Random Initialization Research Articles

Gradient Flow in Sparse Neural Networks and How Lottery Tickets Win

Sparse Neural Networks (NNs) can match the generalization of dense NNs using a fraction of the compute/storage for inference, and have the potential to enable efficient training. However, naively training unstructured sparse NNs from random initialization results in significantly worse generalization, with the notable exceptions of Lottery Tickets (LTs) and Dynamic Sparse Training (DST). In this work, we attempt to answer: (1) why training unstructured sparse networks from random initialization performs poorly and; (2) what makes LTs and DST the exceptions? We show that sparse NNs have poor gradient flow at initialization and propose a modified initialization for unstructured connectivity. Furthermore, we find that DST methods significantly improve gradient flow during training over traditional sparse training methods. Finally, we show that LTs do not improve gradient flow, rather their success lies in re-learning the pruning solution they are derived from — however, this comes at the cost of learning novel solutions.

Novel enhanced Salp Swarm Algorithms using opposition-based learning schemes for global optimization problems

Salp Swarm Algorithm (SSA) is a recent approach with a simple implementation, few parameters, and low computational cost. SSA has been used in different optimization problems and provided competitive results compared to other popular metaheuristics. However, like other metaheuristics, SSA also faces problems such as lack of population diversity and premature convergence in local optima. Its other drawback is that it uses a random population initialization procedure that may produce solutions in a less effective starting point (i.e., initial population) with a poor diversity. In SSA, a single spatial search mode is used to update the position of followers, which makes it easy to fall in the local optima. SSA is more inclined to the exploitation phase, so it cannot always conduct a global search well and, in some cases, cannot find the global optimal solution. This work proposes an in-depth investigation of applying Opposition-based Learning (OBL) and its impact on the search space’s coverage, accuracy, exploration and exploitation, and convergence of the SSA’s search process. Hence, five enhanced hybrid SSA-OBL algorithms were developed and compared to the traditional SSA, Chimp Optimization Algorithm (ChOA), Giza Pyramids Construction (GPC), and Arithmetic Optimization Algorithm (AOA) in 15 IEEE CEC-2015 benchmark global minimization functions with 30, 50, and 100 dimensions. The Wilcoxon Signed Rank Test is used to test the statistical significance of the results. The proposed opposition-based SSAs statistically outperform basic SSA, ChOA, GPC, and AOA, respectively, with an average of 7.6, 15, 13.4, 15 out of 15 functions with 30 dimensions, 8, 15, 15, 15 with 50 dimensions, and 9.2, 15, 15, and 15 with 100 dimensions. Besides, we also applied the proposal in solving real-world optimization problems such as circular antenna array design problem and spacecraft trajectory optimization problem, and the opposition-based SSAs statistically outperform other competitive algorithms. The experiment results showed that incorporating different OBL schemes in SSA enhanced its exploration ability leading to better performance.

An Improved K-means Clustering Algorithm Towards an Efficient Data-Driven Modeling

K-means algorithm is one of the well-known unsupervised machine learning algorithms. The algorithm typically finds out distinct non-overlapping clusters in which each point is assigned to a group. The minimum squared distance technique distributes each point to the nearest clusters or subgroups. One of the K-means algorithm’s main concerns is to find out the initial optimal centroids of clusters. It is the most challenging task to determine the optimum position of the initial clusters’ centroids at the very first iteration. This paper proposes an approach to find the optimal initial centroids efficiently to reduce the number of iterations and execution time. To analyze the effectiveness of our proposed method, we have utilized different real-world datasets to conduct experiments. We have first analyzed COVID-19 and patient datasets to show our proposed method’s efficiency. A synthetic dataset of 10M instances with 8 dimensions is also used to estimate the performance of the proposed algorithm. Experimental results show that our proposed method outperforms traditional kmeans++ and random centroids initialization methods regarding the computation time and the number of iterations.

