311 results about "Algorithms performance" patented technology

The invention provides an intelligent routing decision method based on reinforcement learning, in particular to an intelligent routing decision method based on a DDPG reinforcement learning algorithm.The method aims at designing an intelligent routing decision by utilizing reinforcement learning, balancing an equivalent path flow load and improving the processing capacity of a network for burst flow, an experience decision mechanism based on a sampling probability is adopted, the probability that experience with poorer performance is selected is higher, and the training efficiency of an algorithm is improved. In addition, noise is added into neural network parameters, system exploration is facilitated, and algorithm performance is improved. The method comprises the following steps: 1) constructing a network topology structure; 2) numbering equivalent paths in the network topology structure G0; (3) constructing a routing decision model based on a DDPG reinforcement learning algorithm,(4) initializing a flow demand matrix DM and an equivalent path flow proportion matrix PM, and (5) carrying out iterative training on the routing decision model based on reinforcement learning, and the method can be used for scenes such as a data center network.

The invention discloses a TF-IDF feature extraction based short text classification method. According to the method, short texts are merged into a long text so as to enhance the TF-IDF feature of the short texts; dimension reduction is performed so as to generate a feature word list and a feature word dictionary; a mechanism compensation is established for a class having a relative unobvious feature while the feature word list is established, and the text feature vector weight is enhanced; and other word banks or word vector dictionaries do not have to be constructed or trained, and then the algorithm performance can be improved on the premise of ensuring the feature expression result of the texts. The TF-IDF feature extraction based short text classification method can be widely applied to the field of data processing.

The invention provides a license plate recognition and positioning method based on a deep neural network, and mainly solves the problem of inaccurate license plate recognition and positioning in a complex scene in an existing algorithm. Firstly, a license plate data set meeting specific requirements of license plate detection is established; an anchor frame is generated by using a K-means clustering algorithm, a license plate detection deep convolutional neural network structure is established by combining machine learning and introducing an attention mechanism, a network model is trained by using an established license plate data set, and an Adam algorithm is used as an optimization algorithm in the training process; testing the model by adopting the detection accuracy when the cross-parallel ratio IOU is equal to 0.8 as a measurement index of algorithm performance and adopting a HyperLPR algorithm and a mathematical morphology method as a comparison algorithm. Compared with the priorart, the license plate recognition and positioning method based on the deep neural network has the advantages that a channel attention mechanism is added, so that the detection accuracy is higher, the speed is higher, and the robustness to the environment is very high.

The invention belongs to the field of positioning a cell phone base station, and relates to a method for enhancing cell phone base station positioning precision by using Hidden Markov map-matching algorithm. The cell phone base station positioning precision is enhanced by using Hidden Markov map-matching algorithm. The probability of matching a set point to a candidate matching road by the algorithm is called initial matching probability, the probability of conversion of the candidate matching road is called road transfer probability, and the algorithm rapidly finds the path with maximal product of the initial matching probability and the road transfer probability through dynamic programming. Since a vehicle may not change roads frequently during an actual driving process, constraints are introduced in the algorithm for calculating the set point road transfer probability to improve the algorithm performance.

The invention discloses a method for integrating a covering method to obtain a service node in the SDN based on a greedy algorithm. The method comprises the steps that (1), the balance indexes of nodes are determined, and comprehensive ability values of the nodes in an autonomous domain are obtained according to a data fuzzy normalization method; (2), the node with the highest comprehensive ability value in the covering autonomous domain is selected through the greedy algorithm; (3), the node with the highest comprehensive ability value is determined as the service node, and adjacent nodes become sub-nodes of the service node; (4), a set, composed of the service node and the sub-nodes, in the network is removed; (5), the step (2), the step (3) and the step (4) are carried out repeatedly until no candidate network node exists in the network, wherein the determined service node is the final service node of the SDN. According to the method for integrating the covering method to select the service node, every time when the service node is selected, nodes adjacent to the service node do not need to participate in later selection, and thus the algorithm has the good performance in time complexity.

The invention discloses a social contact-oriented crowdsourcing task allocating method and system. The method comprises the steps of carrying out matching degree estimating algorithm calculations on workers and a crowdsourcing task according to the crowdsourcing task published by a task publisher to obtain a group of workers with the highest crowdsourcing task matching degree; calculating the taskallocating by adopting a greedy algorithm and selecting a group of workers with the largest sum of overall matching degrees from the group of workers with the highest crowdsourcing task matching degree to serve as a final allocation result to ensure that each task has different to-be-assigned workers; and waiting for the workers to receive tasks and completing the crowdsourcing task allocating. The method has the advantages that the reference relationship of information of interests and the like of task recommendation is increased to ensure that the task recommendation is more efficient and more accurate; and the algorithm performance is increased and the quality of crowdsourcing services is increased.

