In order to enhance the performance of Dynamic Differential Evolution(DDE) for solving high dimensional optimization problems,an Orthogonal Dynamic Differential Evolution(ODDE) algorithm is proposed.ODDE is based on the framework of DDE algorithm,so it has very powerful global search ability.At the same time,orthogonal crossover operator based on orthogonal experiment design method is used to enhance the local search ability of algorithm.Nine commonly used benchmark problems with different dimensional size 30,100,300 and 500 are used to evaluate the performance of ODDE which is compared with Differential Evolution(DE),DDE,Orthogonal Crossover Differential Evolution(OXDE) algorithm.Numerical results show that the performance of solving accuracy and convergence rate of ODDE is superior to other algorithms,it can be widely used to solve high-dimensional optimization problems in engineering application.
Microblog is a new type of news media,and its influence and propagation speed surpasses traditional major media.Therefore,it has a great importance to predict hotness in microblog for public opinion monitoring,government propaganda,corporation marketing and popular issues pushing.Through analyzing microblog forward level which combining the effects of the forward index,forward depth and breadth index,this paper gives a new definition of calculating the hotness index of microblog.Then depend on this definition,the hotness index of the microblog is classified as five levels.The goal is to predict the hotness of microblog whose repost count is over 100 to achieve a specified level.By using supervised machine learning algorithm,it successively extracts the static attributes and dynamic repost characteristics of the training samples to train hotness prediction model.The training samples is from Sina microblog is caught by using self-developed BigData open crawler platform.Experimental result by using 10-fold cross-validation shows that,compared with hotness prediction model based on static attributes,the model with dynamic features can effectively improve the prediction performance,and F1-measure achieves 76.9%.
The hotness trend modeling of topic in online forums is one of the main contents of the existing public opinion analysis.Under the situation that the existing methods ignore the influence of opinion tendency on the topic hotness,a hotness trend modeling approach for the topic is proposed considering opinion tendency.The sentimental tendency of topics can be obtained by sentimental classification methods.This sentimental information is leveraged over the hotness computation,which flexibly reflects the influence of sentimental tendency on topic hotness.Gamma distribution is used to fit the hotness curve.Experimental results show that the proposed method can fit the hotness trend more accurately compared with the Gauss model.
For the Garbage Collection(GC) low efficiency in embedded virtual machine environment,this paper proposes an improved generational GC algorithm based on lifespan prediction.Through the prediction of object’s lifespan,objects predicted to be long-lived are allocated directly into the old generation,and the need to copy such objects from the young generation is eliminated,thereby reducing the execution time of GC.In young generation,this paper adopts a kind of un-stop-the-world strategy which objects allocation and promotion perform concurrently.In old generation,it uses lazy-buddy algorithm combining with mark-sweep algorithm to achieve fast allocation and recovery.It not only avoids the copy operation,but also controls the amount of memory fragmentation.Experimental results show that,with this algorithm,the GC time decreases by about 23.9% and the program running time decreases by about 17.2%,the overall system execution performance is significantly improved.
For the larger intra-class variance and inadequate real-time problems caused by factors of imaging scales difference in nighttime pedestrian detection,this paper designs a rapid dentify program for nighttime pedestrians based on entropy weight and header calibration of Fast Classification Support Vector Machine(FCSVM) optimization under the application of statistical learning principles.The program utilizes entropy weight to improve the feature of gradient histogram,introduces three branch structure SVM to identify the target further,and uses rapid classification FCSVM to reduce the overhead required computation and to ensure real-time.Through the header calibration method to analyze and assess error detection goals,it further improves the accuracy of image matching.Experimental results show that the scheme can distinguish far infrared pedestrian goals effectively at night environment,and have good recognition effect in urban,suburban and other different application environments on the basis of ensuring pedestrian real-time fully.