Deep learning algorithm is an effective data mining method and has been used in many fields to solve practical problems.However,the deep learning algorithms often contain some hyper-parameters which may be continuous,...Deep learning algorithm is an effective data mining method and has been used in many fields to solve practical problems.However,the deep learning algorithms often contain some hyper-parameters which may be continuous,integer,or mixed,and are often given based on experience but largely affect the effectiveness of activity recognition.In order to adapt to different hyper-parameter optimization problems,our improved Cuckoo Search(CS)algorithm is proposed to optimize the mixed hyper-parameters in deep learning algorithm.The algorithm optimizes the hyper-parameters in the deep learning model robustly,and intelligently selects the combination of integer type and continuous hyper-parameters that make the model optimal.Then,the mixed hyper-parameter in Convolutional Neural Network(CNN),Long-Short-Term Memory(LSTM)and CNN-LSTM are optimized based on the methodology on the smart home activity recognition datasets.Results show that the methodology can improve the performance of the deep learning model and whether we are experienced or not,we can get a better deep learning model using our method.展开更多
Regularized system identification has become the research frontier of system identification in the past decade.One related core subject is to study the convergence properties of various hyper-parameter estimators as t...Regularized system identification has become the research frontier of system identification in the past decade.One related core subject is to study the convergence properties of various hyper-parameter estimators as the sample size goes to infinity.In this paper,we consider one commonly used hyper-parameter estimator,the empirical Bayes(EB).Its convergence in distribution has been studied,and the explicit expression of the covariance matrix of its limiting distribution has been given.However,what we are truly interested in are factors contained in the covariance matrix of the EB hyper-parameter estimator,and then,the convergence of its covariance matrix to that of its limiting distribution is required.In general,the convergence in distribution of a sequence of random variables does not necessarily guarantee the convergence of its covariance matrix.Thus,the derivation of such convergence is a necessary complement to our theoretical analysis about factors that influence the convergence properties of the EB hyper-parameter estimator.In this paper,we consider the regularized finite impulse response(FIR)model estimation with deterministic inputs,and show that the covariance matrix of the EB hyper-parameter estimator converges to that of its limiting distribution.Moreover,we run numerical simulations to demonstrate the efficacy of ourtheoretical results.展开更多
Due to outstanding performance in cheminformatics,machine learning algorithms have been increasingly used to mine molecular properties and biomedical big data.The performance of machine learning models is known to cri...Due to outstanding performance in cheminformatics,machine learning algorithms have been increasingly used to mine molecular properties and biomedical big data.The performance of machine learning models is known to critically depend on the selection of the hyper-parameter configuration.However,many studies either explored the optimal hyper-parameters per the grid searching method or employed arbitrarily selected hyper-parameters,which can easily lead to achieving a suboptimal hyper-parameter configuration.In this study,Hyperopt library embedding with the Bayesian optimization is employed to find optimal hyper-parameters for different machine learning algorithms.Six drug discovery datasets,including solubility,probe-likeness,h ERG,Chagas disease,tuberculosis,and malaria,are used to compare different machine learning algorithms with ECFP6 fingerprints.This contribution aims to evaluate whether the Bernoulli Na?ve Bayes,logistic linear regression,Ada Boost decision tree,random forest,support vector machine,and deep neural networks algorithms with optimized hyper-parameters can offer any improvement in testing as compared with the referenced models assessed by an array of metrics including AUC,F1-score,Cohen’s kappa,Matthews correlation coefficient,recall,precision,and accuracy.Based on the rank normalized score approach,the Hyperopt models achieve better or comparable performance on 33 out 36 models for different drug discovery datasets,showing significant improvement achieved by employing the Hyperopt library.The open-source code of all the 6 machine learning frameworks employed in the Hyperopt python package is provided to make this approach accessible to more scientists,who are not familiar with writing code.展开更多
A theoretical methodology is suggested for finding the malaria parasites’presence with the help of an intelligent hyper-parameter tuned Deep Learning(DL)based malaria parasite detection and classification(HPTDL-MPDC)...A theoretical methodology is suggested for finding the malaria parasites’presence with the help of an intelligent hyper-parameter tuned Deep Learning(DL)based malaria parasite detection and classification(HPTDL-MPDC)in the smear images of human peripheral blood.Some existing approaches fail to predict the malaria parasitic features and reduce the prediction accuracy.The trained model initiated in the proposed system for classifying peripheral blood smear images into the non-parasite or parasite classes using the available online dataset.The Adagrad optimizer is stacked with the suggested pre-trained Deep Neural Network(DNN)with the help of the contrastive divergence method to pre-train.The features are extracted from the images in the proposed system to train the DNN for initializing the visible variables.The smear images show the concatenated feature to be utilized as the feature vector in the proposed system.Lastly,hyper-parameters are used to fine-tune DNN to calculate the class labels’probability.The suggested system outperforms more modern methodologies with an accuracy of 91%,precision of 89%,recall of 93%and F1-score of 91%.The HPTDL-MPDC has the primary application in detecting the parasite of malaria in the smear images of human peripheral blood.展开更多
