Search

To reduce the number of suspended converters and failure rates for the two degree-of-freedom (DOF) nacelle suspension, a passive pitching stabilizer (PPS) is designed for the Maglev wind yaw system (MWYS) to passively balance the pitching moment. Two DOF suspension modelling is given based on the damping force and restoring torque deriving from the PPS, which is optimized to minimize pitching vibration energy and power loss, as validated by simulation results. However, higher-frequency vibrations of the pitching angle during the convergence process can still influence the axial stability. Coupled with external time-varying disturbances, an adaptive prescribed performance control (PPC)-based sliding mode control (SMC) with current constraints is proposed to further improve transient performance and disturbance rejection capability. Adaptive techniques are also employed to estimate system parameter perturbations and external disturbances. The Lyapunov stability theorem is applied to prove the stability and convergence of the proposed strategy, and all closed-loop signals and adaptive parameters can converge to the desired neighborhood domain. Experimental results validate the optimized PPS and the proposed axial suspension strategy achieves the suspension objective with minimal tracking error and the effective pitching suppression.

Implantable medical devices (IMDs) in medical sciences have provided a quantum leap in network transformation. The communication network with IMDs typically has a wireless radio frequency (RF) telemetry or wired connection. IMDs, being devices, have more computing, communication capabilities and decision-making. Furthermore, these devices are being used to improve patients’ quality of life by medicating various chronic diseases. The captured data is stored in a medical server through a controller node. Our work focuses on wireless communication, so sensitive patient data over a public channel might be tampered with or eavesdropped by unauthorised access. Furthermore, the leakage of health data and malfunctioning of IMDs are vital in constructing cryptographic protocols, particularly in the design of remote user authentication. In this paper, we proposed a novel secure remote user authentication scheme using a lightweight consortium blockchain for the communication network with IMDs.

The photonics-assisted millimeter-wave (MMW) communication technology is attractive to facilitate the MMW application in the upcoming B5G and 6G networks. However, its generated MMW signal by optical heterodyne detection usually suffers from serious laser phase noise, which will severely deteriorate the system performance. In this paper, based on a single dual-drive Mach-Zehnder modulator and a single-end photodetector, we first present a simple and spectrally efficient hybrid fiber-wireless double-sideband transmission by employing an overlapping frequency multiplexing scheme. That is, two independent wireless signals with an identical carrier frequency can be simultaneously transmitted in the hybrid fiber-wireless links. To recover the above two overlapped signals, and cancel the concomitant laser phase noise at the same time, a novel digital signal processing method for carrier extraction and signal recovery is further proposed. A proof-of-concept experiment using two independent 3-GBd quadrature phase shift keying (QPSK) signals at W band (92.5 GHz) is performed. After up to 80-km fiber and 3-m wireless transmission, the two QPSK signals can be successfully demodulated, without using the traditional carrier phase estimation algorithm. The proposed scheme not only can double the spectral efficiency of conventional double sideband transmission scheme, but also is immune to its power fading phenomenon induced by chromatic dispersion and robust to the laser phase noise resulting from two free-running lasers in the photonics-assisted MMW communication link.

This paper addresses the problem of high communication frequency in the multi-rotor unmanned aerial vehicles swarm system. A novel sliding mode control method based on event-triggering is proposed, using a dual-loop structure control model based on sliding mode control of multi-rotor UAVs swarm. The model consists of a position subsystem and an attitude subsystem, where the position subsystem serves as the outer loop control object, and the attitude subsystem serves as the inner loop auxiliary control object. In the position subsystem, a sliding mode controller with event triggers is used to achieve position control of the multi-rotor UAVs swarm by introducing sliding mode surfaces and sliding mode variables. The stability of the system under this control strategy is proved by the Lyapunov stability theorem, and the absence of the Zero phenomenon is also proved. Simulation results demonstrate that the proposed control scheme reduces the reliance on high-frequency sampling of multi-rotor UAVs communication. This research provides a useful reference for the application of multi-rotor UAVs swarms in collaborative path following.