Abstract 5031: Improved deconvolution of combined bulk and single-cell RNA-sequencing data

Abstract Identifying how cell types and their abundances evolve during tumor progression is critical to understanding the mechanisms and identifying predictors of metastasis. Single-cell RNA sequencing (scRNA-seq) has been especially promising in resolving heterogeneity of expression programs at the single cell level but is not always available, for example for large cohort studies or longitudinal analysis of archived samples. In such cases, cell subpopulations must be inferred by deconvolution, a process that can infer single-cell genomic data from bulk data but has limited ability to resolve fine clonal structure. We extend our previous bulk genomic deconvolution tool, Robust and Accurate Deconvolution (RAD), to establish a new method, scRAD, that can use reference scRNA-seq to interpret sample collections for which only bulk RNA-seq is available for some samples, e.g., clonally resolving archived primary (PRM) tissues using scRNA-seq from metastases (METs). We preprocess scRNA-seq data to accurately represent gene expression profiles, yielding a signature matrix S then extend our RAD method via a regularization term to deconvolve bulk data while maximizing consistency with S. We validate our method on semi-synthetic data derived from human PRM breast cancer cases and bone and ovary METs, showing that scRAD improves inference of single-cell gene expression profiles and their frequencies relative to the prior RAD with random initialization or initialization using the single-cell matrix S (Table 1). We then apply scRAD to a collection of paired PRM and MET tumors to quantify progression changes in common cell types. One-sided Kaplan-Meier analysis shows that tumors inferred to increase the mast cell fraction from PRM to MET exhibit lower overall survival (p<0.05), consistent with the role of mast cells in metastatic growth and propagation. Tumors that show increased macrophage cell fraction from PRM to MET show improved overall survival (p<0.04), consistent with the role of immune infiltration in survival. mean square error (MSE) of gene expression and mixture fraction inference on semi-simulated data Method RAD with random initialization RAD with random initialization RAD with random initialization RAD initialized with S RAD initialized with S RAD initialized with S scRAD scRAD scRAD Sample number 2 4 8 2 4 8 2 4 8 Gene Expression MSE 0.41 0.28 0.37 0.31 0.27 0.37 0.22 0.15 0.13 Mixture fraction MSE 0.58 0.82 0.59 0.39 0.83 0.58 0.19 0.20 0.22 Citation Format: Haoyun Lei, Xiaoyan A. Guo, Yifeng Tao, Kai Ding, Xuecong Fu, Steffi Oesterreich, Adrian V. Lee, Russell Schwartz. Improved deconvolution of combined bulk and single-cell RNA-sequencing data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5031.

Neural-based automatic scoring model for Chinese-English interpretation with a multi-indicator assessment

Manual evaluation could be time-consuming, unreliable and unreproducible in Chinese-English interpretation. Therefore, it is necessary to develop an automatic scoring system. This paper proposes an accurate automatic scoring model for Chinese-English interpretation via a multi-indicator assessment. From the three dimensions (i.e. keywords, content, and grammar) of the scoring rubrics, three improved attention-based BiLSTM neural models are proposed to learn the text of the transcribed responses. In the feature vectorisation stage, the pre-training model Bert is utilised to vectorise the keywords and content, and a random initialisation is used for the grammar. In addition, the fluency is also taken into account based on the speech speed. The overall holistic score is obtained by fusing the four scores using the random forest regressor. The experimental results demonstrate that the proposed scoring method is effective and can perform as good as the manual scoring.

