High Learning Speed Research Articles

Multi-Agent Collaborative Target Search Based on the Multi-Agent Deep Deterministic Policy Gradient with Emotional Intrinsic Motivation

Multi-agent collaborative target search is one of the main challenges in the multi-agent field, and deep reinforcement learning (DRL) is a good way to learn such a task. However, DRL always faces the problem of sparse reward, which to some extent reduces its efficiency in task learning. Introducing intrinsic motivation has proved to be a useful way to make the sparse reward in DRL. So, based on the multi-agent deep deterministic policy gradient (MADDPG) structure, a new MADDPG algorithm with the emotional intrinsic motivation name MADDPG-E is proposed in this paper for the multi-agent collaborative target search. In MADDPG-E, a new emotional intrinsic motivation module with three emotions, joy, sadness, and fear, is designed. The three emotions are defined by corresponding psychological knowledge to the multi-agent embodied situations in an environment. An emotional steady-state variable function H is then designed to help judge the goodness of the emotions. Based on H, an emotion-based intrinsic reward function is finally proposed. With the designed emotional intrinsic motivation module, the multi-agent system always tries to make itself joy, which means it always learns to search the target. To show the effectiveness of the proposed MADDPG-E algorithm, two kinds of simulation experiments with a determined initial position and random initial position, respectively, are carried out, and comparisons are performed with MADDPG as well as MADDPG-ICM (MADDPG with an intrinsic curiosity module). The results show that with the designed emotional intrinsic motivation module, MADDPG-E has a higher learning speed and better learning stability, and the advantage is more obvious when facing complex situations.

Elastic-net based robust extreme learning machine for one-class classification

The one-class extreme learning machine (OC-ELM) builds a classification model by learning samples of the known class to detect abnormal samples, and has the advantages of high learning speed and good generalization performance. However, OC-ELM with the square loss function is sensitive to outliers, which leads to poor robustness. To address this issue, an elastic-net based robust extreme learning machine for one-class classification (ER-OCELM) is proposed to achieve both excellent robustness and comparable sparsity. Correntropy loss function is utilized to limit the negative effects of outliers due to its property of bounded. With an elastic-net regularizer, automatic variable selection and continuous shrinkage can be performed simultaneously, and groups of correlated variables can be selected to make the model sparse. The corresponding optimization can be transformed into the linear equation system. We employ an iterative reweighted algorithm to obtain the optimal solution. In each iteration, the solution form of ER-OCELM is similar to that of ELM. Experiments on artificial and benchmark datasets verify that ER-OCELM has superior robustness and sparsity compared to many state-of-the-art methods.

딥러닝기반 토마토 병해 진단 서비스 연구

Tomato crops are easy to expose to disease and spread in a short period of time, so late measures against disease are directly related to production and sales, which can cause damage. Therefore, there is a need for a service that enables early prevention by simply and accurately diagnosing tomato diseases in the field. In this paper, we construct a system that applies a deep learning-based model in which ImageNet transition is learned in advance to classify and serve nine classes of tomatoes for disease and normal cases. We use the input of MobileNet, ResNet, with a deep learning-based CNN structure that builds a lighter neural network using a composite product for the image set of leaves classifying tomato disease and normal from the Plant Village dataset. Through the learning of two proposed models, it is possible to provide fast and convenient services using MobileNet with high accuracy and learning speed.

Adaptive Control of Discrete-time Nonlinear Systems Using ITF-ORVFL

Random vector functional ink (RVFL) networks belong to a class of single hidden layer neural networks in which some parameters are randomly selected. Their network structure in which contains the direct links between inputs and outputs is unique, and stability analysis and real-time performance are two difficulties of the control systems based on neural networks. In this paper, combining the advantages of RVFL and the ideas of online sequential extreme learning machine (OS-ELM) and initial-training-free online extreme learning machine (ITF-OELM), a novel online learning algorithm which is named as initial-training-free online random vector functional link algo rithm (ITF-ORVFL) is investigated for training RVFL. The link vector of RVFL network can be analytically determined based on sequentially arriving data by ITF-ORVFL with a high learning speed, and the stability for nonlinear systems based on this learning algorithm is analyzed. The experiment results indicate that the proposed ITF-ORVFL is effective in coping with nonparametric uncertainty.