The Stackelberg prediction game(SPG)is a bilevel optimization frame-work for modeling strategic interactions between a learner and a follower.Existing meth-ods for solving this problem with general loss functions are ...The Stackelberg prediction game(SPG)is a bilevel optimization frame-work for modeling strategic interactions between a learner and a follower.Existing meth-ods for solving this problem with general loss functions are computationally expensive and scarce.We propose a novel hyper-gradient type method with a warm-start strategy to address this challenge.Particularly,we first use a Taylor expansion-based approach to obtain a good initial point.Then we apply a hyper-gradient descent method with an ex-plicit approximate hyper-gradient.We establish the convergence results of our algorithm theoretically.Furthermore,when the follower employs the least squares loss function,our method is shown to reach an e-stationary point by solving quadratic subproblems.Numerical experiments show our algorithms are empirically orders of magnitude faster than the state-of-the-art.展开更多
Buildings are a major energy consumer and carbon emitter,therefore it is important to improve building energy efficiency to achieve our sustainable development goal.Deep reinforcement learning(DRL),as an advanced buil...Buildings are a major energy consumer and carbon emitter,therefore it is important to improve building energy efficiency to achieve our sustainable development goal.Deep reinforcement learning(DRL),as an advanced building control method,demonstrates great potential for energy efficiency optimization and improved occupant comfort.However,the performance of DRL is highly sensitive to hyper-parameters,and selecting inappropriate hyper-parameters may lead to unstable learning or even failure.This study aims to investigate the design and application of DRL in building energy system control,with a specific focus on improving the performance of DRL controllers through hyper-parameter optimization(HPO)algorithms.It also aims to provide quantitative evaluation and adaptive validation of these optimized controllers.Two widely used algorithms,deep deterministic policy gradient(DDPG)and soft actor-critic(SAC),are used in the study and their performance is evaluated in different building environments based on the BOPTEST virtual testbed.One of the focuses of the study is to compare various HPO techniques,including tree-structured Parzen estimator(TPE),covariance matrix adaptation evolution strategy(CMA-ES),and combinatorial optimization methods,to determine the efficacy of different hyper-parameter optimization methods for DRL.The study enhances HPO efficiency through parallel computation and conducts a comprehensive quantitative assessment of the optimized DRL controllers,considering factors such as reduced energy consumption and improved comfort.The results show that the HPO algorithms significantly improve the performance of the DDPG and SAC controllers.A reduction of 56.94%and 68.74%in thermal discomfort is achieved,respectively.Additionally,the study demonstrates the applicability of the HPO-based approach for enhancing DRL controller performance across diverse building environments,providing valuable insights for the design and optimization of building DRL controllers.展开更多
The popularity of quadrotor Unmanned Aerial Vehicles(UAVs)stems from their simple propulsion systems and structural design.However,their complex and nonlinear dynamic behavior presents a significant challenge for cont...The popularity of quadrotor Unmanned Aerial Vehicles(UAVs)stems from their simple propulsion systems and structural design.However,their complex and nonlinear dynamic behavior presents a significant challenge for control,necessitating sophisticated algorithms to ensure stability and accuracy in flight.Various strategies have been explored by researchers and control engineers,with learning-based methods like reinforcement learning,deep learning,and neural networks showing promise in enhancing the robustness and adaptability of quadrotor control systems.This paper investigates a Reinforcement Learning(RL)approach for both high and low-level quadrotor control systems,focusing on attitude stabilization and position tracking tasks.A novel reward function and actor-critic network structures are designed to stimulate high-order observable states,improving the agent’s understanding of the quadrotor’s dynamics and environmental constraints.To address the challenge of RL hyper-parameter tuning,a new framework is introduced that combines Simulated Annealing(SA)with a reinforcement learning algorithm,specifically Simulated Annealing-Twin Delayed Deep Deterministic Policy Gradient(SA-TD3).This approach is evaluated for path-following and stabilization tasks through comparative assessments with two commonly used control methods:Backstepping and Sliding Mode Control(SMC).While the implementation of the well-trained agents exhibited unexpected behavior during real-world testing,a reduced neural network used for altitude control was successfully implemented on a Parrot Mambo mini drone.The results showcase the potential of the proposed SA-TD3 framework for real-world applications,demonstrating improved stability and precision across various test scenarios and highlighting its feasibility for practical deployment.展开更多
High-dimensional and incomplete(HDI) matrices are primarily generated in all kinds of big-data-related practical applications. A latent factor analysis(LFA) model is capable of conducting efficient representation lear...High-dimensional and incomplete(HDI) matrices are primarily generated in all kinds of big-data-related practical applications. A latent factor analysis(LFA) model is capable of conducting efficient representation learning to an HDI matrix,whose hyper-parameter adaptation can be implemented through a particle swarm optimizer(PSO) to meet scalable requirements.However, conventional PSO is limited by its premature issues,which leads to the accuracy loss of a resultant LFA model. To address this thorny issue, this study merges the information of each particle's state migration into its evolution process following the principle of a generalized momentum method for improving its search ability, thereby building a state-migration particle swarm optimizer(SPSO), whose theoretical convergence is rigorously proved in this study. It is then incorporated into an LFA model for implementing efficient hyper-parameter adaptation without accuracy loss. Experiments on six HDI matrices indicate that an SPSO-incorporated LFA model outperforms state-of-the-art LFA models in terms of prediction accuracy for missing data of an HDI matrix with competitive computational efficiency.Hence, SPSO's use ensures efficient and reliable hyper-parameter adaptation in an LFA model, thus ensuring practicality and accurate representation learning for HDI matrices.展开更多
