WO2024037136A1 - 一种基于图结构特征的路由优化方法与系统 - Google Patents
一种基于图结构特征的路由优化方法与系统 Download PDFInfo
- Publication number
- WO2024037136A1 WO2024037136A1 PCT/CN2023/098735 CN2023098735W WO2024037136A1 WO 2024037136 A1 WO2024037136 A1 WO 2024037136A1 CN 2023098735 W CN2023098735 W CN 2023098735W WO 2024037136 A1 WO2024037136 A1 WO 2024037136A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- network
- graph
- target
- policy
- routing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L45/00—Routing or path finding of packets in data switching networks
- H04L45/02—Topology update or discovery
- H04L45/08—Learning-based routing, e.g. using neural networks or artificial intelligence
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L45/00—Routing or path finding of packets in data switching networks
- H04L45/02—Topology update or discovery
Definitions
- the invention relates to the field of computer network technology, and in particular to a routing optimization method and system based on graph structure characteristics.
- the purpose of the present invention is to provide a routing optimization method and system based on graph structure characteristics, which is suitable for SDN network environments.
- Switches or routing devices support traditional layer 2 network protocols to optimize global routing overhead from multiple network attributes. , adapt to dynamic and complex SDN networks and ensure SDN network performance.
- the present invention designs a routing optimization method based on graph structure characteristics.
- the following steps S1 to S3 are performed to obtain the routing overhead of each link in the target SDN network and adjust each link. weight to complete routing optimization of the target SDN network.
- Step S1 For the target SDN network, based on the southbound interface protocol, obtain the network topology diagram of the target SDN network, and construct a graph adjacency matrix according to the connection relationship between the nodes on each link of the target SDN network in the network topology diagram, respectively. For each node on each link of the target SDN network, construct the information feature vector of each node based on the link bandwidth, traffic, packet loss rate, and transmission delay of each node, and build the target SDN based on the information feature vector of each node Network information feature matrix of the network.
- Step S2 Taking the graph adjacency matrix and the network information feature matrix as the state of the target SDN network, based on the graph learning algorithm
- the graph adjacency matrix and the network information feature matrix are used as input, through the deep graph learning method, the routing strategy and routing cost of the target SDN network in the current state are used as the output, and based on the gradient back propagation method, the graph learning neural network is updated.
- Network parameters, and after a preset number of iterations, the graph learning neural network is trained to obtain a deep graph learning model that minimizes the routing overhead of the target SDN network and maximizes link utilization.
- Step S3 Based on the trained deep graph learning model and the status of the target SDN network, obtain the routing strategy that minimizes the routing cost of the target SDN network, deploy the routing strategy to the target SDN network, and change each link of the target SDN network according to the routing strategy. Route weight to complete routing optimization of the target SDN network.
- step S1 the specific steps of step S1 are as follows:
- Step S1.1 For the target SDN network, based on the southbound interface protocol, obtain the network topology of the target SDN network, where the network topology includes M routers and N links.
- Step S1.2 Based on the network topology of the target SDN network, each router corresponds to a real node, and each link corresponds to an edge. Insert a virtual node on the edge corresponding to each link, and combine the network topology of the target SDN network.
- V real represents the set of real nodes
- V virtual represents the set of virtual nodes
- V real ⁇ v s1 , v s2 ,..., v sM ⁇
- v s1 , v s2 ,..., v sM represent M real nodes;
- V virtual ⁇ v x1 , v x2 ,..., v xN ⁇
- e 1 , e 2 ,..., e 2N represent 2N edges.
- the elements a ij in the graph adjacency matrix A are as follows:
- B wi is the link bandwidth of node i
- T hi is the traffic of node i
- L pi is the packet loss rate of node i
- D ti is the transmission delay of node i
- the network information feature matrix H of the target SDN network is constructed as follows:
- h 1 , h 2 ,..., h i ,..., h x are the information feature vectors of each node.