Camouflaged objects are typically assimilated into their surroundings. Consequently, in contrast to generic object detection/segmentation, camouflaged object detection proves to be considerably more intricate due to the indistinct boundaries and heightened intrinsic similarities between foreground targets and the surrounding environment. Despite the proposition of numerous algorithms that have demonstrated commendable performance across various scenarios, these approaches may still grapple with blurred boundaries, leading to the inadvertent omission of camouflaged targets in challenging scenes. In this paper, we introduce a multi-stage framework tailored for segmenting camouflaged objects through a process of coarse-to-fine refinement. Specifically, our network encompasses three distinct decoders, each fulfilling a unique role in the model. In the initial decoder, we introduce the Bi-directional Locating Module to excavate foreground and background cues, enhancing target localization. The second decoder focuses on leveraging boundary information to augment overall performance, incorporating the Multi-level Feature Fusion Module to generate prediction maps with finer boundaries. Subsequently, the third decoder introduces the Mask-guided Fusion Module, designed to process high-resolution features under the guidance of the second decoder’s results. This approach enables the preservation of structural details and the generation of fine-grained prediction maps. Through the integration of the three decoders, our model effectively identifies and segments camouflaged targets. Extensive experiments are conducted on three commonly used benchmark datasets. The results of these experiments demonstrate that, even without the application of pre-processing or post-processing techniques, our model outperforms 14 state-of-the-art algorithms.

The development of communication and radar fusion systems is critical to enhancing sensor network security. However, in spectrum resource environments, the technology for processing fusion signals may face limitations due to multi-user access and interference issues. Traditional signal processing algorithms may also be influenced by waveform information. As a result, its implementation is limited in low-noise-ratio settings. Based on the advantages of multiple-input multiple-output technology in the field of communication and radar, in order to better ensure the non-interference between the communication and radar systems, the study introduces the orthogonal theory. It proposes a target detection and parameter estimation method based on orthogonal frequency division multiplexing (OFDM) technology in the signal modulation domain. It utilizes sparse tensor theory to obtain low rank features of signals, transform the dimensions of target parameter problems, and solve tensor rank 1 approximation problems. This can better improve the performance of target parameter estimation. The simulation analysis of the proposed method shows that the average error values of the improved decomposition method in Doppler frequency shift estimation and angle estimation are close to 10–2 and 10–1 respectively. The system runs for less than 5 seconds, and its average target detection accuracy exceeds 96%. The proposed orthogonal frequency division multiplexing technique can provide technical support for sensor networks in wireless communication and IOT networks, efficiently reducing computational complexity and interference among users. This can serve as a valuable reference for enhancing the security of communication radar fusion systems.

Due to the problem of a small amount of EEG samples and relatively high dimensionality of electroencephalogram (EEG) features, feature selection plays an essential role in EEG-based emotion recognition. However, current EEG-based emotion recognition studies utilize a problem transformation approach to transform multi-dimension emotional labels into single-dimension labels, and then implement commonly used single-label feature selection methods to search feature subsets, which ignores the relations between different emotional dimensions. To tackle the problem, we propose an efficient EEG feature selection method for multi-dimension emotion recognition (EFSMDER) via local and global label relevance. First, to capture the local label correlations, EFSMDER implements orthogonal regression to map the original EEG feature space into a low-dimension space. Then, it employs the global label correlations in the original multi-dimension emotional label space to effectively construct the label information in the low-dimension space. With the aid of local and global relevance information, EFSMDER can conduct representational EEG feature subset selection. Three EEG emotional databases with multi-dimension emotional labels were used for performance comparison between EFSMDER and fourteen state-of-the-art methods, and the EFSMDER method achieves the best multi-dimension classification accuracies of 86.43, 84.80, and 97.86 percent on the DREAMER, DEAP, and HDED datasets, respectively.

DC corona audible noise (AN) is one of the electromagnetic interference from power grids and the key factors affecting the electrical design of power transmission and distribution equipment. The measurement method is crucial to study the generation mechanism of audible noise. This paper proposed a Schlieren optical measurement with a corresponding numerical reconstruction method to measure the audible noise within DC corona discharge and quantified it with the pressure field. Numerical experiments are conducted to verify the accuracy of the pressure field re-construction method. The relationship between the pressure field and sound pressure during DC corona discharge is experimentally studied. Results show that the reconstruction error is within 5%; the Schlieren system captures the gas expansion process and pressure field in the discharge region at the microsecond timescale. During corona discharge, the radial radius of the discharge region increases while its pressure decreases and overpressure increases. A relationship of non-linear proportional increase appears between the pressure field and audible noise, and only a small portion of the pressure field is converted into audible noise. The new method demonstrated in this paper is an effective way of measuring and analyzing the pressure field in the DC corona discharge, which would be the key to understanding the mechanism of acoustic wave (or AN) generation from DC corona discharge on transmission lines.

