Engineering topics
Biagioni, David
Publications and source records attributed to Biagioni, David.
Non-Stationary Policy Learning for Multi-Timescale Multi-Agent Reinforcement Learning
In multi-timescale multi-agent reinforcement learning (MARL), agents interact across different timescales. In general, policies for time-dependent behaviors, such as those induced by multiple timescales, are non-stationary. Learning non-stationary policies is challenging and typically requires sophisticated or inefficient algorithms. Motivated by the prevalence of this control problem in real-world complex systems, we introduce a simple framework for learning non-stationary policies for multi-timescale MARL. Our approach uses available information about agent timescales to define and learn periodic multi-agent policies. In detail, we theoretically demonstrate that the effects of non-stationarity introduced by multiple timescales can be learned by a periodic multi-agent policy. To learn such policies, we propose a policy gradient algorithm that parameterizes the actor and critic with phase-functioned neural networks, which provide an inductive bias for periodicity. The framework's ability to effectively learn multi-timescale policies is validated on a gridworld and building energy management environment.
Optimization of Distribution Feeder Topology: A Differential Programming Learning Approach
This paper presents a gradient based method for optimizing distribution feeder network topology under load un- certainty. We recast the optimal network reconfiguration problem as a learning problem where edge weights of a graph are learned to produce an optimized spanning tree for a distribution network. Using recent methods published on differentiable programming, we provide a data driven method for learning these weights. We test our method on 100 variations of an IEEE 15-bus test system. Our results show that our method outperforms more traditional mathematical programming-based approaches.
EvoProtGrad (Directed Evolution for Proteins with Gradients) [SWR-23-48]
A Python package for directed evolution on a protein sequence with gradient-based discrete Markov chain monte carlo (MCMC). Users are able to compose custom models that map sequence to function with pretrained models, including protein language models (PLMs), to guide and constrain search. Our package natively integrates with the HuggingFace platform and supports PLMs from transformers. Our MCMC sampler identifies promising amino acids to mutate via model gradients taken with respect to the input (i.e., sensitivity analysis). We allow users to compose their own custom target function for MCMC by leveraging the Product of Experts MCMC paradigm. Each model is an "expert" that contributes its own knowledge about the protein's fitness landscape to the overall target function. The sampler is designed to be more efficient and effective than brute force and random search while maintaining most of the generality and flexibility. Additional information can be found in the related publication: https://iopscience.iop.org/article/10.1088/2632-2153/accacd
Plug & play directed evolution of proteins with gradient-based discrete MCMC
Abstract A long-standing goal of machine-learning-based protein engineering is to accelerate the discovery of novel mutations that improve the function of a known protein. We introduce a sampling framework for evolving proteins in silico that supports mixing and matching a variety of unsupervised models, such as protein language models, and supervised models that predict protein function from sequence. By composing these models, we aim to improve our ability to evaluate unseen mutations and constrain search to regions of sequence space likely to contain functional proteins. Our framework achieves this without any model fine-tuning or re-training by constructing a product of experts distribution directly in discrete protein space. Instead of resorting to brute force search or random sampling, which is typical of classic directed evolution, we introduce a fast Markov chain Monte Carlo sampler that uses gradients to propose promising mutations. We conduct in silico directed evolution experiments on wide fitness landscapes and across a range of different pre-trained unsupervised models, including a 650 M parameter protein language model. Our results demonstrate an ability to efficiently discover variants with high evolutionary likelihood as well as estimated activity multiple mutations away from a wild type protein, suggesting our sampler provides a practical and effective new paradigm for machine-learning-based protein engineering.
PowerGridworld: A Framework for Multi-Agent Reinforcement Learning in Power Systems: Preprint
We present the PowerGridworld software package to provide users with a light-weight, modular, and customizable framework for creating power systems-focused, multi-agent gym environments that readily integrate with existing training frameworks for reinforcement learning (RL). While many frameworks exist for training multi-agent (MA) RL policies, none exist to rapidly prototype and develop the environments themselves, especially in the context of heterogeneous (composite, multi-device) power systems where power flow solutions are required to define grid-level variables and costs. PowerGridworld is an open-source software package that helps to fill this gap. To highlight PowerGridworld's key features, we present two case studies and demonstrate learning multi-agent RL policies using both OpenAI's MADDPG and RLLib's PPO algorithms where, in both cases, at least some subset of agents incorporate elements of the power flow solution at each time step as part of their reward (negative cost) structures.