Benchmarking Deep Learning Models for Tooth Structure Segmentation

A wide range of deep learning (DL) architectures with varying depths are available, with developers usually choosing one or a few of them for their specific task in a nonsystematic way. Benchmarking (i.e., the systematic comparison of state-of-the art architectures on a specific task) may provide guidance in the model development process and may allow developers to make better decisions. However, comprehensive benchmarking has not been performed in dentistry yet. We aimed to benchmark a range of architecture designs for 1 specific, exemplary case: tooth structure segmentation on dental bitewing radiographs. We built 72 models for tooth structure (enamel, dentin, pulp, fillings, crowns) segmentation by combining 6 different DL network architectures (U-Net, U-Net++, Feature Pyramid Networks, LinkNet, Pyramid Scene Parsing Network, Mask Attention Network) with 12 encoders from 3 different encoder families (ResNet, VGG, DenseNet) of varying depth (e.g., VGG13, VGG16, VGG19). On each model design, 3 initialization strategies (ImageNet, CheXpert, random initialization) were applied, resulting overall into 216 trained models, which were trained up to 200 epochs with the Adam optimizer (learning rate = 0.0001) and a batch size of 32. Our data set consisted of 1,625 human-annotated dental bitewing radiographs. We used a 5-fold cross-validation scheme and quantified model performances primarily by the F1-score. Initialization with ImageNet or CheXpert weights significantly outperformed random initialization (P < 0.05). Deeper and more complex models did not necessarily perform better than less complex alternatives. VGG-based models were more robust across model configurations, while more complex models (e.g., from the ResNet family) achieved peak performances. In conclusion, initializing models with pretrained weights may be recommended when training models for dental radiographic analysis. Less complex model architectures may be competitive alternatives if computational resources and training time are restricting factors. Models developed and found superior on nondental data sets may not show this behavior for dental domain-specific tasks.

Hidden-state modeling of a cross-section of geoelectric time series data can provide reliable intermediate-term probabilistic earthquake forecasting in Taiwan

Abstract. Geoelectric time series (TS) have long been studied for their potential for probabilistic earthquake forecasting, and a recent model (GEMSTIP) directly used the skewness and kurtosis of geoelectric TS to provide times of increased probability (TIPs) for earthquakes for several months in the future. We followed up on this work by applying the hidden Markov model (HMM) to the correlation, variance, skewness, and kurtosis TSs to identify two hidden states (HSs) with different distributions of these statistical indexes. More importantly, we tested whether these HSs could separate time periods into times of higher/lower earthquake probabilities. Using 0.5 Hz geoelectric TS data from 20 stations across Taiwan over 7 years, we first computed the statistical index TSs and then applied the Baum–Welch algorithm with multiple random initializations to obtain a well-converged HMM and its HS TS for each station. We then divided the map of Taiwan into a 16-by-16 grid map and quantified the forecasting skill, i.e., how well the HS TS could separate times of higher/lower earthquake probabilities in each cell in terms of a discrimination power measure that we defined. Next, we compare the discrimination power of empirical HS TSs against those of 400 simulated HS TSs and then organized the statistical significance values from this cellular-level hypothesis testing of the forecasting skill obtained into grid maps of discrimination reliability. Having found such significance values to be high for many grid cells for all stations, we proceeded with a statistical hypothesis test of the forecasting skill at the global level to find high statistical significance across large parts of the hyperparameter spaces of most stations. We therefore concluded that geoelectric TSs indeed contain earthquake-related information and the HMM approach is capable of extracting this information for earthquake forecasting.

Distilling ensemble of explanations for weakly-supervised pre-training of image segmentation models

While fine-tuning pre-trained networks has become a popular way to train image segmentation models, such backbone networks for image segmentation are frequently pre-trained using image classification source datasets, e.g., ImageNet. Though image classification datasets could provide the backbone networks with rich visual features and discriminative ability, they are incapable of fully pre-training the target model (i.e., backbone+segmentation modules) in an end-to-end manner. The segmentation modules are left to random initialization in the fine-tuning process due to the lack of segmentation labels in classification datasets. In our work, we propose a method that leverages Pseudo Semantic Segmentation Labels (PSSL), to enable the end-to-end pre-training for image segmentation models based on classification datasets. PSSL was inspired by the observation that the explanation results of classification models, obtained through explanation algorithms such as CAM, SmoothGrad and LIME, would be close to the pixel clusters of visual objects. Specifically, PSSL is obtained for each image by interpreting the classification results and aggregating an ensemble of explanations queried from multiple classifiers to lower the bias caused by single models. With PSSL for every image of ImageNet, the proposed method leverages a weighted segmentation learning procedure to pre-train the segmentation network en masse. Experiment results show that, with ImageNet accompanied by PSSL as the source dataset, the proposed end-to-end pre-training strategy successfully boosts the performance of various segmentation models, i.e., PSPNet-ResNet50, DeepLabV3-ResNet50, and OCRNet-HRNetW18, on a number of segmentation tasks, such as CamVid, VOC-A, VOC-C, ADE20K, and CityScapes, with significant improvements.

