decisions with only local measurements. Comparative tests among various methods on the EPRI Ckt5 test system demonstrate the effectiveness of the proposed method.

distribution system↗

UAS Conflict-Avoidance Using Multiagent RL with Abstract Strategy Type Communication

The use of unmanned aerial systems (UAS) in the national airspace is of growing interest to the research community. Safety and scalability of control algorithms are key to the successful integration of autonomous system into a human-populated airspace. In order to ensure safety while still maintaining efficient paths of travel, these algorithms must also accommodate heterogeneity of path strategies of its neighbors. We show that, using multiagent RL, we can improve the speed with which conflicts are resolved in cases with up to 80 aircraft within a section of the airspace. In addition, we show that the introduction of abstract agent strategy types to partition the state space is helpful in resolving conflicts, particularly in high congestion.

Unmanned Autonomous Systems↗

Announced Strategy Types in Multiagent RL for Conflict-Avoidance in the National Airspace

The use of unmanned aerial systems (UAS) in the national airspace is of growing interest to the research community. Safety and scalability of control algorithms are key to the successful integration of autonomous system into a human-populated airspace. In order to ensure safety while still maintaining efficient paths of travel, these algorithms must also accommodate heterogeneity of path strategies of its neighbors. We show that, using multiagent RL, we can improve the speed with which conflicts are resolved in cases with up to 80 aircraft within a section of the airspace. In addition, we show that the introduction of abstract agent strategy types to partition the state space is helpful in resolving conflicts, particularly in high congestion.

National Airspace↗

A Distributed Reinforcement Learning Yaw Control Approach for Wind Farm Energy Capture Maximization: Preprint

In this paper, we present a reinforcement-learning based distributed approach to wind farm energy capture maximization using yaw control, also known as wake steering. In order to maximize the power output of a wind farm, it is often necessary for individual turbines to decrease their own power output through yaw misalignment so as to deflect their wakes away from downstream turbines. Although using model-based methods to achieve yaw misalignment is one option, a model-free method might be better suited to incorporate factors that are difficult to model or changing conditions. We propose an algorithm that adapts concepts of temporal difference reinforcement learning distributed to a multi-agent environment that allows individual turbines to act so as to optimize overall wind farm output and react to unforeseen disturbances.

controls↗

Discovery and Analysis of Rare High-Impact Failure Modes using Adversarial RL-Informed Sampling

Adaptive learning agents have tremendous potential to handle critical tasks currently performed by humans. Unfortunately, due to their complexity, it can be difficult to verify that these learning agents do not have critical failure modes. Standard verification and validation methods often do not apply directly to learning agents and Monte Carlo methods have difficulty covering even a small fraction of the state space, especially in multiagent systems or over long time horizons. To overcome this difficulty, we demonstrate an adaptive stress-testing method based on reinforcement learning of correlations that raise the probability of failure. This approach has three key properties: (1) it is able to find rare failure modes with far greater sample efficiency than Monte Carlo methods, (2) it can estimate the true probability of a failure mode despite the inherent bias in the learning method, and (3) it is capable of learning and resampling compact representations of multimodal failure spaces. These properties are important in practice as we need to find disparate failure modes while accounting for their actual relevance. This is a significant advantage over traditional adaptive stress testing methods that give abstract likelihoods of particular failure instances, but cannot estimate the probability of a broader failure mode. We test our algorithm on a simple problem from the aviation domain where an autonomous aircraft lands in gusty wind conditions. The results suggest that we can find failure modes with far fewer samples than the Monte Carlo approach and simultaneously estimate the probability of failure.

reinforcement learning↗

Discovery and Analysis of Rare High-Impact Failure Modes using Adversarial RL-Informed Sampling

Adaptive learning agents have tremendous potential to handle critical tasks currently performed by humans. Unfortunately, due to their complexity, it can be difficult to verify that these learning agents do not have critical failure modes. Standard verification and validation methods often do not apply directly to learning agents and Monte Carlo methods have difficulty covering even a small fraction of the state space, especially in multiagent systems or over long time horizons. To overcome this difficulty, we demonstrate an adaptive stress-testing method based on reinforcement learning of correlations that raise the probability of failure. This approach has three key properties: (1) it is able to find rare failure modes with far greater sample efficiency than Monte Carlo methods, (2) it can estimate the true probability of a failure mode despite the inherent bias in the learning method, and (3) it is capable of learning and resampling compact representations of multimodal failure spaces. These properties are important in practice as we need to find disparate failure modes while accounting for their actual relevance. This is a significant advantage over traditional adaptive stress testing methods that give abstract likelihoods of particular failure instances, but cannot estimate the probability of a broader failure mode. We test our algorithm on a simple problem from the aviation domain where an autonomous aircraft lands in gusty wind conditions. The results suggest that we can find failure modes with far fewer samples than the Monte Carlo approach and simultaneously estimate the probability of failure.