The paper studies the problems of developing interconnected models of aggregates of complex chemical-technological systems (CTS) in conditions of scarcity and fuzziness of the initial information. Since the known methods of model development do not allow for solving these problems, we propose a systematic method that allows the development of a package of models of interconnected aggregates of complex CTS using the available information of a different nature. In the proposed method, formal and informal methods of system analysis are jointly used. Due to synergy and emergence, the proposed system allows the synthesis of the most adequate and effective CTS models using the knowledge, experience, and intuition of experts and other available data. Using the method, we have developed a package of models of interconnected reactors and used it in a case study for a reforming reactor and furnace of the Atyrau refinery catalytic reformer based on available statistical and fuzzy information. To study and optimize the operating modes of the reformer unit, a system of computer simulation and optimization of the reformer unit was created. Comparison of the obtained results of computer modeling and optimization with the results of known deterministic models shows the advantages of the proposed approach in the face of scarcity and fuzziness of the initial information. The importance and novelty of the proposed method lie in the possibility of developing a package of models of interrelated CTS units based on various types of available information. This allows one to systematically simulate and optimize the operating modes of the CTS. The practical significance of the results is that this case study can be successfully applied in the development of models of various technological units for oil refining, petrochemistry, and other industries.

The Industrial Internet of Things (IIoT) collects a large amount of data through various types of sensors and intelligently processes this data using cloud computing, which is flexible, efficient, and cost-effective. Since IIoT data is stored on the cloud service provider’s server, the data must be encrypted to protect the user’s privacy. However, the encrypted data faces the search problem, which is usually solved by Public Key Encryption with Keyword Search (PEKS). In addition, most existing PEKS schemes are vulnerable to Inside Keyword Guessing Attacks (IKGA). Recently, some certificateless public key authenticated encryption with keyword search (CLPEKS) schemes have been proposed, which not only avoid the problems of certificate management and key escrow but can also resist IKGA. However, most of them rely on the expensive bilinear pairing. To overcome these problems, in this paper we propose a pairing-free CLPEKS scheme. The security of the proposed scheme is proved in the random oracle model. The analysis results show that the proposed scheme has better overall performance in terms of computational cost, communication cost and security properties.

Gait movement is an important activity in daily human life. The coordination of gait movement is directly affected by the cooperation and functional connectivity between muscles. However, the mechanisms of muscle operation at different gait speeds remain unclear. Therefore, this study addressed the gait speed effect on the changes in cooperative modules and functional connectivity between muscles. To this end, surface electromyography (sEMG) signals were collected from eight key lower extremity muscles of twelve healthy subjects walking on a treadmill at high, middle, and low motion speeds. Nonnegative matrix factorization (NNMF) was applied to the sEMG envelope and intermuscular coherence matrix, yielding five muscle synergies. Muscle functional networks were constructed by decomposing the intermuscular coherence matrix, revealing different layers of functional muscle networks across frequencies. In addition, the coupling strength between cooperative muscles grew with gait speed. Different coordination patterns among muscles with changes in gait speed related to the neuromuscular system regulation were identified.

Nonlinear frequency division multiplexing (NFDM) system is an optional candidate to overcome the fiber nonlinearity limit. A full-spectrum modulated NFDM system, modulating data on combined continuous spectrum (CS) and discrete spectrum (DS) together, was proposed in recent years to improve the data rates and spectral efficiency (SE) by exploiting all the degrees of freedom offered in the nonlinear spectrum. However, the selection of discrete eigenvalues greatly affects the performance of both CS and DS. Designing appropriate eigenvalues of DS is an important issue to ensure the high SE and excellent performance of the system. In this paper, we discussed the selcection principle of eigenvalues and analyzed it from multiple perspectives, 11 eigenvalues with 64-quadrature amplitude modulation (64QAM) are selected for DS. Besides optimizing the eigenvalues at the transmitter, the linear minimum mean-square estimate (LMMSE) method was used at the receiver to furtherly improve the performance of DS. Through the numerical simulation, a 113 Gb/s (SE of 2.8 bits/s/Hz) full-spectrum modulated NFDM system was set up and transmitted 1120 km distance, where the Q-factors of both CS and DS are above the hard-decision forward error correction (HD-FEC) threshold. The results provide a way to design an efficiently full-spectrum modulated NFDM system.