Domain shift is when the data used in training does not match the ones it will be applied to later on under similar conditions.Domain shift will reduce accuracy in results.To prevent this,domain adaptation is done,whi...Domain shift is when the data used in training does not match the ones it will be applied to later on under similar conditions.Domain shift will reduce accuracy in results.To prevent this,domain adaptation is done,which adapts the pre-trained model to the target domain.In real scenarios,the availability of labels for target data is rare thus resulting in unsupervised domain adaptation.Herein,we propose an innovative approach where source-free domain adaptation models and Generative Adversarial Networks(GANs)are integrated to improve the performance of computer vision or robotic vision-based systems in our study.Cosine Generative Adversarial Network(CosGAN)is developed as a GAN that uses cosine embedding loss to handle issues associated with unsupervised source-relax domain adaptations.For less complex architecture,the CosGAN training process has two steps that produce results almost comparable to other state-of-the-art techniques.The efficiency of CosGAN was compared by conducting experiments using benchmarked datasets.The approach was evaluated on different datasets and experimental results show superiority over existing state-of-the-art methods in terms of accuracy as well as generalization ability.This technique has numerous applications including wheeled robots,autonomous vehicles,warehouse automation,and all image-processing-based automation tasks so it can reshape the field of robotic vision with its ability to make robots adapt to new tasks and environments efficiently without requiring additional labeled data.It lays the groundwork for future expansions in robotic vision and applications.Although GAN provides a variety of outstanding features,it also increases the risk of instability and over-fitting of the training data thus making the data difficult to converge.展开更多
Machine learning models may outperform traditional statistical regression algorithms for predicting clinical outcomes.Proper validation of building such models and tuning their underlying algorithms is necessary to av...Machine learning models may outperform traditional statistical regression algorithms for predicting clinical outcomes.Proper validation of building such models and tuning their underlying algorithms is necessary to avoid over-fitting and poor generalizability,which smaller datasets can be more prone to.In an effort to educate readers interested in artificial intelligence and model-building based on machine-learning algorithms,we outline important details on crossvalidation techniques that can enhance the performance and generalizability of such models.展开更多
Due to the fact that the vibration signal of the rotating machine is one-dimensional and the large-scale convolution kernel can obtain a better perception field, on the basis of the classical convolution neural networ...Due to the fact that the vibration signal of the rotating machine is one-dimensional and the large-scale convolution kernel can obtain a better perception field, on the basis of the classical convolution neural network model(LetNet-5), one-dimensional large-kernel convolution neural network(1 DLCNN) is designed. Since the hyper-parameters of 1 DLCNN have a greater impact on network performance, the genetic algorithm(GA) is used to optimize the hyper-parameters, and the method of optimizing the parameters of 1 DLCNN by the genetic algorithm is named GA-1 DLCNN. The experimental results show that the optimal network model based on the GA-1 DLCNN method can achieve 99.9% fault diagnosis accuracy, which is much higher than those of other traditional fault diagnosis methods. In addition, the 1 DLCNN is compared with one-dimencional small-kernel convolution neural network(1 DSCNN) and the classical two-dimensional convolution neural network model. The input sample lengths are set to be 128, 256, 512, 1 024, and 2 048, respectively, and the final diagnostic accuracy results and the visual scatter plot show that the effect of 1 DLCNN is optimal.展开更多
The technology for beneficiation of banded iron ores containing low iron value is a challenging task due to increasing demand of quality iron ore in India. A flotation process has been developed to treat one such ore,...The technology for beneficiation of banded iron ores containing low iron value is a challenging task due to increasing demand of quality iron ore in India. A flotation process has been developed to treat one such ore, namely banded hematite quartzite （BHQ） containing 41.8wt% Fe and 41.5wt% SiO2,by using oleic acid, methyl isobutyl carbinol （MIBC）, and sodium silicate as the collector, frother, and dispersant, respectively. The relative effects of these variables have been evaluated in half-normal plots and Pareto charts using central composite rotatable design. A quadratic response model has been developed for both Fe grade and recovery and optimized within the experimental range. The optimum reagent dosages are found to be as follows： collector concentration of 243.58 g/t, dispersant concentration of 195.67 g/t, pH 8.69, and conditioning time of 4.8 min to achieve the maximum Fe grade of 64.25% with 67.33% recovery. The predictions of the model with regard to iron grade and recovery are in good agreement with the experimental results.展开更多
