Convergence Difficulties Research Articles

Bridging the Gap Between Cumbersome and Light Detectors via Layer-Calibration and Task-Disentangle Distillation in Remote Sensing Imagery

With urgent application requirements such as satellite in-orbit processing and unmanned aerial vehicle tracking, knowledge distillation (KD) following the teacher-student teaching mechanism has shown great potential to obtain lightweight detectors. However, compact students have limited accuracy due to the interference of large scale variations and blurred boundaries in remote sensing objects. Specifically, previous methods mostly force teacher-student responses from the layer of same depth and scale to align. Stereotyped manual inter-layer associations may cause discriminative features of multi-scale objects to be incorrectly bundled. Further, the regression branch follows the identical distillation paradigm as the classification branch, resulting in ambiguous object bounding box deviations. To solve the above two issues, we propose an effective KD framework called layer-calibration and task-disentangle distillation (LTD). First, cross-layer calibration distillation (CCD) structure is innovatively proposed. It adaptively binds a student layer with several related target layers, rather than a fixed layer in the teacher model. Appropriate and clear knowledge of large and small objects is transmitted. Since the CCD structure requires explicit global inner product computation between multiple layers, local implicit calibration (LIC) module is further proposed to reduce distilled convergence difficulty. Second, task-aware spatial disentangle distillation (TASD) structure is devised to transfer task-decoupled semantics and localization knowledge in a divide-and-conquer manner, alleviating objects localization imprecision. Experiments demonstrate that our LTD achieves state-of-the-art performance on several datasets, and is a plug-and-play approach to most detectors. The code will be available soon.

GAIL-PT: An intelligent penetration testing framework with generative adversarial imitation learning

Penetration testing (PT) is an efficient tool for network testing and vulnerability mining by simulating the hackers’ attacks to obtain valuable information applied in operating and database systems. Most of the traditional manual solutions are strongly relying on the domain knowledge of human experts with high penetration costs. Therefore, solutions based on the artificial intelligent algorithm such as reinforcement learning (RL) and deep reinforcement learning (DRL), with less time-consuming and lower labor costs, become a great solution to address the challenge. However, there are still a few challenges for RL/DRL-based PT in real penetration scenarios, such as the large dimension size of the agent’s discrete action space usually causing difficulties in convergence. To address the above issue, this paper proposes a novel framework named Generative Adversarial Imitation Learning based intelligent Penetration Testing (GAIL-PT), which utilizes expert knowledge base and GAIL network to guide the policy generation of RL/DRL agents with lower costs. Specifically, we first construct the expert knowledge bases by collecting state-action pairs from the successful exploitations of pre-trained RL/DRL models. Secondly, we feed the expert knowledge bases generated by different RL/DRL models online into the discriminator of GAIL-PT to guide its training process. Besides, we integrate the losses of the generator and the discriminator in GAIL-PT to optimize the overall objective and use the discriminator’s discounted rewards for policy generation. The extensive experiments conducted on the practical target hosts and simulated network scenarios demonstrate that GAIL-PT achieves outstanding performance, and outperforms the state-of-art method DeepExploit in exploiting Metasploitable2 and Q-learning in different scale networks. It also verified that GAIL-PT is a general leading framework suitable for RL/DRL-based methods. The code of GAIL-PT is open-sourced at https://github.com/Shulong98/GAIL-PT//.

An Indoor UWB 3D Positioning Method for Coplanar Base Stations.

As an indispensable type of information, location data are used in various industries. Ultrawideband (UWB) technology has been used for indoor location estimation due to its excellent ranging performance. However, the accuracy of the location estimation results is heavily affected by the deployment of base stations; in particular, the base station deployment space is limited in certain scenarios. In underground mines, base stations must be placed on the roof to ensure signal coverage, which is almost coplanar in nature. Existing indoor positioning solutions suffer from both difficulties in the correct convergence of results and poor positioning accuracy under coplanar base-station conditions. To correctly estimate position in coplanar base-station scenarios, this paper proposes a novel iterative method. Based on the Newton iteration method, a selection range for the initial value and iterative convergence control conditions were derived to improve the convergence performance of the algorithm. In this paper, we mathematically analyze the impact of the localization solution for coplanar base stations and derive the expression for the localization accuracy performance. The proposed method demonstrated a positioning accuracy of 5 cm in the experimental campaign for the comparative analysis, with the multi-epoch observation results being stable within 10 cm. Furthermore, it was found that, when base stations are coplanar, the test point accuracy can be improved by an average of 63.54% compared to the conventional positioning algorithm. In the base-station coplanar deployment scenario, the upper bound of the CDF convergence in the proposed method outperformed the conventional positioning algorithm by about 30%.