The Brillouin optical time-domain reflectometer (BOTDR) system, which is based on the Brillouin frequency shift (BFS) of the scattering spectrum, has been widely applied to monitor the temperature and strain in many fields. The measurement accuracy of the BFS directly affects the monitoring and location of the temperature change or strain events. In this study, the characteristics of the least-squares fitting optimal solution of the Brillouin power spectrum are theoretically investigated, and a half-peak fitting algorithm is proposed to measure the BFS with high accuracy and stability. In particular, it can precisely determine the double-peak spectrum and detect the position in the transition Section of a temperature change or strain event. Furthermore, it can reduce calculation complexity and enhance measurement speed by dropping most of the data. For the fiber under test (FUT) with a length of 12 km, 200 groups of time-domain data were processed using a probe pulse of width 50 ns and sampling rate 1 GHz. The half-peak fitting algorithm increased the temperature measurement accuracy by $\sim $ 1 fold, with $\sim $ 20% calculation of full fitting. Meanwhile, it effectively eliminated the influence of minor peak in the double-peak spectrum, and optimized the spatial resolution of the temperature change position to 0.1 m, which is the maximum limitation by sampling rate.

Aiming at the existing flexible integrated scheduling algorithms that use unified rules to solve the conflicts between processes and equipment, and ignore the vertical and horizontal characteristics of the process relative positions in the process tree, a flexible integrated scheduling algorithm for dynamic determination processing equipment considering the Root-Subtree vertical and horizontal pre-scheduling (FISA-DEVH) is proposed. The process sorting strategy of the Root-Subtree vertical and horizontal pre-scheduling is proposed to sort the Root-Subtree processes of the same planned processing equipment in groups and load the Root-Subtree processes into the corresponding equipment queues, so as to fully explore the vertical and horizontal characteristics of the process. In order to give full play to the characteristics of the flexible equipment and realize the effective parallel processing between the processes, the flexible equipment optimal scheduling strategy of dynamically determining process is proposed to determine the final processing equipment of the process at each equipment driving moment and complete the production scheduling process from the Root-Subtree static pre-scheduling to dynamic fusion; The strategy uses the scheduling relationship between the unplanned processes and unplanned equipment to reduce the search scope of the combination scheme and improve the efficiency of the algorithm. The example tests show that FISA-DEVH has a better scheduling effect than the flexible integrated scheduling algorithm that processes the process tree with unified rules, and has universal significance for solving the flexible integrated scheduling problem.

This paper presents a Transformer-based Visual Exploration Network (TVENet) that capably serves as a solution for active perception problems, especially the visual exploration problem: How could a robot that is equipped with a camera explore an unknown 3D environment? The TVENet consists of a Mapper, a Global Policy and a Local Policy. The mapper is trained by supervised learning to take the visual observation as input and generate an occupancy grid map for the explored environment. The Global Policy and the Local Policy are trained by reinforcement learning in order to make navigation decision. Most state-of-the-art methods in visual exploration domain use ResNet as feature extractor, and few of them pay attention to the extraction capability of the extractor. Therefore, this paper focuses on enhancing the extraction capability, and proposes a Transformer-based Feature Pyramid Module (TFPM). Moreover, two tricks for training process are introduced to improve the performance (M.F. and Aux.) Our experiments in photo-realistic simulated environment (Habitat) demonstrate the higher-accuracy mapping of TVENet. Experimental results prove that the TFPM and tricks have positive impacts on the mapping accuracy of the visual exploration and increase it by 5.31% compared with the state-of-the-art. Most importantly, the TVENet is deployed on a real robot (NVIDIA Jetbot) to prove the feasibility of Embodied AI approaches. To the authors’ knowledge, this paper is the first one that proves the viability of the Embodied AI style approach for visual exploration tasks and deploys the pre-trained model on the NVIDIA Jetson robot.