Design of Machine Learning Algorithm for Tourism Demand Prediction.

Unused hotel rooms, unused event tickets, and unsold items are all examples of wasted expenses and earnings. Governments require accurate tourism demand forecasting in order to make informed decisions on topics such as infrastructure development and lodging site planning; therefore, accurate tourism demand forecasting becomes vital. Artificial intelligence (AI) models such as neural networks and security violation report (SVR) have been used effectively in tourist demand forecasting as a result of the fast advancement of AI. This paper constructs a tourism demand forecasting model based on machine learning on the basis of the existing forecasting model research. The completed work is as follows: (1) It introduces a large number of domestic and foreign literatures on tourism volume forecasting and proposes the research content of this paper. (2) It is proposed to stack the long short-term memory- (LSTM-) based autoencoders deeply, by adopting a hierarchical greedy pretraining method to replace the random weight initialization method used in the deep network and combining this pretraining stage and fine-tuning network together to form the SAE-LSTM prediction model for improving the performance of deep learning models. (3) This paper uses the monthly search engine strength data of city A's monthly tourist volume and its related influencing factors as the data set; processes the data set to make the model adapt to the data input; uses mean absolute error (MAE), root mean square error (RMSE), MAPE, and other model evaluation indicators; and uses LSTM and the constructed SAE-LSTM model to conduct comparative experiments to predict the number of tourist arrivals in four years. The prediction results of the models proposed in this paper are better than those of the LSTM model. According to the experimental results, the superiority of the proposed LSTM-based unsupervised pretraining method is demonstrated.

PINNup: Robust Neural Network Wavefield Solutions Using Frequency Upscaling and Neuron Splitting

AbstractSeismic wave‐equation based methods, for example, full waveform inversion, are currently used to illuminate the interior of Earth. Solving for the frequency‐domain scattered wavefield via physics‐informed neural network (PINN) has great potential in increasing the flexibility and reducing the computational cost of seismic modeling and inversion. However, when dealing with high‐frequency wavefields using PINN, its accuracy and training cost limit its application. Thus, we propose a novel implementation of PINN using frequency upscaling and neuron splitting, which allows the neural network model to grow in size as we increase the frequency while leveraging the information from the pre‐trained model for lower‐frequency wavefields, resulting in fast convergence to highly accurate wavefield solutions. Numerical results show that, compared to the commonly used PINN with random initialization, the proposed PINN exhibits notable superiority in terms of convergence and accuracy and can achieve neuron based high‐frequency wavefield solutions with a shallow model.

On the Existence of Global Minima and Convergence Analyses for Gradient Descent Methods in the Training of Deep Neural Networks

In this article we study fully-connected feedforward deep ReLU ANNs with an arbitrarily large number of hidden layers and we prove convergence of the risk of the GD optimization method with random initializations in the training of such ANNs under the assumption that the unnormalized probability density function of the probability distribution of the input data of the considered supervised learning problem is piecewise polynomial, under the assumption that the target function (describing the relationship between input data and the output data) is piecewise polynomial, and under the assumption that the risk function of the considered supervised learning problem admits at least one regular global minimum. In addition, in the special situation of shallow ANNs with just one hidden layer and one-dimensional input we also verify this assumption by proving in the training of such shallow ANNs that for every Lipschitz continuous target function there exists a global minimum in the risk landscape. Finally, in the training of deep ANNs with ReLU activation we also study solutions of gradient flow (GF) differential equations and we prove that every non-divergent GF trajectory converges with a polynomial rate of convergence to a critical point (in the sense of limiting Fr\'echet subdifferentiability). Our mathematical convergence analysis builds up on ideas from our previous article Eberle et al., on tools from real algebraic geometry such as the concept of semi-algebraic functions and generalized Kurdyka-Lojasiewicz inequalities, on tools from functional analysis such as the Arzel\`a-Ascoli theorem, on tools from nonsmooth analysis such as the concept of limiting Fr\'echet subgradients, as well as on the fact that the set of realization functions of shallow ReLU ANNs with fixed architecture forms a closed subset of the set of continuous functions revealed by Petersen et al.