Evolutionary algorithm is time-consuming because of the large number of evolutions and much times of finite element analysis, when it is used to optimize the wing structure of a certain high altitude long endurance un...Evolutionary algorithm is time-consuming because of the large number of evolutions and much times of finite element analysis, when it is used to optimize the wing structure of a certain high altitude long endurance unmanned aviation vehicle（UAV）. In order to improve efficiency it is proposed to construct a model management framework to perform the multi-objective optimization design of wing structure. The sufficient accurate approximation models of objective and constraint functions in the wing structure optimization model are built when using the model management framework, therefore in the evolutionary algorithm a number of finite element analyses can he avoided and the satisfactory multi-objective optimization results of the wing structure of the high altitude long endurance UAV are obtained.展开更多
In this paper,we investigate the minimization of age of information(AoI),a metric that measures the information freshness,at the network edge with unreliable wireless communications.Particularly,we consider a set of u...In this paper,we investigate the minimization of age of information(AoI),a metric that measures the information freshness,at the network edge with unreliable wireless communications.Particularly,we consider a set of users transmitting status updates,which are collected by the user randomly over time,to an edge server through unreliable orthogonal channels.It begs a natural question:with random status update arrivals and obscure channel conditions,can we devise an intelligent scheduling policy that matches the users and channels to stabilize the queues of all users while minimizing the average AoI?To give an adequate answer,we define a bipartite graph and formulate a dynamic edge activation problem with stability constraints.Then,we propose an online matching while learning algorithm(MatL)and discuss its implementation for wireless scheduling.Finally,simulation results demonstrate that the MatL is reliable to learn the channel states and manage the users’buffers for fresher information at the edge.展开更多
Gaussian process(GP)has fewer parameters,simple model and output of probabilistic sense,when compared with the methods such as support vector machines.Selection of the hyper-parameters is critical to the performance o...Gaussian process(GP)has fewer parameters,simple model and output of probabilistic sense,when compared with the methods such as support vector machines.Selection of the hyper-parameters is critical to the performance of Gaussian process model.However,the common-used algorithm has the disadvantages of difficult determination of iteration steps,over-dependence of optimization effect on initial values,and easily falling into local optimum.To solve this problem,a method combining the Gaussian process with memetic algorithm was proposed.Based on this method,memetic algorithm was used to search the optimal hyper parameters of Gaussian process regression(GPR)model in the training process and form MA-GPR algorithms,and then the model was used to predict and test the results.When used in the marine long-range precision strike system(LPSS)battle effectiveness evaluation,the proposed MA-GPR model significantly improved the prediction accuracy,compared with the conjugate gradient method and the genetic algorithm optimization process.展开更多
Geological structures often exhibit smooth characteristics away from sharp discontinuities. One aim of geophysical inversion is to recover information about the smooth structures as well as about the sharp discontinui...Geological structures often exhibit smooth characteristics away from sharp discontinuities. One aim of geophysical inversion is to recover information about the smooth structures as well as about the sharp discontinuities. Because no specific operator can provide a perfect sparse representation of complicated geological models, hyper-parameter regularization inversion based on the iterative split Bregman method was used to recover the features of both smooth and sharp geological structures. A novel preconditioned matrix was proposed, which counteracted the natural decay of the sensitivity matrix and its inverse matrix was calculated easily. Application of the algorithm to synthetic data produces density models that are good representations of the designed models. The results show that the algorithm proposed is feasible and effective.展开更多
The lowest order Pl-nonconforming triangular finite element method （FEM） for elliptic and parabolic interface problems is investigated. Under some reasonable regularity assumptions on the exact solutions, the optima...The lowest order Pl-nonconforming triangular finite element method （FEM） for elliptic and parabolic interface problems is investigated. Under some reasonable regularity assumptions on the exact solutions, the optimal order error estimates are obtained in the broken energy norm. Finally, some numerical results are provided to verify the theoretical analysis.展开更多
Accurate prediction of drilling efficiency is critical for developing the earth-rock excavation schedule.The single machine learning(ML)prediction models usually suffer from problems including parameter sensitivity an...Accurate prediction of drilling efficiency is critical for developing the earth-rock excavation schedule.The single machine learning(ML)prediction models usually suffer from problems including parameter sensitivity and overfitting.In addition,the influence of environmental and operational factors is often ignored.In response,a novel stacking-based ensemble learning method taking into account the combined effects of those factors is proposed.Through multiple comparison tests,four models,e Xtreme gradient boosting(XGBoost),random forest(RF),back propagation neural network(BPNN)as the base learners,and support vector regression(SVR)as the meta-learner,are selected for stacking.Furthermore,an improved cuckoo search optimization(ICSO)algorithm is developed for hyper-parameter optimization of the ensemble model.The application to a real-world project demonstrates that the proposed method outperforms the popular single ML method XGBoost and the ensemble model optimized by particle swarm optimization(PSO),with 16.43%and 4.88%improvements of mean absolute percentage error(MAPE),respectively.展开更多