This paper investigates an intelligent method for the motion control of a cable-driven parallel suspension system (CDPSS) in wind tunnel tests, especially at high angles of attack, which is characterized by unsteady and nonlinear aerodynamics. Considering the modeling uncertainties and the complex aerodynamic interference, a composite controller that combines deep deterministic policy gradient (DDPG) and computed-torque is proposed to improve the control performance. The tasks at hand consist in the construction of the training environment based on the dynamic equations and the Markov Decision Process (MDP) design. The supplementary computed-torque control is used to enhance the learning rate of the agent. Then a well-trained agent is applied in the high angles of attack maneuvers control with different examples, including single-DOF and multi-DOF coupled motion. The simulation results show the controller could fulfill the training tasks efficiently, and it turns out to be robust and maintain strong generalization ability despite handling the unlearned tasks.

Remote sensing image change detection is the key technology for monitoring forest windfall damages. A genetic algorithm (GA) is a branch of intelligent optimization techniques available to contribute to the surveys of windstorm and wildfire detection in forest areas. However, traditional GAs remain challenging due to several issues, such as complex calculation, poor noise immunity, and slow convergence. Analysis at the spatial level allows classifications to utilize the contextual and hierarchical information of image objects in addition to solely using spectral information. In addition, ensemble learning presents a possibility for improving classification accuracy. Ensemble classifiers combined with the spatial-based GA offers a promising method [ensemble spatial–spectral genetic algorithm (E-nGA)] for automating the process of monitoring forest loss. The research in this article is presented in four parts. First, block-matching and 3-D filtering is performed to suppress noises while enhancing valuable information. The difference image is, then, generated using the image difference method. Afterward, context-based saliency detection and fuzzy c-means algorithm are conducted on the difference image to reduce the search space. Finally, the proposed E-nGA is executed to further classify the pixels and produce the final change map. Our first proposition is to design improved genetic operators in the GA, relying not only on pixel values but also on spatial information. Our second proposition is to consider an ensemble classification model based on multiple vegetation features for decision integration. Six frequently used classification methods, as well as the simple GA, are executed to demonstrate the effectiveness of the proposed framework in improving the robustness and detection accuracy.

Based on the coupling mode theory and numerical simulation, intramode-group coupling in orbital angular momentum (OAM) fibers with elliptical deformation is analyzed in depth. The elliptical deformation not only introduces a time delay difference between even and odd modes of the same high-order vector mode but also destroys the structure of the high-order vector modes; the structural damage depends on the strength of the elliptical deformation. When the elliptical deformation is relatively small, the structural damage of high-order vector modes can be ignored and coupling can be thought as occurring only between two OAM modes composed of the same high-order vector mode. While the elliptical deformation becomes larger, the structural damage cannot be ignored, coupling appears among four OAM modes in the same OAM mode group, and the coupling between two OAM modes composed of the same high-order vector mode is more serious. The results will pave the way for the establishment of accurate propagation models and the optimization of the OAM fiber design.

The use of metallographic images to predict the mechanical properties of materials and their corrosion behavior is helpful in achieving nondestructive detection and quality control. However, after a long-term attempt, the traditional methods cannot accurately correlate the mechanical properties and corrosion behavior of materials with the corresponding microstructure images. In this study, we propose a deep learning strategy to predict the mechanical property and corrosion behavior of large-scale extruded aluminum profiles using surface optical microstructure images. The proposed models with remarkable properties were established through experimental dataset collection, dataset preprocessing, deep learning network modification, and key parameter screening. Taking extruded Al-Zn-Mg alloys with different surface microstructures as example materials, 4,800 sets of “metallographic image – hardness (HV) – corrosion potential (Ecorr)” data were experimentally collected to establish the HV and Ecorr models with prediction accuracies of 90% and 82%, respectively. The proposed HV and Ecorr models exhibit great generalization ability with mean average errors of 1.8 HV and 7.0 mV on experimental validation sets, respectively. The proposed model can accurately correlate the metallographic images, mechanical property, and corrosion behavior, which can provide theoretical support for intelligent and nondestructive testing methods to further prevent unexpected material failure.