Matrix Neo-Fuzzy-System and its Online Learning in Image Recognition Task

The paper proposes a 2D-hybrid system of computational intelligence, which is based on the generalized neo-fuzzy neuron. The system is characterised by high approximate abilities, simple computational implementation, and high learning speed. The characteristic property of the proposed system is that on its input the signal is fed not in the traditional vector form, but in the image-matrix form. Such an approach allows getting rid of additional convolution-pooling layers that are used in deep neural networks as an encoder. The main elements of the proposed system are a fuzzified multidimensional bilinear model, additional softmax layer, and multidimensional generalized neo-fuzzy neuron tuning with cross-entropy criterion. Compared to deep neural systems, the proposed matrix neo-fuzzy system contains gradually fewer tuning parameters – synaptic weights. The usage of the time-optimal algorithm for tuning synaptic weights allows implementing learning in an online mode.

MEC-Based Jamming-Aided Anti-Eavesdropping with Deep Reinforcement Learning for WBANs

Wireless body area network (WBAN) suffers secure challenges, especially the eavesdropping attack, due to constraint resources. In this article, deep reinforcement learning (DRL) and mobile edge computing (MEC) technology are adopted to formulate a DRL-MEC-based jamming-aided anti-eavesdropping (DMEC-JAE) scheme to resist the eavesdropping attack without considering the channel state information. In this scheme, a MEC sensor is chosen to send artificial jamming signals to improve the secrecy rate of the system. Power control technique is utilized to optimize the transmission power of both the source sensor and the MEC sensor to save energy. The remaining energy of the MEC sensor is concerned to ensure routine data transmission and jamming signal transmission. Additionally, the DMEC-JAE scheme integrates with transfer learning for a higher learning rate. The performance bounds of the scheme concerning the secrecy rate, energy consumption, and the utility are evaluated. Simulation results show that the DMEC-JAE scheme can approach the performance bounds with high learning speed, which outperforms the benchmark schemes.

An efficient epileptic seizure classification system using empirical wavelet transform and multi-fuse reduced deep convolutional neural network with digital implementation

Epilepsy, an incurable brain disorder portrayed by seizures, is the most common neurological disease worldwide. The embryonic detection of epileptic action helps the psychologist for the diagnosis of epileptic seizure and reduces the seizure effect on the patient’s life. Empirical wavelet transform (EWT), and a novel multi-fuse reduced deep convolutional neural network (MF-RDCNN) classifier are integrated to design a computer-aided-diagnosis (CAD) system for meticulous classification of epileptic seizure from electroencephalogram (EEG) signals. The EWT is enforced on recorded EEG signals to extract three distinct band-limited modes (BLMs) and the proposed classifier is used to compute the most informative unsupervised signatures automatically by taking the extracted BLMs as inputs. The experimental results are carried out using three EEG databases provided by Bonn-University, Germany single-channel EEG (dataset-A), Neurology and Sleep Center, New Delhi single-channel EEG (dataset-B), and Boston Children’s Hospital multichannel scalp EEG (dataset-C) to verify the effectiveness of the proposed algorithm. The performance of the proposed method is superior over other prevalent methods and has 100%,99.82% overall mean classification accuracy for two class and three class classification problems respectively using ten-fold cross-validation. Moreover, the proposed method has specificity (SPE) of 99.29%, sensitivity (SEN) of 99.86%, and classification accuracy (ACC) of 99.29% with 0.71% of false positive rate per hour (FPR/h) by taking 40% of data for training, 40% of data for testing, and the remaining 20% of data for validation from the total data incorporated with the database-C. The lesser computational complexity, higher learning speed, short event recognition time, remarkable overall mean classification accuracy, and outstanding overall performance are the major advantages of the proposed EWT-MF-RDCNN method over EWT-RDCNN and RDCNN method for accurate classification of seizure EEG epochs. The digital architecture of the MF-RDCNN classifier is also implemented in a high-speed FPGA processor to develop a CAD system for efficacious diagnosis of epileptic seizure activity online.

BLSTM-Based Adaptive Finite-Time Output-Constrained Control for a Class of AUSs with Dynamic Disturbances and Actuator Faults

In this paper, a BLSTM-based adaptive finite-time control structure has been constructed for a class of aerospace unmanned systems (AUSs). Firstly, a novel neural network structure possessing both the time memory characteristics and high learning speed, broad long short-term memory (BLSTM) network, has been constructed. Secondly, several nonlinear functions are utilized to transform the tracking errors into a novel state vector to guarantee the output constraints of the AUSs. Thirdly, the fractional-order control law and the corresponding adaptive laws are designed, and as a result, the adaptive finite-time control scheme has been formed. Moreover, to handle the uncertainties and the faulty elevator outputs, an inequality of the multivariable systems is utilized. Consequently, by fusing the output of the BLSTM, the transformation of the tracking errors, and the adaptive finite-time control law, a novel BLSTM-based intelligent adaptive finite-time control structure has been established for the AUSs under output constraints. The simulation results show that the proposed BLSTM-based adaptive control algorithm can achieve favorable control results for the AUSs with multiple uncertainties.