The invention discloses a method for taping the internet hot-point subjects correlation. The invention extracts the hot-point subject key words through the search engine inquiry log, and conducts the model building analysis on correlation degree among the hot-point subject key words. The effective Hash method shall be adopted to build a sparse matrix, so as to increase the algorithm efficiency. The invention can conduct the incrementation processing to newly increased hot-point subject key words and data, and is convenient for realization of distributed processing. The algorithm only conducts one-time scan to the data source, so as to update the corresponding sparse matrix area, and finally conduct the sequencing to obtain the sequence of correlation degree among the hot-point subject key words. The invention can accurately and speedily pick out the correlation among the hot-point subjects, thereby overcoming the problems that the prior algorithm performance is low, particularly that the processing effect on the newly increased hot-point subject key words is poor, which can better satisfy the performance requirements recommended by large-scale hot-point subject in internet.

The invention belongs to the wireless communication signal detection technical field and relates to a block compressive sensing non-orthogonal multiple-address system multiuser detection method. The invention provides a sparseness adaptive detection method which can recombine structured sparseness into block sparseness. According to the block compressive sensing non-orthogonal multiple-address system multiuser detection method of the invention, a block sparseness-adaptive sub-space tracking algorithm is adopted for recombined block sparseness signals; according to the algorithm, a noise power setting threshold is utilized to realize the adaptive evaluation of the sparseness of users, and SER (Symbol Error Rate) performance is improved greatly compared with that of the prior art under a condition that the a priori information of sparseness degree is not required; and a simulation result indicates that the performance of the algorithm can reach an optimal theoretical value after a signal-to-noise ratio is increased to a certain extent.

The invention relates to a classification and recognition method for P300 event-related potential based on deep learning, and belongs to the technical field of medical and physiological signal detection and processing analysis. According to the invention, a Butterworth filter is adopted to successively perform high-pass filtering and low-pass filtering on original signals to remove artifacts and power frequency interference; the data is amplified by using a one-time superposition averaging technique, normalization and time domain truncation are performed on EEG signals, and corresponding supervisory signals are formulated according to the signal types; the EEG data is divided into a training set and a verification set after data preprocessing, a deep learning network capable of classifyingand recognizing the P300 event-related potential is constructed, and the feature extraction ability of the network is improved; the probability that an input signal contains the P300 event-related potential is finally predicted through the trained network; and finally, a target character is predicted according to an experimental paradigm and the probability value outputted by the network. The experiment shows that the algorithm is good in performance and can also obtain good character recognition accuracy under the condition of reducing the number of experiments.

The invention relates to a multi-temporal unmanned aerial vehicle video image change area detection and classification method, and belongs to the field of image processing computer vision. The methodspecifically comprises the following steps: detecting a change area of a homologous image, performing multi-source image registration and fusion, and performing change area classification. The methodis based on unmanned aerial vehicle video image characteristics and the defects of the prior art in the aspect of unmanned aerial vehicle video image change area detection and classification. The invention provides the multi-temporal unmanned aerial vehicle video image change area detection and classification method by comprehensively planning algorithm performance and adaptability, and solves a plurality of problems in the aspect of variable reconnaissance of ground object state change of a certain target area by an unmanned aerial vehicle video processing system.

The invention provides a depth pedestrian re-identification method based on positive sample balance constraint. A residual error network employed in the method is simple in structure and can be widely used, and the network structure which is deep sufficiently improves the feature representation capability. Moreover, there is no need to specially design the network structure. A residual error classifier is used for feature extraction of an image, so the accuracy of pedestrian re-identification can be greater than the accuracy of most of well-designed methods. Compared with two-tuple loss and three-tuple loss methods, the method does not need to intentionally generate an effective sample for improving the structural loss and can achieve the similar effect. Moreover, the method enables the learned gradient direction to be more robust and effective through the overall distribution information. On the basis of improving the structural loss, the method improves the positive sample balance constraint, can control the distance of a positive sample pair, also can balance the gradients of the distance of the positive sample pair and the distance of a positive sample pair, enables the algorithm to be easier to train, and improves the performances of an algorithm.