Clothes retrieval based on ResNet and cluster triplet loss

Clothes image retrieval is a task that retrieves exact or very similar clothes from a large clothes image gallery according to a given query image. The high deformability of clothes and similar designs make clothes retrieval very difficult. Convolutional neural networks perform well in the feature extraction field, but cannot extract sufficiently fine-grained features for clothes retrieval tasks. Therefore, a model based on ResNet was designed in this article, with several improvements made to it to ensure better feature discrimination. On the other hand, pair-based loss functions that are commonly used in image retrieval suffer from insufficient discrimination ability, difficulty in convergence, and time-consuming during training. Cluster triplet loss function is proposed in this article, which reduces the computational complexity in the training phase and enables the model to have a certain ability to resist noise labels. Experiments are completed on In-shop clothes retrieval benchmark and Consumer-to-shop clothes retrieval benchmark in Deepfashion, and stanford online products (SOP) to supplement. Our method can have Recall@1 improvements of 1.3–1.7% on all three datasets, surpassing the latest state-of-the-art methods.

Self-Supervised Action Representation Learning Based on Asymmetric Skeleton Data Augmentation.

Contrastive learning has received increasing attention in the field of skeleton-based action representations in recent years. Most contrastive learning methods use simple augmentation strategies to construct pairs of positive samples. When using such pairs of positive samples to learn action representations, deeper feature information cannot be learned, thus affecting the performance of downstream tasks. To solve the problem of insufficient learning ability, we propose an asymmetric data augmentation strategy and attempt to apply it to the training of 3D skeleton-based action representations. First, we carefully study the different characteristics presented by different skeleton views and choose a specific augmentation method for a certain view. Second, specific augmentation methods are incorporated into the left and right branches of the asymmetric data augmentation pipeline to increase the convergence difficulty of the contrastive learning task, thereby significantly improving the quality of the learned action representations. Finally, since many methods directly act on the joint view, the augmented samples are quite different from the original samples. We use random probability activation to transform the joint view to avoid extreme augmentation of the joint view. Extensive experiments on NTU RGB + D datasets show that our method is effective.

Optimal Power Flow Technique for Distribution System Considering Distributed Energy Resources (DER)

Modern electric power systems consist of large-scale, highly complex interconnected systems projected to match the intense demand growth for electrical energy. This involves the decision of generation, transmission, and distribution of resources at different time horizons. They also face challenges in incorporating new forms of generation, distributed generations, which are located close to consumer centers, and new loads such as electric vehicles. Traditionally, the nonlinear Newton–Raphson optimization method is used to support operational decisions in such systems, known as Optimal Power Flow (OPF). Although OPF is one of the most practically important and well-researched sub-fields of constrained nonlinear optimization and has a rich history of research, it faces the convergence difficulties associated with all problems represented using non-linear power flow constraints. The proposal is to present an approach in a software component in cloud Application Programming Interface (API) format, with alternative modeling of the electrical optimization problem as a non-linear objective function and representing electric network constraints modeled through both current and voltage Kirchhoff linear equations. This representation overcomes the non-linearity of the OPF problem considering Distributed Energy Resources (DER). The robustness, scalability, and availability of the method are tested on the IEEE-34 bus system with several modifications to accommodate the DER testing under conditions and in radial or meshed distribution systems under different load scenarios.

A generalized CFD model for evaluating catalytic separation process in structured porous materials

A hybrid multiphase model is developed to simulate the simultaneous momentum, heat and mass transfer and heterogeneous catalyzed reaction in structured catalytic porous materials. The approach relies on the combination of the volume of fluid (VOF) and Eulerian–Eulerian models, and several plug-in field functions. The VOF method is used to capture the gas–liquid interface motion, and the Eulerian–Eulerian framework solves the temperature and chemical species concentration equations for each phase. The self-defined field functions utilize a single-domain approach to overcome convergence difficulty when applying the hybrid multiphase for a multi-domain problem. The method is then applied to investigate selective removal of specific species in multicomponent reactive evaporation process. The results show that the coupling of catalytic reaction and interface species mass transfer at the phase interface is conditional, and the coupling of catalytic reaction and momentum transfer across fluid–porous interface significantly affects the conversion rate of reactants. Based on the numerical results, a strategy is proposed for matching solid catalyst with operating condition in catalytic distillation application.