Optical metasurfaces, i.e., thin planar structures with subwavelength metal or dielectric unit cells, have attracted great interest due to their intriguing capabilities of shaping light in both linear and nonlinear regimes. However, their saturable absorption properties and corresponding applications have so far been rarely exploited. Here we report on passively mode-locked ytterbium-doped fiber lasers utilizing saturable metasurfaces made of periodically arranged gold nanorods. Three 400-nm-period metasurfaces with gold nanorods of different lengths (from 210 to 240 nm) and plasmonic resonances (from 968 to 1044 nm) are fabricated and found to exhibit nonlinear absorption with decent modulation depths, facilitating the formation of mode-locking states at 1060 nm. The corresponding lasers generate typically mode-locked pulses with the duration of ∼ 62.3 ps, repetition rate of 10.33 MHz, and signal-to-noise ratio of ∼74 dB. Our experimental results demonstrate that the gold nanorod-based metasurfaces can be used as relatively broadband mode-lockers in the 1-μm-wavelength range.

Edge computing is a new computing paradigm which distributes tasks to edge networks for processing, it provides effective support for various mobile applications to meet their rapid response requirement. Unmanned Aerial Vehicle (UAV) has been widely used in emergency rescue, mapping, etc., with the advantages of flexible deployment and rapid movement. However, on one hand, mobile applications will be terminated when the battery energy is exhausted. On the other hand, mobile applications will be out of service when mobile devices are out of radio coverage of UAVs. How to achieve low-cost task unloading in the resource limited and location sensitive multi-UAVs edge computing environment is rather challenging. In this paper, we propose a distributed location-aware task offloading scheme, aiming at the above issues. Specifically, we create a nonlinear task allocation problem by combining the limited energy constraints of edge nodes with the random movement of users, where the cost function is divided into static and dynamic costs, respectively. Then, we formulate this problem to a convex optimization one with linear constrains, based on regularization technology. The mathematical proof shows that the scheme can support a parameterized competitive ratio without requiring any prior knowledge of the input task. The simulation results show that the proposed scheme can achieve lower cost edge computing services.

We propose a graph-based modulation format identification (MFI) scheme for elastic optical network (EON), which exploits the trajectory information of the adjacent received symbols to identify six commonly-used modulation formats signals. A uniform grid is constructed in the first quadrant of two-dimensional (2D) Stokes plane to capture the received symbols sequence, and then the corresponding trajectory information is converted into the adjacent matrix via graph theory. The eigenvector associated with the largest eigenvalue of the adjacent matrix is selected as the discriminated-feature of the corresponding modulation format signal. Subsequently, we employ cosine similarity algorithm to obtain the modulation format of the unknown signal by analyzing the angle between the discriminated-features of the canonical modulation formats signals and the unknown signal. Then, the effectiveness of the proposed MFI scheme is verified through 28 GBaud polarization division multiplexing (PDM)-binary phase shift keying (BPSK), PDM-quadrature phase shift keying (QPSK), PDM-8 quadrature amplitude modulation (QAM), PDM-16QAM, PDM-32QAM, and PDM-64QAM simulation systems. The simulation results show that the proposed MFI scheme achieves well performance on the required minimum optical signal-to-noise ratio (OSNR) value, the robustness to the variation of the transmission environment, residual chromatic dispersion (CD) and different group delay (DGD). Finally, the proposed MFI scheme is further verified by 20 GBaud PDM-QPSK/8QAM/-16QAM/-32QAM long-haul transmission experiments, and the results demonstrate that the proposed MFI scheme exhibits good resilience towards fiber nonlinear impairments.

Reading data at a temperature which different from writing can cause a large number of failed bits in 3D NAND Flash memory. In this work, the threshold voltage (Vth) temperature effect of 3D NAND flash memory cell was investigated and a method was proposed to optimize its temperature coefficient (Tco). The results suggested that the Tco of low temperature programmed high temperature read is larger than high temperature programmed low temperature read and this difference is about 2-3mv/°C, which makes against the temperature compensation of reading voltage and lead to irreparable data errors even there is error correction codes. Furthermore, the critical factor of this difference was demonstrated that is the charge loss during low temperature programmed high temperature read and the Vth shift caused by charge loss is up to 30 percent. Based on this, we proposed to optimize the Tco consistency by shortening the interval time of erase to program in endurance cycling stage and confirmed that this method is effective by experiment. As the time interval becomes smaller, the difference of the Tco decreases as well, which have great reference signification for improving the reliability of 3D NAND flash memory.