During the Twelfth Five-Year plan,large-scale construction of smart grid with safe and stable operation requires a timely and accurate short-term load forecasting method.Moreover,along with the full-scale smart grid c...During the Twelfth Five-Year plan,large-scale construction of smart grid with safe and stable operation requires a timely and accurate short-term load forecasting method.Moreover,along with the full-scale smart grid construction,the power supply mode and consumption mode of the whole system can be optimized through the accurate short-term load forecasting;and the security,stability and cleanness of the system can be guaranteed.展开更多
基金Supported by the Anhui Province Sports Health Information Monitoring Technology Engineering Research Center Open Project (KF2023012)。
文摘Deep learning algorithm is an effective data mining method and has been used in many fields to solve practical problems.However,the deep learning algorithms often contain some hyper-parameters which may be continuous,integer,or mixed,and are often given based on experience but largely affect the effectiveness of activity recognition.In order to adapt to different hyper-parameter optimization problems,our improved Cuckoo Search(CS)algorithm is proposed to optimize the mixed hyper-parameters in deep learning algorithm.The algorithm optimizes the hyper-parameters in the deep learning model robustly,and intelligently selects the combination of integer type and continuous hyper-parameters that make the model optimal.Then,the mixed hyper-parameter in Convolutional Neural Network(CNN),Long-Short-Term Memory(LSTM)and CNN-LSTM are optimized based on the methodology on the smart home activity recognition datasets.Results show that the methodology can improve the performance of the deep learning model and whether we are experienced or not,we can get a better deep learning model using our method.
基金supported in part by the National Natural Science Foundation of China(No.62273287)by the Shenzhen Science and Technology Innovation Council(Nos.JCYJ20220530143418040,JCY20170411102101881)the Thousand Youth Talents Plan funded by the central government of China.
文摘Regularized system identification has become the research frontier of system identification in the past decade.One related core subject is to study the convergence properties of various hyper-parameter estimators as the sample size goes to infinity.In this paper,we consider one commonly used hyper-parameter estimator,the empirical Bayes(EB).Its convergence in distribution has been studied,and the explicit expression of the covariance matrix of its limiting distribution has been given.However,what we are truly interested in are factors contained in the covariance matrix of the EB hyper-parameter estimator,and then,the convergence of its covariance matrix to that of its limiting distribution is required.In general,the convergence in distribution of a sequence of random variables does not necessarily guarantee the convergence of its covariance matrix.Thus,the derivation of such convergence is a necessary complement to our theoretical analysis about factors that influence the convergence properties of the EB hyper-parameter estimator.In this paper,we consider the regularized finite impulse response(FIR)model estimation with deterministic inputs,and show that the covariance matrix of the EB hyper-parameter estimator converges to that of its limiting distribution.Moreover,we run numerical simulations to demonstrate the efficacy of ourtheoretical results.
基金financial support provided by the National Key Research and Development Project(2019YFC0214403)Chongqing Joint Chinese Medicine Scientific Research Project(2021ZY023984)。
文摘Due to outstanding performance in cheminformatics,machine learning algorithms have been increasingly used to mine molecular properties and biomedical big data.The performance of machine learning models is known to critically depend on the selection of the hyper-parameter configuration.However,many studies either explored the optimal hyper-parameters per the grid searching method or employed arbitrarily selected hyper-parameters,which can easily lead to achieving a suboptimal hyper-parameter configuration.In this study,Hyperopt library embedding with the Bayesian optimization is employed to find optimal hyper-parameters for different machine learning algorithms.Six drug discovery datasets,including solubility,probe-likeness,h ERG,Chagas disease,tuberculosis,and malaria,are used to compare different machine learning algorithms with ECFP6 fingerprints.This contribution aims to evaluate whether the Bernoulli Na?ve Bayes,logistic linear regression,Ada Boost decision tree,random forest,support vector machine,and deep neural networks algorithms with optimized hyper-parameters can offer any improvement in testing as compared with the referenced models assessed by an array of metrics including AUC,F1-score,Cohen’s kappa,Matthews correlation coefficient,recall,precision,and accuracy.Based on the rank normalized score approach,the Hyperopt models achieve better or comparable performance on 33 out 36 models for different drug discovery datasets,showing significant improvement achieved by employing the Hyperopt library.The open-source code of all the 6 machine learning frameworks employed in the Hyperopt python package is provided to make this approach accessible to more scientists,who are not familiar with writing code.
文摘A theoretical methodology is suggested for finding the malaria parasites’presence with the help of an intelligent hyper-parameter tuned Deep Learning(DL)based malaria parasite detection and classification(HPTDL-MPDC)in the smear images of human peripheral blood.Some existing approaches fail to predict the malaria parasitic features and reduce the prediction accuracy.The trained model initiated in the proposed system for classifying peripheral blood smear images into the non-parasite or parasite classes using the available online dataset.The Adagrad optimizer is stacked with the suggested pre-trained Deep Neural Network(DNN)with the help of the contrastive divergence method to pre-train.The features are extracted from the images in the proposed system to train the DNN for initializing the visible variables.The smear images show the concatenated feature to be utilized as the feature vector in the proposed system.Lastly,hyper-parameters are used to fine-tune DNN to calculate the class labels’probability.The suggested system outperforms more modern methodologies with an accuracy of 91%,precision of 89%,recall of 93%and F1-score of 91%.The HPTDL-MPDC has the primary application in detecting the parasite of malaria in the smear images of human peripheral blood.