Validation↗

Discovery and Analysis of Rare High-Impact Failure Modes using Adversarial RL-Informed Sampling

Adaptive learning agents have tremendous potential to handle critical tasks currently performed by humans. Unfortunately, due to their complexity, it can be difficult to verify that these learning agents do not have critical failure modes. Standard verification and validation methods often do not apply directly to learning agents and Monte Carlo methods have difficulty covering even a small fraction of the state space, especially in multiagent systems or over long time horizons. To overcome this difficulty, we demonstrate an adaptive stress-testing method based on reinforcement learning of correlations that raise the probability of failure. This approach has three key properties: (1) it is able to find rare failure modes with far greater sample efficiency than Monte Carlo methods, (2) it can estimate the true probability of a failure mode despite the inherent bias in the learning method, and (3) it is capable of learning and resampling compact representations of multimodal failure spaces. These properties are important in practice as we need to find disparate failure modes while accounting for their actual relevance. This is a significant advantage over traditional adaptive stress testing methods that give abstract likelihoods of particular failure instances, but cannot estimate the probability of a broader failure mode. We test our algorithm on a simple problem from the aviation domain where an autonomous aircraft lands in gusty wind conditions. The results suggest that we can find failure modes with far fewer samples than the Monte Carlo approach and simultaneously estimate the probability of failure.

Validation↗

A Novel LDPP-MADDPG Approach for Distributed Power Allocation in mmWave Cellular Networks

This paper considers the problem of distributed beam scheduling and power allocation problem in millimeter- Wave (mmWave) cellular networks, in which multiple Base Stations (BSs) operate as individual operators over a shared spectrum. We propose a novel learning-aided approach that integrates the Lyapunov Drift-Plus-Penalty (LDPP) framework and Multi-agent Deep Deterministic Policy Gradient (MADDPG) reinforcement learning algorithms. This offers a powerful approach to learning stable and constraint-aware policies, reaping the joint benefit of both LDPP and MADDPG, in complex multiagent environments. The major challenge for this approach is to integrate these two approaches in a meaningful and effective manner. The key idea to solve this problem is to introduce a novel feature of local observation that incorporates potential negative value of the reward function due to the stochastic constraints introduced by the LDPP framework. Empirical results demonstrate that our proposed scheme outperforms the baseline methods under various conditions.

99 - GENERAL AND MISCELLANEOUS↗

Learning Sequences of Actions in Collectives of Autonomous Agents

In this paper we focus on the problem of designing a collective of autonomous agents that individually learn sequences of actions such that the resultant sequence of joint actions achieves a predetermined global objective. We are particularly interested in instances of this problem where centralized control is either impossible or impractical. For single agent systems in similar domains, machine learning methods (e.g., reinforcement learners) have been successfully used. However, applying such solutions directly to multi-agent systems often proves problematic, as agents may work at cross-purposes, or have difficulty in evaluating their contribution to achievement of the global objective, or both. Accordingly, the crucial design step in multiagent systems centers on determining the private objectives of each agent so that as the agents strive for those objectives, the system reaches a good global solution. In this work we consider a version of this problem involving multiple autonomous agents in a grid world. We use concepts from collective intelligence to design goals for the agents that are 'aligned' with the global goal, and are 'learnable' in that agents can readily see how their behavior affects their utility. We show that reinforcement learning agents using those goals outperform both 'natural' extensions of single agent algorithms and global reinforcement, learning solutions based on 'team games'.

Turner, Kagan↗

A Modular and Transferable Reinforcement Learning Framework for the Fleet Rebalancing Problem

Mobility on demand (MoD) systems show great promise in realizing flexible and efficient urban transportation. However, significant technical challenges arise from operational decision making associated with MoD vehicle dispatch and fleet rebalancing. For this reason, operators tend to employ simplified algorithms that have been demonstrated to work well in a particular setting. To help bridge the gap between novel and existing methods, we propose a modular framework for fleet rebalancing based on model-free reinforcement learning (RL) that can leverage an existing dispatch method to minimize system cost. In particular, by treating dispatch as part of the environment dynamics, a centralized agent can learn to intermittently direct the dispatcher to reposition free vehicles and mitigate against fleet imbalance. We formulate RL state and action spaces as distributions over a grid partitioning of the operating area, making the framework scalable and avoiding the complexities associated with multiagent RL. Numerical experiments, using real-world trip and network data, demonstrate that RL reduces waiting time by 28% to 38% for the same-day evaluation, 17% to 44% for cross-day evaluation, and 22% to 25% for cross-season evaluation compared with no rebalancing scenarios. This approach has several distinct advantages over baseline methods including: improved system cost; high degree of adaptability to the selected dispatch method; and the ability to perform scale-invariant transfer learning between problem instances with similar vehicle and request distributions.

33 ADVANCED PROPULSION SYSTEMS↗

PowerGridworld: A Framework for Multi-Agent Reinforcement Learning in Power Systems [SWR-22-07]

NREL's PowerGridworld provides a modular simulation environment for training heterogenous, grid-aware, multi-agent reinforcement learning (RL) policies at scale. The package enables the user to create component gym environments that can be composed into more complex agents. For example, a grid interactive building environment can be created by composing together component environments each encapsulating the building, PV, and battery physics. These multi-component environments can then be combined into multi-agent simulation where each agent's power consumption/injection becomes an input for solving the optimal power flow on a distribution feeder modeled in OpenDSS. Information from OpenDSS, such as bus voltages and line flows, can be included in the agents' observation spaces to enable grid-aware rewards. The default API for the PowerGridworld simulator conforms to RLLib's MultiAgent API and thus enables distributed training using HPC and cloud resources.

Biagioni, David↗