Ensemble of Artificial Intelligence Techniques for Bacterial Antimicrobial Resistance (AMR) Estimation Using Topic Modeling and Similarity Measure

In recent times, bacterial Antimicrobial Resistance (AMR) analyses becomes a hot study topic. The AMR comprises information related to the antibiotic product name, class name, subclass name, type, subtype, gene type, etc., which can fight against the illness. However, the tagging language used to determine the data is of free context. These contexts often contain ambiguous data, which leads to a hugely challenging issue in retrieving, organizing, merging, and finding the relevant data. Manually reading this text and labelling is not time-consuming. Hence, topic modeling overcomes these challenges and provides efficient results in categorizing the topic and in determining the data. In this view, this research work designs an ensemble of artificial intelligence for categorizing the AMR gene data and determine the relationship between the antibiotics. The proposed model includes a weighted voting based ensemble model by the incorporation of Latent Dirichlet Allocation (LDA) and Hierarchical Recurrent Neural Networks (HRNN), shows the novelty of the work. It is used for determining the amount of “topics” that cluster utilizing a multidimensional scaling approach. In addition, the proposed model involves the data pre-processing stage to get rid of stop words, punctuations, lower casing, etc. Moreover, an explanatory data analysis uses word cloud which assures the proper functionality and to proceed with the model training process. Besides, three approaches namely perplexity, Harmonic mean, and Random initialization of K are employed to determine the number of topics. For experimental validation, an openly accessible Bacterial AMR reference gene database is employed. The experimental results reported that the perplexity provided the optimal number of topics from the AMR gene data of more than 6500 samples. Therefore, the proposed model helps to find the appropriate antibiotic for bacterial and viral spread and discover how to increase the proper antibiotic in human bodies

An Optimization Method for Enterprise Resource Integration Based on Improved Particle Swarm Optimization.

An enterprise's development and growth are inextricably linked to rational and efficient resource integration and optimization. This study focuses on the reorganization and integration of industrial elements inside the firm from the standpoint of resource integration. The ideal resource integration strategy is investigated by integrating the industrial parts of a certain enterprise in order to increase the efficiency of project completion and lower enterprise expenses. The enterprise's internal material and human resources are limited, but it is frequently necessary to execute numerous activities at the same time, and each activity must meet multiple goals. This research investigates how to properly integrate and schedule resources while attaining different goals. This research proposes using an enhanced particle swarm optimization technique (IPSO) to combine firms' internal resources. In order to address the issue of uneven particle dispersion caused by random population initialization, IPSO incorporates chaos theory into particle population initialization. The logistic mapping sequence generates a huge number of particles, and the particles with the highest quality are chosen for initialization. This can increase particle quality, allowing particles to be spread equally during setup. In the late stage, the classic particle swarm optimization algorithm (PSO) has a slow convergence rate, causing the algorithm to readily slip into a local optimal solution. This research proposes a dynamic inertia weight update approach based on fitness value. In the later stages of the algorithm, this strategy can improve the convergence speed and quality of the global optimal solution, allowing the particles to do a global search and eventually identify the population's ideal solution. Furthermore, IPSO creates a fitness function depending on task completion time. IPSO is used to test the performance of an enterprise's resource integration case. Experiments show that the method utilized can swiftly locate the ideal solution, complete the integration, and optimization of enterprise resources in the shortest job completion time, and for the least amount of money.