- node i described in step S1.4 if node i is a virtual node, then the traffic T hi , packet loss rate L pi , and transmission delay D ti of node i are 0, If node i is a real node, the link bandwidth B wi of node i is 0.
- the deep graph learning method in step S2 includes four graph learning neural networks and an experience pool.
- the four graph learning neural networks are respectively an online graph policy network, an online graph value network, and a target Graph policy network, target graph value network, and the four graph learning neural networks each include an input layer, two hidden layers, and an output layer.
- the input layer of the online graph policy network and the target graph policy network is based on the graph adjacency matrix A and the network information feature matrix H
- the outputs of the online graph policy network and the target graph policy network are used as the inputs of the online graph value network and the target graph value network respectively.
- Each graph learns the propagation formula from the input layer of the neural network to the hidden layer and between hidden layers. Similarly, if the input layer is recorded as layer 0, the first hidden layer is recorded as layer 1, and the second hidden layer is recorded as layer 2, the propagation formula is as follows:
- ⁇ ( ⁇ ) means normalizing the formula inside the brackets
- H l is the network information feature matrix of the l-th layer
- W l+1 is the weight matrix of the l+1-th layer
- H 0 H
- I is the x-order unit matrix, for The degree matrix of As follows:
- W 1 is a 4 ⁇ 4 matrix
- W 2 is a 4 ⁇ 1 matrix
- the output layer is a fully connected layer
- its output value is an x ⁇ 1 matrix
- K is the weight matrix of the output layer of the online graph policy network and the target graph policy network
- H 2 is the network information feature matrix of the second layer
- W 1 and W 2 are both A 1 ⁇ 1 matrix
- the output layer is the aggregation layer
- its output value is a 1 ⁇ 1 matrix, recorded as Value, specifically as follows:
- Q is the weight value of the output layer, is the i-th value in the layer 2 network information feature matrix H2 ; according to the routing policy Policy output by the online graph policy network, the routing cost of each link in the target SDN network is updated.
- step S2 the specific steps of step S2 are as follows:
- Step S2.1 Initialize the weight matrices of the online graph policy network, online graph value network, target policy network, and target graph value network.
- the weight matrix of the online graph policy network is W ⁇ and the weight matrix of the online graph value network is W ⁇ ′
- the weight matrix of the target graph policy network is W ⁇
- the weight matrix of the target graph value network is W ⁇ ′ .
- Step S2.2 Initialize the experience pool. The specific steps are as follows:
- Step S2.2.2 Definition are the outputs of the output layer of the online graph policy network, target graph policy network, online graph value network, and target graph value network at time t respectively; calculate the output routing policy of the online graph policy network according to the following formula
- U(B w , Th , L p , D t ) is the link utilization rate
- B w , Th , L p , and D t are the link bandwidth, traffic, and packet loss rate of the target SDN network respectively.
- Transmission delay, K f is the proportional coefficient
- the objective function to construct the target SDN network link utilization maximization is U max (B w , Th , L p , D t ).
- Step S2.2.3 Define the experience pool R as follows:
- s t+1 represents the status of the target SDN network at time t+1, that is, the online graph policy network outputs the routing policy The obtained status of the target SDN network.
- Step S2.3 For the target SDN network, perform a preset number of iterations, where the preset number of iterations is T.
- the specific steps are as follows:
- Step S2.3.2 Based on the status s t of the target SDN network at time t, the online graph policy network outputs the routing policy.
- the process is recorded as Among them, ⁇ is the network parameter of the online graph policy network;
- Step S2.3.3 According to routing policy Update the routing costs of each link in the target SDN network
- Step S2.3.4 Obtain the routing policy Updated target SDN network state s t+1 , and obtain environmental feedback f t at the same time;
- Step S2.3.5 Place Stored in the experience pool R as a set of historical records
- Step S2.3.6 Randomly select Y groups of historical records from the experience pool R Among them, the subscript m represents any set of historical records in the experience pool R;
- Step S2.3.7 Based on the historical records extracted in step S2.3.6 Calculate target map value The output corresponding to the network As follows:
- ⁇ ′ is the network parameter of the target graph policy network
- ⁇ ′ is the network parameter of the target graph value network
- ⁇ is the discount.