graphenv: a Python library for reinforcement learning on graph search spaces
Many important and challenging problems in combinatorial optimization (CO) can be expressed as graph search problems, in which graph vertices represent full or partial solutions and edges represent decisions that connect them. Graph structure not only introduces strong relational inductive biases for learning (Battaglia et al., 2018) - in this context, by providing a way to explicitly model the value of transitioning (along edges) between one search state (vertex) and the next - but lends itself to problems both with and without clearly defined algebraic structure. For example, classic CO problems on graphs such as the Traveling Salesman Problem (TSP) can be expressed as either pure graph search or integer programs. Other problems, however, such as molecular optimization, do no have concise algebraic formulations and yet are readily implemented as a graph search (V. et al., 2022; Zhou et al., 2019). Such "model-free" problems constitute a large fraction of modern reinforcement learning (RL) research owing to the fact that it is often much easier to write a forward simulation that expresses all of the state transitions and rewards, than to write down the precise mathematical expression of the full optimization problem. In the case of molecular optimization, for example, one can use domain knowledge alongside existing software libraries to model the effect of adding a single bond or atom to an existing but incomplete molecule, and let the RL algorithm build a model of how good a given decision is by "experiencing" the simulated environment many times through. In contrast, a model-based mathematical formulation that fully expresses all the chemical and physical constraints is intractable. In recent years, RL has emerged as an effective paradigm for optimizing searches over graphs and led to state-of-the-art heuristics for games like Go and chess, as well as for classical CO problems such as the TSP. This combination of graph search and RL, while powerful, requires non-trivial software to execute, especially when combining advanced state representations such as Graph Neural Networks (GNN) with scalable RL algorithms.
Graph-Env
The Graph-Env library provides a framework for adapting graph search problems into OpenAI Gym environments for reinforcement learning. In other words, Graph-Env enables reinforcement learning algorithms to be applied to graph search problems. Graph search problems include molecule and crystal structure design problems; route planning problems, including traveling salesperson problem; puzzles and games, such as chess and go; shortest path problems; minimum spanning tree; vehicle path generation; and others.
LBC (Learning Building Control)
LBC encompasses the source code and data to reproduce and extend results for a manuscript that compares demand responsive control schemes for multi-zone buildings. It includes examples of model predictive control (MPC), value function approximation via CVXPYLAYERS, and differentiable predictive control (DPC). The goal of the study is to evaluate state-of-the-art controllers and establish the efficacy (if any) of learning-based approaches that leverage deep neural networks in one way or another.
Two-Stage Reinforcement Learning Policy Search for Grid-Interactive Building Control
This paper develops an intelligent grid-interactive building controller, which optimizes building operation during both normal hours and demand response (DR) events. To avoid costly on-demand computation and to adapt to non-linear building models, the controller utilizes reinforcement learning (RL) and makes real-time decisions based on a near-optimal control policy. Learning such a policy typically amounts to solving a hard non-convex optimization problem. We propose to address this problem with a novel global-local policy search method. In the first stage, an RL algorithm based on zero-order gradient estimation is leveraged to search for the optimal policy globally, due to its scalability and the potential to escape some poor performing local optima. The obtained policy is then fine-tuned locally to bring the first-stage solution closer to that of the original unsmoothed problem. Experiments on a simulated five-zone commercial building demonstrate the advantages of the proposed method over existing learning approaches. They also show that the learned control policy outperforms a pragmatic linear model predictive controller (MPC) and approaches the performance of an oracle MPC in testing scenarios. Using a state-of-the-art advanced computing system, we demonstrate that the controller can be learned and deployed within hours of training.
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.
A Cyber-Physical System for Freeway Ramp Meter Signal Control Using Deep Reinforcement Learning in a Connected Environment
Freeway bottlenecks such as on-ramp merging areas account for about 40% of recurring freeway congestion. It is generally agreed that building more roads and adding more lanes to existing infrastructure does not solve the congestion problem, and so dynamic traffic control measures offer a more cost-effective alternative. Ramp meters, traffic signal devices that regulate traffic flow entering freeways, are among the most effective measures to mitigate congestion at on-ramp merging areas on freeways. The confluence of deep reinforcement learning (RL) and connectivity provides a possible solution to advance ramp meter signal control. Deep RL is a group of machine-learning methods that enables an agent learning from the environment to improve its performance. In this study, three deep RL methods-proximal policy optimization (PPO), Ape-X deep Q-network (DQN), and asynchronous advantage actor-critic agents (A3C)-are explored for ramp meter signal control to maximize vehicle speed and traffic throughput, as well as to minimize energy consumption and emissions at freeway on-ramp merging areas in a connected environment. The low computational requirement and scalability of deep RL for deployment make it a powerful optimization tool for time-sensitive applications such as ramp meter signal control. The results of this study show that deep RL methods yield superior performance to both a fixed-time controller and ALINE A, a state-of-the-art feedback controller.
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.