The invention discloses a reinforcement learning-based named data network congestion control method and belongs to the network information transmission communication technical field. According to thereinforcement learning-based named data network congestion control method of the invention, from the perspective of a forwarding strategy, the process of the forwarding of an interest packet by a routing node is mapped into a Markov decision-making process; with minimum response time adopted as an objective, intra-network buffer and multi-path forwarding in an NDN (named data network) are considered; a reinforcement learning method is used to solve an optimal strategy under a condition that any assumptions about link capacity or packet size are not made; in the reinforcement learning solutionmethod, a Sarsa algorithm based on eligibility traces, namely the Sarsa(lambda) algorithm, is used to realize theoretically good algorithm performance; a dynamic forwarding policy is adopted to intelligently select a forwarding port, and therefore, a condition that traffic is sent to a congested link can be avoided as much as possible, and congestion can be actively avoided and alleviated.

The invention discloses a path planning method and system for an unmanned vehicle. The method comprises the following steps: randomly generating a plurality of paths from a starting point to a targetpoint, wherein the paths are obtained by connecting a plurality of path points; encoding the path by using the serial numbers of the path points through which the path sequentially passes from the starting point to the target point; dividing paths with the same number of path points into the same sub-population; performing parallel genetic operation on each sub-population; selecting the sub-population with the optimal moderate value from the sub-populations after genetic manipulation, and recording the sub-population as the optimal sub-population; and performing genetic manipulation on each individual in the optimal sub-population, determining an optimal path in combination with the fitness value of each individual in the optimal sub-population, wherein the fitness function, the crossoverprobability function and the mutation probability function in the method are all adaptive functions. According to the method, the algorithm can be prevented from falling into local optimum, the algorithm performance is further improved, the adaptability is higher, and the use experience of vehicle passengers is effectively improved.

The invention discloses an algorithm integration and evaluation platform and method based on SLURM scheduling. The platform comprises a packaging module, a scheduling interface module, an uploading and downloading module, a compiling module, an algorithm integration module and an algorithm performance statistics module. In the running process of an SLURM center daemon process and a monitoring process, a user can dynamically conduct scheduling method integration through the platform. When using the platform, the user only needs to know about a public variable, a structural body and a foundation function library file provided by the packaging module of the platform to achieve the aim of submitting two external sub-function interfaces of the platform to a server, and does not need to care about source codes of other parts of software or the cooperative relationship between all the modules in the platform, and therefore an SLURM developer or a high-performance computing user can more conveniently integrate scheduling algorithms and ignore research of the software on other module source codes and can detect the performance of the algorithms under a real environment and flexibly use the various scheduling algorithms.

The invention provides an integrated system which can realize the synthesis integration of various algorithms, demonstrate the overall effect of a scene, and verify the performance of the algorithms. The system generates data through a calculation function, integrates the algorithms through a frame function, and shows the result through a display function. Parameter testing, performance testing, function verification and effect demonstration are realized by constructing a unified environment which shows uniform size and practicable project in all procedures from data generation to algorithm calculation, and the problems of various kinds of interfaces, repeated code editing and inconsistent performance and effect during traditional algorithm integration are solved. The technical scheme of the invention is a demonstration and verification integrated system based on algorithm synthesis integration, comprising a version algorithms base management module, a scene and process setting module, a system main control module, a group communication module, an analysis processing frame module, a data access and management module and a database module.

The invention relates to a method and a system for realizing remote sensing image target detection based on a deep neural network, and a storage medium thereof, so as to realize detection of horizontal and rotary arrangement targets of a remote sensing image. According to the method, an anchor point box generation module is designed, an anchor point box is generated in a self-adaptive mode throughfeature information of different positions, and the influence of the difference of preset anchor point boxes on detection precision is reduced; aiming at the characteristic that more small targets exist in a remote sensing image, an improved feature pyramid structure is provided, and deep and shallow layer feature information is fused by adopting a transposed convolution method; aiming at difficulties such as complex background of a remote sensing image, a receptive field expanding module is adopted to extract more characteristic information, and the detection precision of a small target under a complex background is improved; a SmoothLn function is adopted as regression loss, so that the algorithm performance is further improved; for a rotation arrangement target, regression of an anglefactor is introduced to realize rotation frame detection. In addition, in order to facilitate the use of a user, the remote sensing image target detection system designed by the invention has the functions of horizontal frame and rotary frame detection and result statistics.