- Factor is a constant, and ⁇ (0,1);
- Step S2.3.8 Calculate the loss Loss ogvn of the online graph value network output value according to the following formula:
- the online graph value network representing the network parameter ⁇ is in the state s m of the target SDN network, and when the routing policy output by the online graph policy network is ⁇ (s m
- Step S2.3.9 According to the loss Loss ogvn of the output value of the online graph value network, based on the gradient backpropagation method, update the network parameters ⁇ of the online graph value network;
- Step S2.3.10 Calculate gradient value According to the gradient value Based on the gradient backpropagation method, the network parameters ⁇ of the online graph policy network are updated, where Indicates finding the gradient of the formula in parentheses;
- ⁇ is a constant, and ⁇ (0,1);
- Step S2.3.12 Repeat S2.3.2 to step S2.3.11 until the number of iterations reaches the preset number T, and the target SDN is obtained.
- step S3 the specific steps of step S3 are as follows:
- Step S31 Obtain the graph adjacency matrix A and network information feature matrix H of the target SDN network;
- Step S32 Based on the trained deep graph learning model and according to the status [A, H] of the target SDN network, obtain the routing strategy that minimizes the routing cost of the target SDN network;
- Step S33 Deploy to the target SDN network according to the routing policy obtained in step S32, and change the link weights of the target SDN network according to the routing policy;
- Step S34 During the traffic transmission process, the updated weight of each link is used for traffic transmission according to the shortest path scheme.
- the present invention also designs a system for route optimization method based on graph structure characteristics.
- the target SDN network includes a control plane and a data plane, where the control plane includes an information acquisition module, a policy deployment module, and a DGL module; so that the method based on graph structure characteristics
- the system of routing optimization method implements the routing optimization method based on graph structure characteristics.
- Each link and node of the target SDN network is deployed on the data plane.
- the information acquisition module on the control plane is used to obtain the network topology diagram of the target SDN network, generate a graph adjacency matrix and a network information feature matrix, and send them to the DGL module.
- the DGL module is based on the graph learning neural network. It takes the graph adjacency matrix and the network information feature matrix as inputs. Through the deep graph learning method, it uses the routing cost of the target SDN network in the current state as the output. Based on the gradient back propagation method, it updates the graph learning neural network. Network parameters of the network, and after a preset number of iterations, the graph learning neural network is trained to obtain a deep graph learning model that minimizes the routing overhead of the target SDN network and maximizes link utilization.
- the policy deployment module on the control plane is used to obtain the routing strategy that minimizes the routing cost of the target SDN network based on the trained deep graph learning model obtained by the DGL module and based on the status of the target SDN network, and combine the routing strategy with the target SDN network Routing overhead is sent to the data plane.
- the advantages of the present invention include:
- the deep graph learning model has strong generalization ability.
- the trained deep graph learning model is still effective when the network topology changes, and can adapt to large-scale dynamic and complex networks.
- Figure 1 is an overall block diagram of a system based on a route optimization method based on graph structure features provided according to an embodiment of the present invention
- Figure 2 is a DGL algorithm framework diagram provided according to an embodiment of the present invention.
- Figure 3 is a structural diagram of a graph learning neural network provided according to an embodiment of the present invention.
- the embodiment of the present invention provides a routing optimization method based on graph structure characteristics. For the target SDN network, the following steps S1 to S3 are performed to obtain the routing overhead of each link in the target SDN network and adjust the weight of each link. , complete the routing optimization of the target SDN network.
