19 February 2021 in npj Digital Medicine
The aim of this work was to develop and evaluate the reinforcement learning algorithm VentAI, which is able to suggest a dynamically optimized mechanical ventilation regime for critically-ill patients. We built, validated and tested its performance on 11,943 events of volume-controlled mechanical ventilation derived from 61,532 distinct ICU admissions and tested it on an independent, secondary dataset (200,859 ICU stays; 25,086 mechanical ventilation events). A patient “data fingerprint” of 44 features was extracted as multidimensional time series in 4-hour time steps. We used a Markov decision process, including a reward system and a Q-learning approach, to find the optimized settings for positive end-expiratory pressure (PEEP), fraction of inspired oxygen (FiO2) and ideal body weight-adjusted tidal volume (Vt). The observed outcome was in-hospital or 90-day mortality. VentAI reached a significantly increased estimated performance return of 83.3 (primary dataset) and 84.1 (secondary dataset) compared to physicians’ standard clinical care (51.1). The number of recommended action changes per mechanically ventilated patient constantly exceeded those of the clinicians. VentAI chose 202.9% more frequently ventilation regimes with lower Vt (5–7.5 mL/kg), but 50.8% less for regimes with higher Vt (7.5–10 mL/kg). VentAI recommended 29.3% more frequently PEEP levels of 5–7 cm H2O and 53.6% more frequently PEEP levels of 7–9 cmH2O. VentAI avoided high (>55%) FiO2 values (59.8% decrease), while preferring the range of 50–55% (140.3% increase). In conclusion, VentAI provides reproducible high performance by dynamically choosing an optimized, individualized ventilation strategy and thus might be of benefit for critically ill patients.
15 April 2019 in IEEE Transactions on Communications
In this paper, we study an unmanned aerial vehicle (UAV)-enabled wireless power transfer network, where a UAV flies at a constant altitude in the sky to provide wireless energy supply for a set of ground nodes with a linear topology. Our objective is to maximize the minimum received energy among all ground nodes by optimizing the UAV’s one-dimensional (1D) trajectory, subject to the maximum UAV flying speed constraint. Different from previous works that only provided heuristic and locally optimal solutions, this paper is the first time to present the globally optimal 1D UAV trajectory solution to the considered min-energy maximization problem. Towards this end, we first show that for any given speed-constrained UAV trajectory, we can always construct a maximum-speed trajectory and a speed-free trajectory, such that their combination can achieve the same received energy at all these ground nodes. Next, we transform the UAV-speed-constrained trajectory design problem into an equivalent UAV-speed-free problem, which is then optimally solved via the Lagrange dual method. The optimal 1D UAV trajectory solution follows the so-called successive hover-and-fly structure, i.e., the UAV successively hovers at a finite number of hovering points each for an optimized hovering duration, and flies among these hovering points at the maximum speed. Building upon the optimal UAV trajectory structure, we further present a low-complexity UAV trajectory design by first transforming the original problem into an equivalent non-convex problem with only the UAV hovering locations and durations as optimization variables, and then updating the trajectory via the successive convex approximation technique. Our analysis shows that the low-complexity design is guaranteed to converge to a suboptimal solution at a significantly lower complexity irrespective of the geographical network size. Numerical results show that the proposed low-complexity design actually achieves the same performance as the proposed optimal solution, and both of them outperform the benchmark algorithms in prior works under different scenarios.
20 October 2020 in IEEE Transactions on Wireless Communications
In this paper, we study an unmanned aerial vehicle (UAV)-enabled multiuser wireless power transfer (WPT) network, where a UAV is responsible for providing wireless energy for a set of ground devices (GDs) deployed in an area. We focus on the design of UAV trajectory subject to the maximum flight speed limit, in order to maximize the minimum harvested energy among GDs over a particular charging duration. Different from prior works that considered simplified linear energy harvesting models, this paper for the first time takes into account the realistic nonlinear energy harvesting model for the UAV trajectory design. However, the formulated trajectory design problem is highly non-convex and has infinite number of variables, thus making it be challenging to be solved optimally. To tackle this difficulty, we adopt the following three-step approach to obtain an efficient solution. First, we rigorously characterize that the optimal trajectory follows a new successive-hover-and-fly (SHF) structure, where the UAV hovers at a certain set of points for efficiently transferring energy, and flies among these hovering points with the maximum speed following certain arcs (not necessarily straight lines). Next, based on this SHF structure, we transform the original problem to a new one for finding a set of turning point variables during the maximum-speed flight, at which the UAV changes the flight direction without hovering. Finally, we use the techniques of convex approximation to solve the transformed problem. According to the convexity of the nonlinear energy harvesting model, we iteratively solve a series of convex optimization problems to update the UAV trajectory towards a high-quality solution. Numerical results show the convergence of the proposed approach, and validate its performance gain over conventional designs.