文摘The Stackelberg prediction game(SPG)is a bilevel optimization frame-work for modeling strategic interactions between a learner and a follower.Existing meth-ods for solving this problem with general loss functions are computationally expensive and scarce.We propose a novel hyper-gradient type method with a warm-start strategy to address this challenge.Particularly,we first use a Taylor expansion-based approach to obtain a good initial point.Then we apply a hyper-gradient descent method with an ex-plicit approximate hyper-gradient.We establish the convergence results of our algorithm theoretically.Furthermore,when the follower employs the least squares loss function,our method is shown to reach an e-stationary point by solving quadratic subproblems.Numerical experiments show our algorithms are empirically orders of magnitude faster than the state-of-the-art.
基金supported by the National Natural Science Foundation of China(No.72371072)Jiangsu Association for Science&Technology Youth Science&Technology Talents Lifting Project(No.JSTJ-2023-JS001).
文摘Buildings are a major energy consumer and carbon emitter,therefore it is important to improve building energy efficiency to achieve our sustainable development goal.Deep reinforcement learning(DRL),as an advanced building control method,demonstrates great potential for energy efficiency optimization and improved occupant comfort.However,the performance of DRL is highly sensitive to hyper-parameters,and selecting inappropriate hyper-parameters may lead to unstable learning or even failure.This study aims to investigate the design and application of DRL in building energy system control,with a specific focus on improving the performance of DRL controllers through hyper-parameter optimization(HPO)algorithms.It also aims to provide quantitative evaluation and adaptive validation of these optimized controllers.Two widely used algorithms,deep deterministic policy gradient(DDPG)and soft actor-critic(SAC),are used in the study and their performance is evaluated in different building environments based on the BOPTEST virtual testbed.One of the focuses of the study is to compare various HPO techniques,including tree-structured Parzen estimator(TPE),covariance matrix adaptation evolution strategy(CMA-ES),and combinatorial optimization methods,to determine the efficacy of different hyper-parameter optimization methods for DRL.The study enhances HPO efficiency through parallel computation and conducts a comprehensive quantitative assessment of the optimized DRL controllers,considering factors such as reduced energy consumption and improved comfort.The results show that the HPO algorithms significantly improve the performance of the DDPG and SAC controllers.A reduction of 56.94%and 68.74%in thermal discomfort is achieved,respectively.Additionally,the study demonstrates the applicability of the HPO-based approach for enhancing DRL controller performance across diverse building environments,providing valuable insights for the design and optimization of building DRL controllers.
基金supported by Princess Nourah Bint Abdulrahman University Researchers Supporting Project number(PNURSP2024R135)Princess Nourah Bint Abdulrahman University,Riyadh,Saudi Arabia.
文摘The popularity of quadrotor Unmanned Aerial Vehicles(UAVs)stems from their simple propulsion systems and structural design.However,their complex and nonlinear dynamic behavior presents a significant challenge for control,necessitating sophisticated algorithms to ensure stability and accuracy in flight.Various strategies have been explored by researchers and control engineers,with learning-based methods like reinforcement learning,deep learning,and neural networks showing promise in enhancing the robustness and adaptability of quadrotor control systems.This paper investigates a Reinforcement Learning(RL)approach for both high and low-level quadrotor control systems,focusing on attitude stabilization and position tracking tasks.A novel reward function and actor-critic network structures are designed to stimulate high-order observable states,improving the agent’s understanding of the quadrotor’s dynamics and environmental constraints.To address the challenge of RL hyper-parameter tuning,a new framework is introduced that combines Simulated Annealing(SA)with a reinforcement learning algorithm,specifically Simulated Annealing-Twin Delayed Deep Deterministic Policy Gradient(SA-TD3).This approach is evaluated for path-following and stabilization tasks through comparative assessments with two commonly used control methods:Backstepping and Sliding Mode Control(SMC).While the implementation of the well-trained agents exhibited unexpected behavior during real-world testing,a reduced neural network used for altitude control was successfully implemented on a Parrot Mambo mini drone.The results showcase the potential of the proposed SA-TD3 framework for real-world applications,demonstrating improved stability and precision across various test scenarios and highlighting its feasibility for practical deployment.