- Step S1 Referring to Figure 1, for the target SDN network, obtain the network topology diagram of the target SDN network based on the southbound interface protocol, and construct a diagram based on the connection relationships between the nodes on each link of the target SDN network in the network topology diagram.
- the adjacency matrix is for each node on each link of the target SDN network, and based on the link bandwidth, traffic, packet loss rate, and transmission delay of each node, the information feature vector of each node is constructed, and based on the information feature vector of each node , construct the network information feature matrix of the target SDN network.
- step S1 The specific steps of step S1 are as follows:
- Step S1.1 For the target SDN network, based on the southbound interface protocol, obtain the network topology of the target SDN network, where the network topology includes M routers and N links.
- Step S1.2 Based on the network topology of the target SDN network, each router corresponds to a real node, and each link corresponds to an edge. Insert a virtual node on the edge corresponding to each link, and combine the network topology of the target SDN network.
- V real represents the set of real nodes
- V virtual represents the set of virtual nodes
- V real ⁇ v s1 , v s2 ,..., v sM ⁇
- v s1 , v s2 ,..., v sM represent M real nodes;
- V virtual ⁇ v x1 , v x2 ,..., v xN ⁇
- e 1 , e 2 ,..., e 2N represent 2N edges.
- the elements a ij in the graph adjacency matrix A are as follows:
- B wi is the link bandwidth of node i
- T hi is the traffic of node i
- L pi is the packet loss rate of node i
- D ti is the transmission delay of node i.
- the node i For the node i, if the node i is a virtual node, the traffic T hi , the packet loss rate L pi , and the transmission delay D ti of the node i are 0. If the node i is a real node, the link bandwidth B wi of the node i is 0.
- the network information feature matrix H of the target SDN network is constructed as follows:
- h 1 , h 2 ,..., h i ,..., h x are the information feature vectors of each node.
- Step S2 Take the graph adjacency matrix and the network information feature matrix as the target SDN network state, learn the neural network based on the graph, take the graph adjacency matrix and the network information feature matrix as input, and use the deep graph learning method (Deep Graph Learning, DGL) to Taking the routing strategy and routing cost of the target SDN network in the current state as the output, based on the gradient back propagation method, the network parameters of the graph learning neural network are updated, and after a preset number of iterations, the graph learning neural network is Conduct training to obtain a deep graph learning model that minimizes routing overhead and maximizes link utilization in the target SDN network.
- DGL Deep Graph Learning
- the deep graph learning method described in step S2 includes four graph learning neural networks and an experience pool.
- the four graph learning neural networks are the Online Graph Strategy Network (OGSN) and the Online Graph Value Network. (Online Graph Value Network, OGVN), Target Graph Strategy Network (TGSN), Target Graph Value Network (Target Graph Value Network, TGVN), referring to Figure 3, each of the four graph learning neural networks includes an input layer, two hidden layers, and an output layer.
- the input layer of the online graph policy network and the target graph policy network takes the graph adjacency matrix A and the network information feature matrix H as inputs, and the outputs of the online graph policy network and the target graph policy network serve as the inputs of the online graph value network and the target graph value network respectively.
- the propagation formulas from the input layer to the hidden layer and between hidden layers of each graph learning neural network are the same.
- the input layer is recorded as layer 0
- the first hidden layer is recorded as layer 1
- the second hidden layer Denoted as layer 2
- the propagation formula is as follows:
- ⁇ ( ⁇ ) means normalizing the formula inside the brackets
- H l is the network information feature matrix of the l-th layer
- W l+1 is the weight matrix of the l+1-th layer
- H 0 H
- I is the x-order unit matrix, for The degree matrix of As follows:
- W 1 is a 4 ⁇ 4 matrix
- W 2 is a 4 ⁇ 1 matrix
- the output layer is a fully connected layer
- its output value is an x ⁇ 1 matrix
- K is the weight matrix of the output layer of the online graph policy network and the target graph policy network
- H 2 is the network information feature matrix of the second layer.