Reinforcement learning has been widely applied for sequential decision making problems in various fields of the real world, including recommendation, e-learning, etc. The features of multi-policy, latent mixture environments and offline learning implied by many real applications bring a new challenge for reinforcement learning. To this challenge, the paper proposes a reinforcement learning approach called offline multi-policy gradient for latent mixture environments. The proposed method uses an objective of expected return of trajectory with respect to the joint distribution of trajectory and model, and adopts a multi-policy searching algorithm to find the optimal policies based on expectation maximization. We also prove that the off-policy technique of importance sampling and advantage function can be used by offline multi-policy learning with fixed historical trajectories. The effectiveness of our approach is demonstrated by the experiments on both synthetic and real datasets.

Fuzzy Cognitive Map is recognized as an important model in soft computing. The Fuzzy Cognitive Map theory studies often treat uncertainty modeling and dynamic modeling separately. So far, there is no single Fuzzy Cognitive Map that can deal with uncertain data and dynamic environments simultaneously. This paper proposed an environment model to describe the link between changing weights and the dynamic environment. As the extensions of the classic Fuzzy Grey Cognitive Map, two dynamic models were designed and implemented in this work: Dynamic Fuzzy Grey Cognitive Map and Dynamic Fuzzy General Grey Map. In this work, we also analyze the characteristics of the two models and evaluate their performance using industrial control problems. The results showed that the proposed models could well handle uncertain data in dynamic environments.

Restricted Boltzmann machines (RBMs) are commonly used as pre-training methods for deep learning models. Contrastive divergence (CD) and parallel tempering (PT) are traditional training algorithms of RBMs. However, these two algorithms have shortcomings in processing high-dimensional and complex data. In particular, the number of temperature chains in PT has a significant impact on the training effect, and the PT algorithm cannot fully utilize parallel sampling from multiple temperature chains for the divergence of the algorithm. The training can quickly converge with fewer temperature chains, but this impacts the accuracy. More temperature chains can help PT achieve higher accuracy in theory, but severe divergence at the beginning of the training may ruin the training result. To exploit fully the advantages of PT and improve the ability of RBMs to process high-dimensional and complex models, this article proposes dynamic tempering chains (DTC). By dynamically changing the number of temperature chains during the training process, DTC starts training with fewer temperature chains and gradually increase the number of temperature chains with training going on, and finally get an accurate RBM. And one-step reconstruction error is proposed to measure the convergence, which can decrease the influence of the dynamic training strategy on reconstruction error. Experiments on MNIST, MNORB, Cifar 10, and Cifar 100 indicate that, compared with PT, the classification accuracy of DTC algorithm improved by up to 8%. DTC quickly converges in the early stage of training because of few exchanges among temperature chains and produces higher accuracy at the end for the global optimum model learned by more temperature chains, especially when learning high-dimensional and complex data. This proves that the DTC algorithm effectively utilizes parallel sampling of multiple temperature chains, overcomes divergence challenges, and further improves the training effect of the RBM.

Recent years have seen many advances based on Deep Convolutional Neural Networks (DCNNs) in the tasks of face recognition, most of which are developed to pursue high recognition accuracy. In this paper, we propose a novel Fast FAce Recognizer (Fast-FAR), learning to improve the speed of DCNN-based face recognition model without sacrificing recognition accuracy. Our fundamental insight is that the computation increases exponentially with the depth of a network, the easily identifiable face images can be accurately recognized by the cheap features (pixel values at shallow layers), while the challenging samples that exhibit low quality, large pose variations or occlusions need to be processed by the expensive deep layers. The major contribution of this paper is the Reinforcement Learning Agent (RLA), which is proposed to learn a decision policy determined by a reward function. The policy adaptively decides whether the recognition should be performed at an early layer with a high recognition confidence, or proceeding to the subsequent layers, thus significantly reducing feed-forward cost for the easy faces. According to the extensive experiments on the popular face recognition benchmarks, Fast-FAR reduces the inference time by 14.22%, 20.61%, and 7.84% on the dataset LFW, AgeDB-30 and CFP-FP, respectively.