The invention discloses an intelligent evaluation and diagnosis method and a system for heart disease types and severity degrees. The method comprises the steps of acquiring disease characteristic data and demographic characteristic data, and analyzing the acquired ultrasonic echocardiogram report data and the patient demographic characteristic data by utilizing a learning model to obtain a modelevaluation index, a heart disease type and a heart disease severity. According to the invention, a data mining method is adopted, so that data preprocessing, data screening and other operations are carried out on data through the data mining correlation method. The method is adopted for selecting a noise ratio during the characteristics selection process. A random forest model is adopted for carrying out the classification prediction of the heart disease severity. Meanwhile, an effective research method is obtained through comparing and analyzing the algorithm performances and the learning effects of the random forest model, a naive Bayes classifier, a decision tree model and a BP neural network model. Moreover, a standard for the severity classification of heart disease patients and a prediction method for predicting the treatment risk of the heart disease operation are provided.

The invention relates to the technical field of data compression method, and discloses a precision-controlled self-adaptive data compression method. The compression method comprises the steps of: step A, carrying out a compression processing of data by a revolving door; step B, judging whether a self-adaptive calculation adjustment of the threshold width is needed or not; if a self-adaptive calculation adjustment of the threshold width is needed, carrying out a self-adaptive calculation adjustment of the threshold width; otherwise, carrying out compression processing of next data. The compression method of the invention adjusts the threshold width gradually by introducing feedback of standard errors, thereby assists in avoiding decrease of compression performance caused by artificial blind setting of threshold width value, and assists in avoiding a process of repeated massive tests. In addition, data validity judgments and dynamically adjusted data compression time window parameters are added, influence of exception data on algorithm performance is reduced. For slowly changing steady-state data, time interval of self-adaptive calculation is changed dynamically, pointless calculating actions are reduced, and compression performance is improved.

The invention provides a single-piece job-shop scheduling method based on multi-Agent deep reinforcement learning, aiming at the characteristics that the single-piece job-shop scheduling problem is complex in constraint and various in solution space types, and the traditional mathematical programming algorithm and meta-heuristic algorithm cannot meet the quick solution of the large-scale job-shopscheduling problem. The method comprises the following steps: firstly, designing a communication mechanism among multiple Agents, and carrying out reinforcement learning modeling on a single job shopscheduling problem by adopting a multi-Agent method; secondly, constructing a deep neural network to extract a workshop state, and designing an operation workshop action selection mechanism on the basis of the deep neural network to realize interaction between a workshop processing workpiece and a workshop environment; thirdly, designing a reward function to evaluate the whole scheduling decision,and updating the scheduling decision by using a PolicyGraphic algorithm to obtain a more excellent scheduling result; and finally, performing performance evaluation and verification on the algorithmperformance by using the standard data set. The job shop scheduling problem can be solved, and the method system of the job shop scheduling problem is enriched.

The invention relates to designing and testing technology field of very large scale integrating circuit, namely equivalence testing method. The method uses satisfiability arithmetic as the engine, tests the whole circuit design, the method has two segments, the first segment is to determine the candidate equal signal in the two circuits, and through the circuits structure analysis, carries on static sift to the candidate equal signal, the second segment is to realize gain satisfiability method through clause group, and tests the equivalence of the candidate signal. The main character of the invention is: 1) selects the candidate equal signal in the circuit through circuit structure analysis, selects the cut or centralized equal signal dynamically in the process of test, and guarantees the signal is independent from each other. 2) Realizes the satisfiability arithmetic through grouping the clause saves the calculating resources, upgrades the arithmetic performance and the processing ability.

System, method and computer product for incremental improvement of algorithm performance during algorithm development. In one aspect of this disclosure, there is a system, method and computer readable medium that stores instructions for instructing a computer system, to assess the performance of an algorithm during development. In this embodiment, a design of experiments component establishes an acceptable number of experiments for analyzing incremental improvements of the algorithm. An experiment performance component runs the established number of experiments for the algorithm. A simulation component simulates the behavior of the algorithm using results from the experiment performance component.

The embodiment of the invention discloses a text clustering method, device and equipment based on density and a storage medium, and relates to the technical field of text data analysis. The method comprises the steps of receiving a target data set; determining a target distance formula; generating a distance matrix about the whole target data set; calculating the local density of each data point;separately extracting the minimum value of the distance between each data point and each data point in the sample point set, and recording the minimum value as the minimum point distance; establishinga clustering decision diagram according to the local density and the minimum point distance; determining the number of class clusters and a class cluster center in the clustering decision diagram; and dividing each data point into class clusters of the clustering decision diagram. According to the method, in the whole clustering process, the non-spherical data can be clustered only by calculatingthe distance between the sample points once without iterative calculation, the algorithm performance is greatly improved, the clustering decision diagram is used for scientifically selecting the number of the class clusters, and the situation that the number of the class clusters is manually set without basis is avoided.