A Multi-Agent Reinforcement Learning Approach for Capacity Sharing in Multi-Tenant Scenarios

5G is envisioned to simultaneously provide diverse service types with heterogeneous needs under very different application scenarios and business models. Therefore, network slicing is included as a key feature of the 5G architecture to allow sharing a common infrastructure among different tenants, such as mobile communication providers, vertical market players, etc. In order to provide the Radio Access Network (RAN) with network slicing capabilities, mechanisms that efficiently distribute the available capacity among the different tenants while satisfying their needs are required. For this purpose, this paper proposes a multi-agent reinforcement learning approach for RAN capacity sharing. It makes use of the Deep Q-Network algorithm in a way that each agent is associated to a different tenant and learns the capacity to be provided to this tenant in each cell while ensuring that the service level agreements are satisfied and that the available radio resources are efficiently used. The consideration of multiple agents contributes to a better scalability and higher learning speed in comparison to single-agent approaches. In this respect, results show that the policy learnt by the agent of one tenant can be generalised and directly applied by other agents, thus reducing the complexity of the training and making the proposed solution easily scalable, e.g., to add new tenants in the system. The proposed approach is well aligned with the on-going 3GPP standardization work and guidelines for the parametrization of the solution are provided, thus enforcing its practical applicability.

FPGA implementation of epileptic seizure detection using semisupervised reduced deep convolutional neural network

In this article, an optimized variational mode decomposition (OVMD), reduced deep convolutional neural network (RDCNN), and multi-kernel random vector functional link network (MKRVFLN) are combined to recognize the epileptic seizure using electroencephalogram (EEG) signals. An improved particle swarm optimization based on the maximum value of log energy entropy is introduced to compute the optimized values of the number of band-limited intrinsic mode functions (BLIMFs) and the data-fidelity factor. The Kurtosis and correlation coefficient are used to extract the most efficient BLIMF from highly nonlinear and non-stationary multi-class epilepsy EEG signals. The RDCNN structure is designed to extract the most discriminative unsupervised features and feed into the novel supervised MKRVFLN classifier to train efficiently by reducing the cross-entropy loss for recognizing the seizure epochs efficaciously. The Bonn University, Germany, Neurology and Sleep centre, New Delhi single-channel EEG, and Boston Children’s Hospital multichannel scalp EEG (sEEG) datasets are considered to evaluate the overall efficiency of the proposed method. The proposed method has produced 100% classification accuracy (CA) for classification problem (CP) 1 to 4 present in database-A, as well as for all CP incorporated with database-B and also it has 99.88% CA for CP-5 of the Bonn University database using a ten-fold cross-validation strategy. Furthermore, the proposed method has also the sensitivity of 100%, specificity of 99.34%, and CA of 99.37% with a negligible false positive per hour (FPR/h) of 0.663% by adopting 50% training, 30% testing and 20% validation of total data present in Boston database and its overall performance outperforms as compared to other state-of-the-art methods. The less computational complexity, higher learning speed, accurate epileptic seizure recognition, remarkable classification accuracy, short event recognition time, and negligible FPR/h are the main advantages of the proposed OVMD-RDCNN-MKRVFLN method over RDCNN, OVMD-RDCNN, and OVMD-RDCNN-KRVFLN methods. The novel RDCNN-MKRVFLN digital architecture is designed and implemented on a high-speed field-programmable gate array (FPGA) hardware environment to design a computer-aided-diagnosis (CAD) system for online epileptic seizure diagnosis. The simplicity, feasibility, and practicability of the proposed method validate the effectiveness of automatic seizure recognition.

Epileptic Seizure Recognition Using Reduced Deep Convolutional Stack Autoencoder and Improved Kernel RVFLN From EEG Signals.