EXPLORING LABEL INITIALIZATION FOR WEAKLY SUPERVISED ALS POINT CLOUD SEMANTIC SEGMENTATION

Abstract. Although a number of emerging point-cloud semantic segmentation methods achieve state-of-the-art results, acquiring fully interpreted training data is a time-consuming and labor-intensive task. To reduce the burden of data annotation in training, semiand weakly supervised methods are proposed to address the situation of limited supervisory sources, achieving competitive results compared to full supervision schemes. However, given a fixed budget, the effective annotation of a few points is typically ignored, which is referred to as weak-label initialization in this study. In practice, random selection is typically adopted by default. Because weakly supervised methods largely rely on semantic information supplied by initial weak labels, this studies explores the influence of different weak-label initialization strategies. In addition to random initialization, we propose a feature-constrained framework to guide the selection of initial weak labels. A feature space of point clouds is first constructed by feature extraction and embedding. Then, we develop a density-biased strategy to annotate points by locating highly dense clustered regions, as significant information distinguishing semantic classes is often concentrated in such areas. Our method outperforms random initialization on ISPRS Vaihingen 3D data when only using sparse weak labels, achieving an overall accuracy of 78.06% using 1‰ of labels. However, only a minor increase is observed on the LASDU dataset. Additionally, the results show that initialization with category-wise uniformly distributed weak labels is more effective when incorporated using a weakly supervised method.

Improving image classification robustness using self‐supervision

Self‐supervised learning allows training of neural networks without immense, high‐quality or labelled data sets. We demonstrate that self‐supervision furthermore improves robustness of models using small, imbalanced or incomplete data sets which pose severe difficulties to supervised models. For small data sets, the accuracy of our approach is up to 12.5% higher using MNIST and 15.2% using Fashion‐MNIST compared to random initialization. Moreover, self‐supervision influences the way of learning itself, which means that in case of small or strongly imbalanced data sets, it can be prevented that classes are not or insufficiently learned. Even if input data are corrupted and large image regions are missing from the training set, self‐supervision significantly improves classification accuracy (up to 7.3% for MNIST and 2.2% for Fashion‐MNIST). In addition, we analyse combinations of data manipulations and seek to generate a better understanding of how pretext accuracy and downstream accuracy are related. This is not only important to ensure optimal pretraining but also for training with unlabelled data in order to find an appropriate evaluation measure. As such, we make an important contribution to learning with realistic data sets and making machine learning accessible to application areas that require expensive and difficult data collection.

Interference Nulling Using Reconfigurable Intelligent Surface

This paper investigates the interference nulling capability of reconfigurable intelligent surface (RIS) in a multiuser environment where multiple single-antenna transceivers communicate simultaneously in a shared spectrum. From a theoretical perspective, we show that when the channels between the RIS and the transceivers have line-of-sight and the direct paths are blocked, it is possible to adjust the phases of the RIS elements to null out all the interference completely and to achieve the maximum <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> degrees-of-freedom (DoF) in the overall <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> -user interference channel, provided that the number of RIS elements exceeds some finite value that depends on <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$K$ </tex-math></inline-formula> . Algorithmically, for any fixed channel realization we formulate the interference nulling problem as a feasibility problem, and propose an alternating projection algorithm to efficiently solve the resulting nonconvex problem with local convergence guarantee. Numerical results show that the proposed alternating projection algorithm can null all the interference if the number of RIS elements is only slightly larger than a threshold of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2K(K-1)$ </tex-math></inline-formula> . For the practical sum-rate maximization objective, this paper proposes to use the zero-forcing solution obtained from alternating projection as an initial point for subsequent Riemannian conjugate gradient optimization and shows that it has a significant performance advantage over random initializations. For the objective of maximizing the minimum rate, this paper proposes a subgradient projection method which is capable of achieving excellent performance at low complexity.