Comparative studies of EV fleet smart charging approaches for demand response in solar-powered building communities

The use of electric vehicles (EVs) has been on the rise during the past decade, and the number is expected to rapidly increase in the future. At aggregated level, the large EV charging loads, if not well regulated, will cause great stress on the existing grid infrastructures. On the other hand, considered as a resource-efficient and cost-effective demand response resource, EV fleet smart charging control methods have been developed and applied to mitigate power issues of the grid while avoiding expensive upgrade of power grid infrastructure. Until now, there is no systematic study on how different coordination mechanisms affecting the EV fleet's charging demand response performance. Thus, it is still unclear which one may perform better in the increasingly common solar-powered building communities, especially as demand response is increasingly concerned. Aiming to fill in such knowledge gaps, this study conducted systematic comparative studies of three representative control methods selected from the non-coordinated, bottom-up coordinated, and top-down coordinated control categories. Their power regulation performances have been comparatively investigated in two perspectives: minimizing peak power exchanges with the grid and maximizing PV self-utilization, based on a real building community in Sweden. Meanwhile, their computational performances have also been investigated. The study results show that due to the ability to schedule and coordinate all the EVs simultaneously, the top-down coordinated control is superior to the other two control methods in the considered demand response performances. Note that its better performance is realized with a higher computational load, leading to possible convergence difficulties in practice. The study results will help improve understanding of how coordination affect the EV smart charging control performances. It will pave the way for developments of more sophisticated control methods for EV smart charging in more complex scenarios.

Numerical solution of a singularly perturbed boundary value problem of supersonic flow transformed to the modified best argument

When solving problems of aerodynamics, researchers often need to numerically solve singularly perturbed boundary value problems. In some cases, the problem can be reduced to solving a boundary value problem for an ordinary differential equation. Then it is possible to apply various numerical methods such as the grid method, the shooting method, as well as a number of projection methods, which, in turn, can form the basis of the finite element method. The grid method requires solving a system of algebraic equations, that are often nonlinear, which leads to an increase in the calculation time and to the difficulties in convergence of the approximate solution. According to the shooting method, the solution of boundary value problem is reduced to solving a certain set of Cauchy problems. When solving stiff Cauchy problems, implicit schemes are used as a rule, but in this case the same difficulties arise as for the grid method. The transformation of the problem to the best argument λ, calculated tangentially along the integral curve, makes it possible to increase the efficiency of explicit numerical methods. However, in cases where the growth rate of integral curves is close to exponential, the transformation to the best argument is not efficient enough. Then the best argument is modified in such a way as to smooth out this flaw. This paper investigates the application of modified best argument to the solution of the boundary value problem of an aerodynamic flow movement in case when the gas is injected at supersonic speed into a channel of variable cross-section.

A simple arithmetic mean to ensure the convergence of models based on Blade Element Momentum theory applied to Vertical Axis Wind Turbines simulations

The computer models based on the Blade Element Momentum (BEM) theory are able to predict the aerodynamic performance of wind turbines with good accuracy and efficiency. However, these methods may present some convergence difficulties. This work presents the development of a methodology adopting arithmetic mean in situations of flow instability to ensure the convergence of BEM models applied to Vertical Axis Wind Turbines (VAWT) simulations. This methodology was used in the creation of a program written in Fortran, called FASTEEVSIM. Tests were conducted in a Darrieus wind turbine using NACA0018 and DU06W200 airfoils as blade profiles. A comparison of the results using the FASTEEVSIM with experimental data from well-known literature sources has been done in order to validate the method. When comparing the power coefficients obtained for two types of blade profiles, one symmetrical (NACA0018) and another asymmetrical (DU06W200), the final validationresults showed a mean square deviation of 3.62% and 6.09%, respectively. The results showed that the developed methodology can assist the convergence effectively and that the FASTEEVSIM program is able to predict the performance of VAWT, providing data close to experimental values in a simple and fast way.

A feasible path-based branch and bound algorithm for strongly nonconvex MINLP problems

In this paper, a feasible path-based branch and bound (B&B) algorithm is proposed to solve mixed-integer nonlinear programming problems with highly nonconvex nature through integration of the previously proposed hybrid feasible-path optimisation algorithm and the branch and bound method. The main advantage of this novel algorithm is that our previously proposed hybrid steady-state and time-relaxation-based optimisation algorithm is employed to solve a nonlinear programming (NLP) subproblem at each node during B&B. The solution from a parent node in B&B is used to initialize the NLP subproblems at the child nodes to improve computational efficiency. This approach allows circumventing complex initialisation procedure and overcoming difficulties in convergence of process simulation. The capability of the proposed algorithm is illustrated by several process synthesis and intensification problems using rigorous models.

