本文選題：速度趨勢 + SVR　； 參考：《北京郵電大學(xué)》2017年碩士論文

【摘要】：隨著互聯(lián)網(wǎng)快速成熟的發(fā)展,從互聯(lián)網(wǎng)獲取信息已經(jīng)成為人們?nèi)粘Ｉ钪蝎@取信息的重要渠道之一。伴隨著2G網(wǎng)絡(luò),3G網(wǎng)絡(luò),4G網(wǎng)絡(luò)的逐代更新,網(wǎng)絡(luò)的訪問速度也在不斷的提升,近年來國家也發(fā)布了關(guān)于提升網(wǎng)絡(luò)速度的相關(guān)方案。在實際的應(yīng)用中網(wǎng)絡(luò)的訪問會經(jīng)常隨著不同的訪問時間而呈現(xiàn)出速度上的巨大差異,甚至由于此類的原因?qū)е孪嚓P(guān)資源無法及時獲取到。因此,我們需要提供一種監(jiān)控手段對網(wǎng)絡(luò)的實際狀況進(jìn)行評估和反饋,對接下來網(wǎng)絡(luò)的穩(wěn)定狀態(tài)有一個清晰的認(rèn)識。目前機(jī)器學(xué)習(xí)算法被廣泛的應(yīng)用到各個領(lǐng)域來解決分類和回歸問題,常見的機(jī)器學(xué)習(xí)算法諸如支持向量回歸(SVR)和神經(jīng)網(wǎng)絡(luò)都發(fā)展的很成熟。本文應(yīng)用機(jī)器學(xué)習(xí)算法對網(wǎng)絡(luò)的速度趨勢進(jìn)行預(yù)測。將機(jī)器學(xué)習(xí)算法應(yīng)用到網(wǎng)絡(luò)速度趨勢預(yù)測當(dāng)中一方面提供了預(yù)測的科學(xué)性和理論性,另一方面也促使了機(jī)器學(xué)習(xí)算法多分支快速的發(fā)展。本文的大工作如下:1、設(shè)計了一套網(wǎng)絡(luò)速度采集模型,該采集模型會對指定運營商提供的網(wǎng)絡(luò)速度進(jìn)行采集,采集后的網(wǎng)絡(luò)速度作為趨勢預(yù)測的元數(shù)據(jù)。該采集模型對采集的功能模塊進(jìn)行了分割,采用云端對采集服務(wù)器進(jìn)行管理和任務(wù)的分配,一方面減小了單一服務(wù)器負(fù)責(zé)全部任務(wù)的負(fù)載,另一方面也方便了日后采集服務(wù)器數(shù)量的擴(kuò)展。2、提出了針對本課題的輸入向量的選取方式,選取輸入向量的方式有很多,本課題從前兩個月的網(wǎng)絡(luò)速度數(shù)據(jù)中提取一部分作為預(yù)測第三個月網(wǎng)絡(luò)速度數(shù)據(jù)的輸入。不同的選取方式對預(yù)測的效果影響很大,在實驗對比的基礎(chǔ)上最終確定了六維的輸入向量。3、綜合對比了 PSO優(yōu)化的SVR和神經(jīng)網(wǎng)絡(luò)的預(yù)測效果,其中對于神經(jīng)網(wǎng)絡(luò)的選取沒有固定在特定的結(jié)構(gòu)上,而是在不同的實驗基礎(chǔ)上采用不同結(jié)構(gòu)的神經(jīng)網(wǎng)絡(luò)進(jìn)行綜合對比,最終確定預(yù)測效果最好的神經(jīng)網(wǎng)絡(luò)作為與PSO優(yōu)化的SVR對比的參考。4、本文的主題是網(wǎng)絡(luò)速度趨勢的預(yù)測,趨勢預(yù)測本身是一個分類的問題,本文采用準(zhǔn)確率,召回率和F1-socre作為趨勢預(yù)測的一個評估標(biāo)準(zhǔn)。除此之外,本文還采用回歸的方式對網(wǎng)絡(luò)的速度進(jìn)行預(yù)測,這里提供了網(wǎng)絡(luò)速度值的參考,以MAE, MSE, MAPE作為衡量的標(biāo)準(zhǔn),在訓(xùn)練的過程中以MSE作為適應(yīng)度函數(shù)。5、綜合以上實現(xiàn)了一個完整的網(wǎng)絡(luò)速度趨勢預(yù)測系統(tǒng),并在實際的數(shù)據(jù)基礎(chǔ)上進(jìn)行了實驗驗證和性能的測試,完成了課題提出的目標(biāo)。
[Abstract]:With the rapid development of the Internet, obtaining information from the Internet has become one of the important channels for people to obtain information in their daily life. With the generation update of 2G network and 3G network, the access speed of the network is improving constantly. In recent years, the country has also issued the related plan to improve the network speed. In practical applications, network access often presents a huge difference in speed with different access times, even due to such reasons, related resources can not be obtained in time. Therefore, we need to provide a monitoring means to evaluate and feedback the actual situation of the network, and have a clear understanding of the stability of the network. At present machine learning algorithms are widely used in various fields to solve classification and regression problems. Common machine learning algorithms such as support vector regression (SVR) and neural networks are developed very mature. In this paper, the machine learning algorithm is used to predict the speed trend of the network. The application of machine learning algorithm to network speed trend prediction not only provides scientific and theoretical prediction, but also promotes the rapid development of multi-branch machine learning algorithm. The main work of this paper is as follows: 1. A set of network speed acquisition model is designed. The collection model will collect the network speed provided by the designated operator and the network speed will be used as the metadata to predict the trend. The collection model divides the function module of the collection and uses the cloud to manage and distribute the tasks of the collection server. On the one hand, it reduces the load of the single server which is responsible for all the tasks. On the other hand, it also facilitates the expansion of the number of collection servers in the future. 2. The selection of input vectors for this topic is proposed. There are many ways to select input vectors. This paper extracts part of the network speed data from the first two months as input to predict the third month network speed data. Different selection methods have great influence on the effect of prediction. On the basis of experimental comparison, the six-dimensional input vector .3is finally determined, and the prediction effect of PSO optimized SVR and neural network is compared synthetically. The selection of neural network is not fixed on the specific structure, but on the basis of different experiments, the neural network with different structure is used for comprehensive comparison. Finally determine the best prediction effect of neural network as a reference compared with PSO optimized SVR. The theme of this paper is network speed trend prediction, trend prediction itself is a classification problem, this paper adopts accuracy. Recall rates and F 1-socre are used as an evaluation criterion for trend forecasting. In addition, this paper also uses regression method to predict the speed of the network, which provides a reference for the network speed value, with mae, MSE, MAPE as the measurement standard, In the process of training, MSE is taken as the fitness function. 5, a complete network speed trend prediction system is realized, and the experimental verification and performance test are carried out on the basis of the actual data, and the target proposed by the project is completed.
【學(xué)位授予單位】：北京郵電大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP18;TP393.0

【參考文獻(xiàn)】

相關(guān)博士學(xué)位論文 前2條

1 田野;基于社會化媒體的話題檢測與傳播關(guān)鍵問題研究[D];北京郵電大學(xué);2013年

2 韓毅;社會網(wǎng)絡(luò)分析與挖掘的若干關(guān)鍵問題研究[D];國防科學(xué)技術(shù)大學(xué);2011年

，

本文編號：1998206