In this paper, reduced deep convolutional stack autoencoder (RDCSAE) and improved kernel random vector functional link network (IKRVFLN) are combined to recognize the epileptic seizure using both the multichannel scalp and single-channel electroencephalogram (EEG) signals. The novel RDCSAE structure is designed to extract the most discriminative unsupervised features from EEG signals and fed into the proposed supervised IKRVFLN classifier to train efficiently by reducing the mean-square error cost function for recognizing the epileptic seizure activity with promising accuracy. The proposed RDCSAE-IKRVFLN algorithm is tested over the benchmark Boston Children's Hospital multichannel scalp EEG (sEEG) and Boon University, Germany single-channel EEG databases. The less computational complexity, higher learning speed, better model generalization, accurate epileptic seizure recognition, remarkable classification accuracy, negligible false positive rate per hour (FPR/h) and short event recognition time are the main advantages of the proposed RDCSAE-IKRVFLN method over reduced deep convolutional neural network (RDCNN), RDCSAE and RDCSAE-KRVFLN methods. The proposed RDCSAE-IKRVFLN method is implemented in a high-speed reconfigurable field-programmable gate array (FPGA) hardware environment to design a computer-aided-diagnosis (CAD) system for automatic epileptic seizure diagnosis. The simplicity, feasibility, and practicability of the proposed method validate the stable and reliable performances of epilepsy detection and recognition.

Multifuse multilayer multikernel RVFLN+ of process modes decomposition and approximate entropy data from iEEG/sEEG signals for epileptic seizure recognition

In this paper, the extracted features using variational mode decomposition (VMD) and approximate entropy (ApEn) privileged information of the input EEG signals are combined with multilayer multikernel random vector functional link network plus (MMRVFLN+) classifier to recognize the epileptic seizure epochs efficaciously. In our experiment Bonn University single-channel intracranial electroencephalogram (iEEG) and Children's Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) multichannel scalp EEG (sEEG) recordings are considered to evaluate the efficacy of the proposed method. The VMD is applied on chaotic, non-stationary, nonlinear, and complex EEG signal to decompose it into three band-limited intrinsic mode functions (BLIMFs). The Hilbert transform (HT) is applied on BLIMFs to extract informative spectral and temporal features. The ApEn is computed from the raw EEG signals as the privileged information and given to the multi-hidden layer structure to obtain the most discriminative compressed form. The scatter plots show the distinct nature of compressed privileged ApEn information among the seizure pattern classes. The linear as well as nonlinear mapping, local and global kernel function, high-learning speed, less computationally complex MMRVFLN+ classifier is proposed to recognize the seizure events accurately by importing the efficacious features with ApEn as the input. The advanced signal processing algorithm i.e., Hilbert Huang transform (HHT) with ApEn and MMRVFLN+ are combined to compare the performance with the proposed VMDHTApEn-MMRVFLN+ method. The proposed method has remarkable recognition ability, superior classification accuracy, and excellent overall performance as compared to other methods. The digital architecture of the multifuse MMRVFLN+ is developed and implemented on a high-speed reconfigurable FPGA hardware platform to validate the effectiveness of the proposed method. The superior classification accuracy, the negligible false positive rate per hour (FPR/h), simplicity, feasibility, robustness, and practicability of the proposed method validate its ability to recognize the epileptic seizure epochs automatically.

Dynamically Generated Hierarchical Neural Networks Designed With the Aid of Multiple Support Vector Regressors and PNN Architecture With Probabilistic Selection.

The two issues on dynamically generated hierarchical neural networks such as the sort of basic neurons and how to compose a layer are considered in this article. On the first issue, a variant version of the least-square support vector regression (SVR) is chosen as a basic neuron. Support vector machine (SVM) is a representative classifier which usually shows good classification performance. Along with the SVMs, SVR was introduced to deal with the regression problem. Especially, least-square SVR has the advantages of high learning speed due to the substitution of the inequality constraints by the equality constraint in the formulation of the optimization problem. Based on the least-square SVR, the multiple least-square (MLS) SVR, which is a type of a linear combination of least-square SVRs with fuzzy clustering, is proposed to improve the modeling performance. In addition, a hierarchical neural network, where the MLS SVR is utilized as the generic node instead of the conventional polynomial, is developed. The key issues of hierarchical neural networks, which are generated dynamically layer by layer, are discussed on how to retain the diversity of the nodes located at the same layer according to the increase of the layer. In order to maintain the diversity of the nodes, various selection methods such as truncation selection and roulette wheel selection (RWS) to choose the nodes among candidate nodes are proposed. In addition, in order to reduce the computational overhead to determine all candidates which exhibit all compositions of the input variables, a new implementation method is proposed. From the viewpoint of the diversity of the selected nodes and the computational aspects, it is shown that the proposed method is preferred over the conventional design methodology.