基金supported in part by the National Natural Science Foundation of China (62372385, 62272078, 62002337)the Chongqing Natural Science Foundation (CSTB2022NSCQ-MSX1486, CSTB2023NSCQ-LZX0069)the Deanship of Scientific Research at King Abdulaziz University, Jeddah, Saudi Arabia (RG-12-135-43)。
文摘High-dimensional and incomplete(HDI) matrices are primarily generated in all kinds of big-data-related practical applications. A latent factor analysis(LFA) model is capable of conducting efficient representation learning to an HDI matrix,whose hyper-parameter adaptation can be implemented through a particle swarm optimizer(PSO) to meet scalable requirements.However, conventional PSO is limited by its premature issues,which leads to the accuracy loss of a resultant LFA model. To address this thorny issue, this study merges the information of each particle's state migration into its evolution process following the principle of a generalized momentum method for improving its search ability, thereby building a state-migration particle swarm optimizer(SPSO), whose theoretical convergence is rigorously proved in this study. It is then incorporated into an LFA model for implementing efficient hyper-parameter adaptation without accuracy loss. Experiments on six HDI matrices indicate that an SPSO-incorporated LFA model outperforms state-of-the-art LFA models in terms of prediction accuracy for missing data of an HDI matrix with competitive computational efficiency.Hence, SPSO's use ensures efficient and reliable hyper-parameter adaptation in an LFA model, thus ensuring practicality and accurate representation learning for HDI matrices.
文摘Domain shift is when the data used in training does not match the ones it will be applied to later on under similar conditions.Domain shift will reduce accuracy in results.To prevent this,domain adaptation is done,which adapts the pre-trained model to the target domain.In real scenarios,the availability of labels for target data is rare thus resulting in unsupervised domain adaptation.Herein,we propose an innovative approach where source-free domain adaptation models and Generative Adversarial Networks(GANs)are integrated to improve the performance of computer vision or robotic vision-based systems in our study.Cosine Generative Adversarial Network(CosGAN)is developed as a GAN that uses cosine embedding loss to handle issues associated with unsupervised source-relax domain adaptations.For less complex architecture,the CosGAN training process has two steps that produce results almost comparable to other state-of-the-art techniques.The efficiency of CosGAN was compared by conducting experiments using benchmarked datasets.The approach was evaluated on different datasets and experimental results show superiority over existing state-of-the-art methods in terms of accuracy as well as generalization ability.This technique has numerous applications including wheeled robots,autonomous vehicles,warehouse automation,and all image-processing-based automation tasks so it can reshape the field of robotic vision with its ability to make robots adapt to new tasks and environments efficiently without requiring additional labeled data.It lays the groundwork for future expansions in robotic vision and applications.Although GAN provides a variety of outstanding features,it also increases the risk of instability and over-fitting of the training data thus making the data difficult to converge.
文摘Machine learning models may outperform traditional statistical regression algorithms for predicting clinical outcomes.Proper validation of building such models and tuning their underlying algorithms is necessary to avoid over-fitting and poor generalizability,which smaller datasets can be more prone to.In an effort to educate readers interested in artificial intelligence and model-building based on machine-learning algorithms,we outline important details on crossvalidation techniques that can enhance the performance and generalizability of such models.
基金The National Natural Science Foundation of China(No.51675098)
文摘Due to the fact that the vibration signal of the rotating machine is one-dimensional and the large-scale convolution kernel can obtain a better perception field, on the basis of the classical convolution neural network model(LetNet-5), one-dimensional large-kernel convolution neural network(1 DLCNN) is designed. Since the hyper-parameters of 1 DLCNN have a greater impact on network performance, the genetic algorithm(GA) is used to optimize the hyper-parameters, and the method of optimizing the parameters of 1 DLCNN by the genetic algorithm is named GA-1 DLCNN. The experimental results show that the optimal network model based on the GA-1 DLCNN method can achieve 99.9% fault diagnosis accuracy, which is much higher than those of other traditional fault diagnosis methods. In addition, the 1 DLCNN is compared with one-dimencional small-kernel convolution neural network(1 DSCNN) and the classical two-dimensional convolution neural network model. The input sample lengths are set to be 128, 256, 512, 1 024, and 2 048, respectively, and the final diagnostic accuracy results and the visual scatter plot show that the effect of 1 DLCNN is optimal.
文摘The technology for beneficiation of banded iron ores containing low iron value is a challenging task due to increasing demand of quality iron ore in India. A flotation process has been developed to treat one such ore, namely banded hematite quartzite （BHQ） containing 41.8wt% Fe and 41.5wt% SiO2,by using oleic acid, methyl isobutyl carbinol （MIBC）, and sodium silicate as the collector, frother, and dispersant, respectively. The relative effects of these variables have been evaluated in half-normal plots and Pareto charts using central composite rotatable design. A quadratic response model has been developed for both Fe grade and recovery and optimized within the experimental range. The optimum reagent dosages are found to be as follows： collector concentration of 243.58 g/t, dispersant concentration of 195.67 g/t, pH 8.69, and conditioning time of 4.8 min to achieve the maximum Fe grade of 64.25% with 67.33% recovery. The predictions of the model with regard to iron grade and recovery are in good agreement with the experimental results.