Solving phase retrieval with random initial guess is nearly as good as by spectral initialization

The problem of recovering a signal x∈Rn from a set of magnitude measurements yi=|〈ai,x〉|,i=1,…,m is referred as phase retrieval, which has many applications in fields of physical sciences and engineering. In this paper we show that the smoothed amplitude flow based model for phase retrieval has benign geometric structure under the optimal sampling complexity. In particular, we show that when the measurements ai∈Rn are Gaussian random vectors and the number of measurements m≥Cn, our smoothed amplitude flow based model has no spurious local minimizers with high probability, i.e., the target solution x is the unique global minimizer (up to a global phase) and the loss function has a negative directional curvature around each saddle point. Due to this benign geometric landscape, the phase retrieval problem can be solved by the gradient descent algorithms without spectral initialization. Numerical experiments show that the gradient descent algorithm with random initialization performs well even comparing with state-of-the-art algorithms with spectral initialization in empirical success rate and convergence speed.

Fast costate estimation for indirect trajectory optimization using Bezier-curve-based shaping approach

In this study, a fast costate estimation method derived from the Bezier-curve-based shaping approach is proposed for indirect low-thrust transfer trajectory optimization. The trajectory is designed rapidly by means of the shape-based method, which can be discretized on the Gauss points and substituted into the discrete first-order optimal condition equation for costate estimation. The estimated costate at the starting point is substituted into the indirect method to solve the optimal control problem. The costate estimation performance for two interplanetary rendezvous missions from Earth to Mars and Earth to Dionysus is compared via random initialization and Newton's iteration method. The initial guesses for the transfer time are given by the Hohmann transfer and the Lambert two-pulse traversal search. The results show that the Lambert traversal search can provide a more accurate launch time estimation in the Bezier-curve-based shaping approach, which can estimate the costate for indirect trajectory optimization in an order of magnitude less time than that required by random initialization.

Maneuvering penetration strategies of ballistic missiles based on deep reinforcement learning

In this paper, a ballistic missile terminal penetration scenario is studied, which contains three participants: target, missile, and defender. The ballistic missile attempts to hit the target while evading the defender. A maneuvering penetration guidance strategy that balances both the guidance accuracy and penetration capability is proposed through deep reinforcement learning. Reward shaping and random initialization are applied to improve training speed and generalization, respectively. The proposed strategy is developed based on the twin delayed deep deterministic policy gradient algorithm. It directly maps observations to actions and is an end-to-end guidance scheme that does not require an accurate model. The simulation results show that the proposed strategy has higher penetration probabilities than conventional strategies for different initial heading errors and even for defenders with different guidance laws, which indicates its good robustness and generalization. For different initial heading errors, it has learned different maneuvering modes and has certain intelligence. In addition, it is computationally small, does not consume much memory, and can be easily applied on modern flight computers.

Multi-Sonar Distributed Fusion for Target Detection and Tracking in Marine Environment.

The multi-sonar distributed fusion system has been pervasively deployed to jointly detect and track marine targets. In the realistic scenario, the origin of locally transmitted tracks is uncertain due to clutter disturbance and the presence of multi-target. Moreover, attributed to the different sonar internal processing times and diverse communication delays between sonar and the fusion center, tracks unavoidably arrive in the fusion center with temporal out-of-sequence (OOS), both problems pose significant challenges to the fusion system. Under the distributed fusion framework with memory, this paper proposes a novel multiple forward prediction-integrated equivalent measurement fusion (MFP-IEMF) method, it fuses the multi-lag OOST with track origin uncertainty in an optimal manner and is capable to be implemented in both the synchronous and asynchronous multi-sonar tracks fusion system. Furthermore, a random central track initialization technique is also proposed to detect the randomly born marine target in time via quickly initiating and confirming true tracks. The numerical results show that the proposed algorithm achieves the same optimality as the existing OOS reprocessing method, and delivers substantially improved detection and tracking performance in terms of both ANCTT and estimation accuracy compared to the existing OOST discarding fusion method and the ANF-IFPFD method.