FPGA-Based Semisupervised Multifusion RDCNN of Process Robust VMD Data With Online Kernel RVFLN for Power Quality Events Recognition

The improved particle swarm optimization algorithm is integrated with variational mode decomposition (VMD) to extract the efficient band-limited intrinsic mode function (BLIMF) of the single and combined power quality events (PQEs). The selected BLIMF of the robust VMD (RVMD) and the privileged Fourier magnitude spectrum (FMS) information are fed to the proposed reduced deep convolutional neural network (RDCNN) for the extraction of the most discriminative unsupervised features. The RVMD-FMS-RDCNN method shows minimum feature overlapping compared with RDCNN and RVMD-RDCNN methods. The feature vector is imported to the novel supervised online kernel random vector functional link network (OKRVFLN) for quick and accurate categorization of complex PQEs. The proposed RVMD-FMS-RDCNN-OKRVFLN method produces excellent recognition capability over RDCNN, RVMD-RDCNN, and RVMD-RDCNN-OKRVFLN methods in noise-free and noisy environments. The unique BLIMF selection, clear detection, descriptive feature extraction, higher learning speed, superior classification accuracy, and robust antinoise performances are considerable importance of the proposed RVMD-FMS-RDCNN-OKRVFLN method. Finally, the proposed method architecture is developed and implemented in a very-high-speed ML506 Virtex-5 FPGA to text, examine, and validate the feasibility, performances, and practicability for online monitoring of the PQEs.

Randomized Neural Networks Based Decentralized Multi-Task Learning via Hybrid Multi-Block ADMM

In multi-task learning (MTL), related tasks learn jointly to improve generalization performance. To exploit the high learning speed of feed-forward neural networks (FNN), we apply the randomized single-hidden layer FNN (RSF) to the MTL problem, where the output weights of RSFs for all the tasks are learned collaboratively. We first present the RSF based MTL problem in the centralized setting, which is solved by the proposed MTL-RSF algorithm. Due to the fact that many data sets of different tasks are geo-distributed, decentralized machine learning is studied. We formulate the decentralized MTL problem based on RSF as majorized multi-block optimization with coupled bi-convex objective functions. To solve the problem, we propose the DMTL-RSF algorithm, which is a hybrid Jacobian and Gauss-Seidel Proximal multi-block alternating direction method of multipliers (ADMM). Further, to reduce the computation load of DMTL-RSF, DMTL-RSF with first-order approximation (FO-DMTL-RSF) is presented. Theoretical analysis shows that the convergence to the stationary point of proposed decentralized algorithms can be guaranteed conditionally. Through simulations, we demonstrate the convergence of presented algorithms, and also show that they can outperform existing MTL methods. Moreover, by adjusting the dimension of hidden feature space, there exists a trade-off between communication load and learning accuracy for DMTL-RSF.

A Regional NWP Tropospheric Delay Inversion Method Based on a General Regression Neural Network Model.

Tropospheric delay is a major error source that affects the initialization and re-initialization speed of the Global Navigation Satellite System’s (GNSS) medium-/long-range baseline in Network Real-Time Kinematic (NRTK) positioning. Fusing the meteorological data from the Numerical Weather Prediction (NWP) model to estimate the zenith tropospheric delay (ZTD) is one of the current research hotspots. However, research has shown that the ZTD derived from NWP models is still not accurate enough for high-precision GNSS positioning applications without the estimation of the residual tropospheric delay. To date, General Regression Neural Network (GRNN) has been applied in many fields. It has a high learning speed and simple structure, and can approximate any function with arbitrary precision. In this study, we developed a regional NWP tropospheric delay inversion method based on a GRNN model to improve the accuracy of the tropospheric delay derived from the NWP model. The accuracy of the tropospheric delays derived from reanalysis data of the European Center for Medium-Range Weather Forecasts (ECMWF) and the US National Centers for Environmental Prediction (NCEP) was assessed through comparisons with the results of the International GPS Service (IGS). The variation characteristics of the residual of the ZTD inverted by NWP data were analyzed considering the factors of temperature, humidity, latitude, and season. To evaluate the performance of this new method, the National Center Atmospheric Research (NCAR) troposphere data of 650 stations in Japan in 2005 were collected as a reference to compare the accuracy of the ZTD before and after using the new method. The experimental results showed that the GRNN model has obvious advantages in fitting the NWP ZTD residual. The mean residual and the root mean square deviation (RMSD) of the ZTD inverted using the method of this study were 9.5 mm and 12.7 mm, respectively, showing reductions of 20.8% and 19.1%, respectively, as compared to the standard NWP model. For long-range baseline (155 km and 207 km), the corrected NWP-constrained RTK showed a reduction of over 43% in the initialization time compared with the standard RTK, and showed a reduction of over 24% in the initialization time compared with the standard NWP-constrained RTK.