文摘Evolutionary algorithm is time-consuming because of the large number of evolutions and much times of finite element analysis, when it is used to optimize the wing structure of a certain high altitude long endurance unmanned aviation vehicle（UAV）. In order to improve efficiency it is proposed to construct a model management framework to perform the multi-objective optimization design of wing structure. The sufficient accurate approximation models of objective and constraint functions in the wing structure optimization model are built when using the model management framework, therefore in the evolutionary algorithm a number of finite element analyses can he avoided and the satisfactory multi-objective optimization results of the wing structure of the high altitude long endurance UAV are obtained.
基金supported in part by Shanghai Pujiang Program under Grant No.21PJ1402600in part by Natural Science Foundation of Chongqing,China under Grant No.CSTB2022NSCQ-MSX0375+4 种基金in part by Song Shan Laboratory Foundation,under Grant No.YYJC022022007in part by Zhejiang Provincial Natural Science Foundation of China under Grant LGJ22F010001in part by National Key Research and Development Program of China under Grant 2020YFA0711301in part by National Natural Science Foundation of China under Grant 61922049。
文摘In this paper,we investigate the minimization of age of information(AoI),a metric that measures the information freshness,at the network edge with unreliable wireless communications.Particularly,we consider a set of users transmitting status updates,which are collected by the user randomly over time,to an edge server through unreliable orthogonal channels.It begs a natural question:with random status update arrivals and obscure channel conditions,can we devise an intelligent scheduling policy that matches the users and channels to stabilize the queues of all users while minimizing the average AoI?To give an adequate answer,we define a bipartite graph and formulate a dynamic edge activation problem with stability constraints.Then,we propose an online matching while learning algorithm(MatL)and discuss its implementation for wireless scheduling.Finally,simulation results demonstrate that the MatL is reliable to learn the channel states and manage the users’buffers for fresher information at the edge.
基金Project(513300303)supported by the General Armament Department,China
文摘Gaussian process(GP)has fewer parameters,simple model and output of probabilistic sense,when compared with the methods such as support vector machines.Selection of the hyper-parameters is critical to the performance of Gaussian process model.However,the common-used algorithm has the disadvantages of difficult determination of iteration steps,over-dependence of optimization effect on initial values,and easily falling into local optimum.To solve this problem,a method combining the Gaussian process with memetic algorithm was proposed.Based on this method,memetic algorithm was used to search the optimal hyper parameters of Gaussian process regression(GPR)model in the training process and form MA-GPR algorithms,and then the model was used to predict and test the results.When used in the marine long-range precision strike system(LPSS)battle effectiveness evaluation,the proposed MA-GPR model significantly improved the prediction accuracy,compared with the conjugate gradient method and the genetic algorithm optimization process.
基金Projects(41174061,41374120)supported by the National Natural Science Foundation of China
文摘Geological structures often exhibit smooth characteristics away from sharp discontinuities. One aim of geophysical inversion is to recover information about the smooth structures as well as about the sharp discontinuities. Because no specific operator can provide a perfect sparse representation of complicated geological models, hyper-parameter regularization inversion based on the iterative split Bregman method was used to recover the features of both smooth and sharp geological structures. A novel preconditioned matrix was proposed, which counteracted the natural decay of the sensitivity matrix and its inverse matrix was calculated easily. Application of the algorithm to synthetic data produces density models that are good representations of the designed models. The results show that the algorithm proposed is feasible and effective.
基金Project supported by the National Natural Science Foundation of China(No.11271340)
文摘The lowest order Pl-nonconforming triangular finite element method （FEM） for elliptic and parabolic interface problems is investigated. Under some reasonable regularity assumptions on the exact solutions, the optimal order error estimates are obtained in the broken energy norm. Finally, some numerical results are provided to verify the theoretical analysis.
基金supported by the Yalong River Joint Funds of the National Natural Science Foundation of China(No.U1965207)the National Natural Science Foundation of China(Nos.51839007,51779169,and 52009090)。
文摘Accurate prediction of drilling efficiency is critical for developing the earth-rock excavation schedule.The single machine learning(ML)prediction models usually suffer from problems including parameter sensitivity and overfitting.In addition,the influence of environmental and operational factors is often ignored.In response,a novel stacking-based ensemble learning method taking into account the combined effects of those factors is proposed.Through multiple comparison tests,four models,e Xtreme gradient boosting(XGBoost),random forest(RF),back propagation neural network(BPNN)as the base learners,and support vector regression(SVR)as the meta-learner,are selected for stacking.Furthermore,an improved cuckoo search optimization(ICSO)algorithm is developed for hyper-parameter optimization of the ensemble model.The application to a real-world project demonstrates that the proposed method outperforms the popular single ML method XGBoost and the ensemble model optimized by particle swarm optimization(PSO),with 16.43%and 4.88%improvements of mean absolute percentage error(MAPE),respectively.
文摘During the Twelfth Five-Year plan,large-scale construction of smart grid with safe and stable operation requires a timely and accurate short-term load forecasting method.Moreover,along with the full-scale smart grid construction,the power supply mode and consumption mode of the whole system can be optimized through the accurate short-term load forecasting;and the security,stability and cleanness of the system can be guaranteed.
