An intelligent routing method based on network partition

Abstract

New routing strategies are urgently needed to control exploding traffic and provide precise routing for the next-generation complex network. Traditional routing methods cannot satisfy its routing requirements of this complex traffic control situation. One of the important reasons is that it does not consider learning from traffic features. Using deep learning method to predict routing path is an emerging and promising solution, however, the existing deep learning-based methods in network traffic control field still have some disadvantages, such as low routing accuracy, high routing time complexity and need large data set to train complex deep learning system. To solve these problems, this paper proposes a block-based deep learning intelligent routing strategy (DLBR strategy), which divides the network into multiple sub-blocks according to a recursive partition method and uses three deep learning models to train and test them. Experiments show that the proposed network DLBR strategy has the ability to combine with different deep learning intelligent structures, and achieves higher accuracy and lower time complexity under the training of smaller training data.

Introduction

In the next-generation of wired and wireless communication network, network traffic will increase enormously with the growing number of users and the rapid increase in connected devices. Network routing strategy is urgently needed to control traffic and provide excellent routing results. However, traditional routing protocols typically meet routing requirements based on characteristics of the network such as using delay, hops to search for the shortest path [1], without learning from the different features of traffic. In the background of modern complex networks, it is difficult to meet the different control requirements of diverse flows.

In recent years, deep learning [2] has made breakthrough in a variety of application areas, such as computer vision, speech recognition and text categorization. So why deep learning is favored by academia and industry? On the one hand, the deep learning structure has a powerful ability to learn complex relationships automatically. It achieves significant results through the expression of some nonlinear complex relationships and the capture of hidden features by unique deep neural network structures. On the other hand, the deep learning method is a promising end-to-end learning model whose structure consists of only three parts: input, deep learning network and output, which is easy to implement.

Influenced by the breakthrough progress of deep learning in other fields, researchers are beginning to use deep learning method to solve problems in the field of network, such as traffic control [3], however, it still suffers from some disadvantages:

• Next-generation networks are often complex networks with multi-dimensional features control requirements, which require large scale data set to train and adjust parameters for complex deep learning systems. However, it is difficult to find a large scale training dataset in the field of traffic control based on deep learning, which is still in its infancy.

• The Deep Learning method combined with routing next router node-by-node (DLNR) [3], which uses multiple serial deep learning systems for node-by-node prediction, can easily lead to local optimum, because only the next router is predicted at each time. Any route prediction error will inevitably lead to the failure of the entire routing path prediction.

• The existing Deep Learning (DL) method for directly predicting an overall routing path (DLDR) in a medium-size network, has high error rates [3]. This is because it is difficult for users to train a complex deep learning system with limited computing resources and training data.

To solve these problems, an intelligent routing strategy, named as Deep-Learning-Block-Routing strategy (DLBR strategy), is proposed in this paper. This method provides a solution that enables users to perform complex routing calculations in next-generation networks using existing computing resources and limited training data, and to obtain routing paths faster and more accurately. Our strategy consists of three phases: the initialization phase, the training phase and the running phase. In initialization phase, we divide the original network recursively into multiple sub-blocks according to different computing conditions and requirements of users. The input patterns of each sub-block are collected. In training phase, the main task is to train parameters of deep learning systems in each sub-block. In the running phase, first, a new input pattern is fed into the trained deep learning system, which outputs the routing path in each sub-block. Next, we get the best path in each sub-block through the deep learning system, and the corresponding nodes are added to the node set of the best path in each sub-block until the destination router is reached. Finally, the entire best path is computed from these paths according to the evaluation function. The main contributions of this work are summarized as follows:

• The proposed method can avoid the prediction from falling into local optimum and perform global prediction as much as possible under limited computing resources, ensuring the accuracy of routing path prediction. Compared to DLDR method or DLNR method [3] in the experiments, our method gets higher accuracy while achieving lower time complexity.

• The proposed method can achieve high accuracy prediction under a smaller training data.

• Three different deep learning intelligent structures are utilized to verify the effectiveness of the proposed DLBR strategy, namely block-based fully connected intelligent routing method (i.e., DLBR-fully connected method), block-based convolutional neural network intelligent routing method (i.e., DLBR-CNN method) and block-based residual network intelligent routing method (i.e., DLBR-RES method). All three proposed methods have achieved excellent results. The experimental results show that the proposed DLBR strategy has the ability to combine with different deep learning intelligent structures.

The remainder of the paper is organized as follows. In Section 2, we discuss related research work. In Section 3, we present the design of DLBR strategy in detail. Following that, the proposed method is evaluated using a set of experiments in Section 4. Finally, we give the conclusion.