Optimization of dividing wall columns based on online Kriging model and improved particle swarm optimization algorithm

Dividing wall columns (DWCs) can effectively improve the thermodynamic efficiency of traditional distillation columns. However, DWCs have intricate structures and strong internal interactions. Numerous structural and operational variables are interrelated. This work presents an improved cellular particle swarm optimization based on the online Kriging model (KCPSO) algorithm, and applies it to the optimization of DWC with the objective of minimizing the total annual cost. The algorithm uses the information of particle swarm search to act on the online Kriging model, and reacts on the particle search through the information of the online Kriging model. Calculations demonstrate that the KCPSO algorithm is superior to standard particle swarm optimization (PSO) and cellular PSO (CPSO) algorithms due to its higher quality of iteration. The KCPSO algorithm can effectively overcome the difficulty of early convergence of the CPSO algorithm and the problem that the PSO algorithm is prone to falling into local optimal solutions.

Influence-aware memory architectures for deep reinforcement learning in POMDPs

AbstractDue to its perceptual limitations, an agent may have too little information about the environment to act optimally. In such cases, it is important to keep track of the action-observation history to uncover hidden state information. Recent deep reinforcement learning methods use recurrent neural networks (RNN) to memorize past observations. However, these models are expensive to train and have convergence difficulties, especially when dealing with high dimensional data. In this paper, we propose influence-aware memory, a theoretically inspired memory architecture that alleviates the training difficulties by restricting the input of the recurrent layers to those variables that influence the hidden state information. Moreover, as opposed to standard RNNs, in which every piece of information used for estimating Q values is inevitably fed back into the network for the next prediction, our model allows information to flow without being necessarily stored in the RNN’s internal memory. Results indicate that, by letting the recurrent layers focus on a small fraction of the observation variables while processing the rest of the information with a feedforward neural network, we can outperform standard recurrent architectures both in training speed and policy performance. This approach also reduces runtime and obtains better scores than methods that stack multiple observations to remove partial observability.

A Novel Vehicle Destination Prediction Model With Expandable Features Using Attention Mechanism and Variational Autoencoder

The daily passage of vehicles generates a huge amount of location-aware social data, which provides a rich source of data for analyzing vehicle travel behavior. Being able to accurately predict the future destinations of vehicle travel has great economic value and social impact. The presence of larger sparsity, fewer features and error information in the real dataset led to difficulties in convergence of previous models. Therefore, we propose a Novel Vehicle Destination Prediction Model with Expandable Features Using Attention Mechanism and Variational Autoencoder (EFAMVA). The EFAMVA model combines the autoencoder model and the attention mechanism has overcome the above mentioned problems. The variational autoencoder model obtains the hidden features conforming to the characteristics of the data from the structured vehicle driving data. And the attention mechanism can learn the appropriate combination of weight parameters. The comprehensive experimental results with other comparison models show that the EFAMVA model achieved the best index score, with the MSE value of 0.750, the RMSE value of 1.215, and the MAE value of 0.955. Therefore, it can be shown that the EFAMVA model has a better predictive effect on the future destination of the vehicle.

A Hybrid Chebyshev Approximation Technique and Subdomain AIM Method for Wideband RCS Computation

An efficient hybrid method based on subdomain adaptive integral method (SAIM) and Chebyshev approximation technology (CAT) is proposed for analyzing the wideband radar cross section (RCS) of complex multiscale objects in this letter. The SAIM cannot only be used to effectively deal with the difficulty of slow convergence caused by the mutiscale problems, but also greatly reducing the computational memory. By introducing the CAT into SAIM, we can use the currents at a few frequency sampling points to predict the currents at any frequency point over the entire band. In addition, the application of the Maehly approximation can obtain a better accuracy. Numerical examples show that the proposed SAIM-CAT can significantly improve the calculation efficiency without loss of accuracy.

Using a new coefficient conjugate gradient method for solving unconstrained optimization problems

The conjugate gradient technique is a numerical solution strategy for finding minimization in mathematics. We present a simple, straightforward, efficient, and resilient conjugate gradient technique in this study. To address the convergence difficulty and descent property, the new technique is built on the quadratic model. Under some assumptions, the new improved approach meets the convergence characteristics and the adequate descent criterion. The suggested unique strategy is substantially more efficient than the classic FR method, according to our numerical analysis. The number of function evaluations, iterations and restarts are all included in the numerical results. The computational efficiency of the proposed approach is proved by comparative results.