A Novel Isomap-SVR Soft Sensor Model and Its Application in Rotary Kiln Calcination Zone Temperature Prediction

Soft sensing technology has been proved to be an effective tool for the online estimation of unmeasured or variables that are difficult to directly measure. The performance of a soft sensor depends heavily on its convergence speed and generalization ability to a great extent. Based on this idea, we propose a new soft sensor model, Isomap-SVR. First, the sample data set is divided into training set and testing set by using self-organizing map (SOM) neural network to ensure the fairness and symmetry of data segmentation. Isometric feature mapping (Isomap) method is used for dimensionality reduction of the model input data, which could not only reduce the structure complexity of the proposed model but speed up learning speed, and then the Support Vector Machine Regression (SVR) is applied to the regression model. A novel bat algorithm based on Cauchy mutation and Lévy flight strategy is used to optimize parameters of Isomap and SVR to improve the accuracy of the proposed model. Finally, the model is applied to the prediction of the temperature of rotary kiln calcination zone, which is difficult to measure directly. The simulation results show that the proposed soft sensor modeling method has higher learning speed and better generalization ability. Compared with other algorithms, this algorithm has obvious advantages and is an effective modeling method.

An EWT-PCA and Extreme Learning Machine Based Diagnosis Approach for Hydraulic Pump

Aiming at solving the problem that the fault characteristic of hydraulic pumps in the signal is weak, a fault diagnosis method combining empirical wavelet transform (EWT), principal component analysis (PCA) and extreme learning machine (ELM) is proposed. The vibration signal is firstly decomposed into several components that contain different frequency adaptively by EWT. After extracting features from the decomposed signals, the dimension of the feature vector is reduced by PCA. Finally, the dimensionally-reduced feature vectors that retain key fault characteristics are fed into ELM, which is a classifier with high learning speed and high generalization ability, to obtain the fault modes classification result. Experiments show that the proposed method can capture fault patterns of the hydraulic pump well and achieve high-accuracy fault diagnosis results.

Efficient hindsight reinforcement learning using demonstrations for robotic tasks with sparse rewards

The goal of reinforcement learning is to enable an agent to learn by using rewards. However, some robotic tasks naturally specify with sparse rewards, and manually shaping reward functions is a difficult project. In this article, we propose a general and model-free approach for reinforcement learning to learn robotic tasks with sparse rewards. First, a variant of Hindsight Experience Replay, Curious and Aggressive Hindsight Experience Replay, is proposed to improve the sample efficiency of reinforcement learning methods and avoid the need for complicated reward engineering. Second, based on Twin Delayed Deep Deterministic policy gradient algorithm, demonstrations are leveraged to overcome the exploration problem and speed up the policy training process. Finally, the action loss is added into the loss function in order to minimize the vibration of output action while maximizing the value of the action. The experiments on simulated robotic tasks are performed with different hyperparameters to verify the effectiveness of our method. Results show that our method can effectively solve the sparse reward problem and obtain a high learning speed.

Ensemble of extreme learning machines for multivariate calibration of near-infrared spectroscopy.

Inspired by the attractive features of extreme learning machine (ELM), a simple ensemble ELM algorithm, named EELM, is proposed for multivariate calibration of near-infrared spectroscopy. Such an algorithm takes full advantage of random initialization of the weights of the hidden layer in ELM for obtaining the diversity between member models. Also, by combining a large number of member models, the stability of the final prediction can be greatly improved and the ensemble model outperforms its best member model. Compared with partial least-squares (PLS), the superiority of EELM is attributed to its inherent characteristics of high learning speed, simple structure and excellent predictive performance. Three NIR spectral datasets concerning solid samples are used to verify the proposed algorithm in terms of both the accuracy and robustness. The results confirmed the superiority of EELM to classic PLS. Also, even if the experiment is done on NIR datasets, it provides a good reference for other spectral calibration.