Computational Math Problems for a Clean Energy Future
Cutting edge computational mathematics are ubiquitous in renewable energy research. Problems in resilient and reliable electric grid operations, infrastructure planning, wind farm yaw control, and more demand sophisticated and scalable computational tools that enable the transition of renewable energy technologies from proof of concept to deployment into our energy system. The mission of the Computational Science Center at NREL is to lead the lab's efforts to solve energy challenges using high-performance computing (HPC), computational science, applied mathematics, scientific data management, visualization, and informatics. In this poster, we provide a short overview of three areas of computational mathematics research at NREL: wind power scenario generation for stochastic grid operations and infrastructure planning, improved rational function approximations for electromagnetic transients codes, and wind farm yaw control using a combination of the Alternating Direction Method of Multipliers (ADMM) and reinforcement learning (RL). Increasing penetrations of renewable energy into power grids motivate the investigation of new approaches to characterizing uncertainty for five-minute economic dispatch problems. Similarly, as the penetration of distributed energy resources on power grids increases, it becomes important to revisit our methods of modelling transient phenomena, i.e. electromagnetic transients programs. Finally, the combination of ADMM and RL for wind farm yaw control presented here can potentially increase the efficiency of the deployed distributed controllers by orders of magnitude.
Learning-Accelerated ADMM for Distributed DC Optimal Power Flow
We propose a novel data-driven method to accelerate the convergence of Alternating Direction Method of Multipliers (ADMM) for solving distributed DC optimal power flow (DC-OPF) where lines are shared between independent network partitions. Using previous observations of ADMM trajectories for a given system under varying load, the method trains a recurrent neural network (RNN) to predict the converged values of dual and consensus variables. Given a new realization of system load, a small number of initial ADMM iterations is taken as input to infer the converged values and directly inject them into the iteration. We empirically demonstrate that the online injection of these values into the ADMM iteration accelerates convergence by a significant factor for partitioned 14-, 118- and 2848-bus test systems under differing load scenarios. The proposed method has several advantages: it maintains the security of private decision variables inherent in consensus ADMM; inference is fast and so may be used in online settings; RNN-generated predictions can dramatically improve time to convergence but, by construction, can never result in infeasible ADMM subproblems; it can be easily integrated into existing software implementations. While we focus on the ADMM formulation of distributed DC-OPF in this paper, the ideas presented are naturally extended to other distributed optimization problems.
Advanced Computing, Data Science, and Artificial Intelligence Research Opportunities for Energy-Focused Transportation Science
The Energy Efficient Mobility Systems (EEMS) technology landscape is complex and rapidly evolving, which provides both tremendous opportunities and formidable challenges. Significant alterations to the mobility landscape are underway due to the advent of vehicle and infrastructure connectivity, autonomous driving, and rapid passenger- and freight-vehicle electrification. Advanced computing will play an increasingly important role in enabling the EEMS program to understand and identify the most important levers to improve the energy productivity of future integrated mobility systems. It is also driving new approaches to mobility and the research to unlock an affordable, efficient, safe, and accessible transportation future. Driving much of this change is the collection, analysis, and strategic use of massive amounts of diverse, complex data from infrastructure and vehicles with on-board sensors and data storage and transmission capabilities. Diverse and representative data are key to implementing approaches to maximize mobility energy productivity. While high-fidelity modeling of integrated transportation networks has strengthened our understanding of dynamic movement and behavior patterns, existing tools must be expanded beyond their current focus. This work necessitates data infrastructure investments (e.g., secure-streaming data platforms driven by ubiquitous sensors and video analytics) as well as investments in critical capabilities for large-scale automated analysis and organization using modern machine learning, statistics, and artificial intelligence. Other chief needs include agile, large-scale storage that can be quickly searched and queried for relevant data to support validation and model development, data-sharing agreements, and formatting standards for key data types. The future of public transit must be explored in greater detail, research must inform design, and opportunities must be identified for improving the mobility productivity of public transit in both urban and rural America.
rlmolecule: A library for general-purpose material and molecular optimization using AlphaZero-style reinforcement learning
This library includes in-progress code for the optimization of materials and molecules using reinforcement learning. This project seeks to develop a generalized machine learning approach for optimizing targeted, and often complex, functional properties over the space of possible material structures, allowing faster identification of promising candidates. Electronic and transport properties of materials are determined by their molecular structure (for molecules) or their crystal structure (for in-organic crystalline materials). The atomic composition, bonding, and 3-D-spatial arrangement in these systems therefore represent a vast, discontinuous search space for optimal candidates. This library implements AlphaZero style reinforcement learning applied to two worked examples: the optimization of stable organic radicals for redox flow batteries, and the determination of stable electrolyte species for solid state batteries. The current code contains proof-of-concept results run on the Eagle HPC system, while the completed package will be a flexible tool for material optimization across of number of energy-relevant fields, which leverages cloud compute resources for the majority of the computational heavy lifting.
On the Computational Viability of Quantum Optimization for PMU Placement
Using optimal phasor measurement unit placement as a prototypical problem, we assess the computational viability of the current generation D-Wave Systems 2000Q quantum annealer for power systems design problems. We reformulate minimum dominating set for the annealer hardware, solve the reformulation for a standard set of IEEE test systems, and benchmark solution quality and time to solution against the CPLEX optimizer and simulated annealing. For some problem instances the 2000Q outpaces CPLEX. For instances where the 2000Q underperforms with respect to CPLEX and simulated annealing, we suggest hardware improvements for the next generation of quantum annealers.