Full-waveform inversion with adversarial losses via deep learning

Full-waveform inversion (FWI) is a powerful technology for determining geophysical properties, nevertheless, with high computational cost and difficulty in convergence. Motivated by the powerful tools available in modern deep learning system to deal with complex non-linear inverse problems, we implement the FWI workflow with deep neural networks and propose the use of two different neural networks with adversarial losses to provide fast and accurate solutions. Specifically, a deep neural network is used to generate simulated seismic data based on the finite-difference method, and another fully connected neural network is used to fit the Wasserstein distance between simulated and observed seismic data. The Wasserstein distance and the L2 metric together form an adversarial loss function. The parallel-layer model and the SEAM model are used to verify the benefits of the proposed method and its performance in the face of seismic data lack low-frequency components and noise interference. Numerical experiments demonstrate that, compared with other algorithms, the adaptive gradient (Adagrad) algorithm performs best. In addition, the Adagrad algorithm with adversarial losses has faster convergence speed and more inverting details than with L2 losses. The proposed method provides reliable inversion results, even when seismic data without low-frequency components. Moreover, the adversarial losses are less sensitive to noise, although still difficult to converge.

Effect of exchange-correlation functionals on the estimation of migration barriers in battery materials

Facile ionic mobility within host frameworks is crucial to the design of high-energy-density batteries with high-power-densities, where the migration barrier (Em) is the governing factor. Here, we assess the accuracy and computational performance of generalized gradient approximation (GGA), the strongly constrained and appropriately normed (SCAN), and their Hubbard U corrections, GGA+U and SCAN+U, within the density functional theory-nudged elastic band framework, in the prediction of Em as benchmarked against experimental data. Importantly, we observe SCAN to be more accurate than other frameworks, on average, albeit with higher computational costs and convergence difficulties, while GGA is a feasible choice for “quick” and “qualitative” Em predictions. Further, we quantify the sensitivity of Em with adding uniform background charge and/or the climbing image approximation in solid electrolytes, and the Hubbard U correction in electrodes. Our findings will improve the quality of Em predictions which will enable identifying better materials for energy storage applications.

A Zero-injection Initialization Numerical Method for Ill-conditioned Power Flow

Regarding the initial value sensitivity issues commonly found in Newton-like power flow calculation algorithms, a zero-injection initialization numerical method for ill-conditioned power flow is proposed in this paper, which improves the convergence of power flow calculation in ill-conditioned power systems. Based on the initial power flow provided by this method, the unbalanced active power of each bus (the slack bus excluded) relates to the net injection of active power of that bus and the unbalanced reactive power of each PQ bus relates to the net injection of reactive power of that bus, thus avoiding significant power unbalance. The method put forward herein can provide a reasonable initial value for the power flow calculation so that difficulty in power flow convergence caused by the unreasonable initial power flow is effectively avoided. The Newton’s method and the optimal multiplier method are implemented respectively on a small-scale ill-conditioned test system to verify the feasibility of the proposed zero-injection initialization numerical method.

Deep Reinforcement Learning for Intelligent Dual-UAV Reconnaissance Mission Planning

The reconnaissance of high-value targets is prerequisite for effective operations. The recent appreciation of deep reinforcement learning (DRL) arises from its success in navigation problems, but due to the competitiveness and complexity of the military field, the applications of DRL in the military field are still unsatisfactory. In this paper, an end-to-end DRL-based intelligent reconnaissance mission planning is proposed for dual unmanned aerial vehicle (dual UAV) cooperative reconnaissance missions under high-threat and dense situations. Comprehensive consideration is given to specific mission properties and parameter requirements through the whole modelling. Firstly, the reconnaissance mission is described as a Markov decision process (MDP), and the mission planning model based on DRL is established. Secondly, the environment and UAV motion parameters are standardized to input the neural network, aiming to deduce the difficulty of algorithm convergence. According to the concrete requirements of non-reconnaissance by radars, dual-UAV cooperation and wandering reconnaissance in the mission, four reward functions with weights are designed to enhance agent understanding to the mission. To avoid sparse reward, the clip function is used to control the reward value range. Finally, considering the continuous action space of reconnaissance mission planning, the widely applicable proximal policy optimization (PPO) algorithm is used in this paper. The simulation is carried out by combining offline training and online planning. By changing the location and number of ground detection areas, from 1 to 4, the model with PPO can maintain 20% of reconnaissance proportion and a 90% mission complete rate and help the reconnaissance UAV to complete efficient path planning. It can adapt to unknown continuous high-dimensional environmental changes, is generalizable, and reflects strong intelligent planning performance.