21 November 2022 in IEEE Transactions on Wireless Communications
We consider an unmanned aerial vehicle (UAV) assisting data collection from multiple sensor nodes (SNs). We provide a completion time minimization design via jointly deciding the UAV trajectory and the SN assignment scheme. In particular, we first characterize the fundamental features of the joint optimal solution to the formulated problem. On the one hand, the optimal UAV trajectory is proved following a successive-hover-fly (SHF) structure. Namely, in an optimal solution, the UAV successively visits multiple hovering points and performs hovering with designated duration, while the maximum speed is achieved during the whole flying period between each two hovering points. On the other hand, the optimal SN assignment is characterized to follow a segment-based scheme. Based on the two characterizations, we are motivated to implement SHF structure with turning points in trajectory design and reasonably assume each segment in SHF structure having constant SN assignment. Afterwards, we relax the binary constraints for SN assignments and establish a convex approximation for the reformulated problem, which enables an iterative algorithm. A suboptimal joint solution is obtained via iteratively optimizing the completion time. A realization strategy is also provided for the relaxed solution while assuring the completion of data collection tasks. Finally, the proposed solution is validated and evaluated through numerical results. Both a low complexity and an accurate task completion guarantee of our proposed solution are observed in comparison with the benchmarks.
27 March 2024 in IEEE Transactions on Vehicular Technology
Inthis paper, we investigate an unmanned aerial vehicle (UAV)-assisted edge computing network in a sophisticated three-dimensional (3D) scenario, e.g., a post-disaster urban area with multiple mobile ground terminals (GTs) for resilience. The mission of the UAV is to first collect a set of computation tasks from multiple GT sources, then process them and make joint decisions (computation results) which are finally transmitted to another set of GT destinations. An efficient design is provided aiming at minimizing the total time cost including costs for the task offloading, computations, and decision transmissions, via jointly designing the UAV's trajectory and the scheduling of communications from and to different GTs. In particular, we take into account the practical assumptions on the GTs' mobility, on the obstacle avoidance to 3D buildings, and on that the UAV is possible to fly among buildings, i.e., the flying altitude can be lower than the buildings' heights. To address the established complex and non-convex problem, we first provide a problem transformation by exploiting Markov decision process (MDP), and then propose a deep reinforcement learning (DRL) approach-based solution with a multi-step dueling DDQN (D3QN) method. As a result, an efficient solution is achieved, where the UAV acts as the agent to continuously explore and improve its mobile strategy during interacting with the environment under the proposed method. The simulation results show the advantages of the proposed 3D trajectory design in comparison to a 2D one. In addition, the results confirm the robustness of the proposed design in terms of different GTs' mobility models and UAV flying height/area limits, and illustrate the superiority of unrestricted feasible domains for the UAV flight.
14 September 2023 in IEEE Transactions on Artificial Intelligence
With the exponential growth of computational power and the availability of large-scale datasets in recent years, remarkable advancements have been made in the field of artificial intelligence (AI), leading to complex models and innovative applications. However, these models consume a significant unprecedented amount of energy, contributing to greenhouse gas emissions and a growing carbon footprint in the AI industry. In response, the concept of green AI has emerged, prioritizing energy efficiency and sustainability alongside accuracy and related measures. To this end, data-centric approaches are very promising to reduce the energy consumption of AI algorithms. This paper presents a comprehensive overview of data-centric technologies and their impact on the energy efficiency of AI algorithms. Specifically, it focuses on methods that utilize training data in an efficient manner to improve the energy efficiency of AI algorithms. We have identified multiple data-centric approaches, such as active learning, knowledge transfer/sharing, dataset distillation, data augmentation, and curriculum learning that can contribute to the development of environmentally-friendly implementations of machine learning algorithms. Finally, the practical applications of these approaches are highlighted, and the challenges and future directions in the field are discussed.