- W 1 and W 2 are both 1 ⁇ 1 matrices
- the output layer is the aggregation layer
- its output value is a 1 ⁇ 1 matrix, recorded as Value, as follows:
- Q is the weight value of the output layer, is the i-th value in the layer 2 network information feature matrix H2 ; according to the routing policy Policy output by the online graph policy network, the routing cost of each link in the target SDN network is updated.
- step S2 the specific steps of step S2 are as follows:
- Step S2.1 Initialize the weight matrices of the online graph policy network, online graph value network, target policy network, and target graph value network.
- the weight matrix of the online graph policy network is W ⁇ and the weight matrix of the online graph value network is W ⁇ ′
- the weight matrix of the goal graph policy network is W ⁇
- the weight matrix of the goal graph value network is W ⁇ ′ .
- the network parameters of the online graph policy network and the goal policy network are consistent, and the online graph value network and goal graph The network parameters of the value network are consistent.
- Step S2.2 Initialize the experience pool. The specific steps are as follows:
- Step S2.2.2 Definition are the outputs of the output layer of the online graph policy network, target graph policy network, online graph value network, and target graph value network at time t respectively; calculate the output routing policy of the online graph policy network according to the following formula
- U(B w , Th , L p , D t ) is the link utilization rate
- B w , Th , L p , and D t are the link bandwidth, traffic, and packet loss rate of the target SDN network respectively.
- Transmission delay, K f is the proportional coefficient.
- the objective function to construct the target SDN network link utilization maximization is U max (B w , Th , L p , D t ).
- Step S2.2.3 Define the experience pool R as follows:
- s t+1 represents the status of the target SDN network at time t+1, that is, the online graph policy network outputs the routing policy The obtained status of the target SDN network.
- Step S2.3 For the target SDN network, perform a preset number of iterations, where the preset number of iterations is T.
- the specific steps are as follows:
- Step S2.3.2 Based on the status s t of the target SDN network at time t, the online graph policy network outputs the routing policy.
- the process is recorded as Among them, ⁇ is the network parameter of the online graph policy network;
- Step S2.3.3 According to routing policy Update the routing costs of each link in the target SDN network
- Step S2.3.4 Obtain the routing policy Updated target SDN network state s t+1 , and obtain environmental feedback f t at the same time;
- Step S2.3.5 Place Stored in the experience pool R as a set of historical records
- Step S2.3.6 Randomly select Y groups of historical records from the experience pool R Among them, the subscript m represents any set of historical records in the experience pool R;
- Step S2.3.7 Based on the historical records extracted in step S2.3.6 Calculate the output corresponding to the target graph value network As follows:
- ⁇ ′ is the network parameter of the target graph policy network
- ⁇ ′ is the network parameter of the target graph value network
- ⁇ is the discount.
- the factor is a constant, and ⁇ (0,1).
- Step S2.3.8 Calculate the loss Loss ogvn of the online graph value network output value according to the following formula:
- the online graph value network representing the network parameter ⁇ is in the state s m of the target SDN network, and when the routing policy output by the online graph policy network is ⁇ (s m
- Step S2.3.9 According to the loss Loss ogvn of the output value of the online graph value network, based on the gradient backpropagation method, update the network parameters ⁇ of the online graph value network.
- Step S2.3.10 Calculate gradient value According to the gradient value Based on the gradient backpropagation method, the network parameters ⁇ of the online graph policy network are updated, where Indicates finding the gradient of the formula in parentheses.
- ⁇ is a constant, and ⁇ (0,1).
- Step S2.3.12 Repeat S2.3.2 to Step S2.3.11 until the number of iterations reaches the preset number T, and obtain the routing strategy that minimizes the routing cost of the target SDN network.
- Step S3 Based on the trained deep graph learning model and the status of the target SDN network, obtain the routing strategy that minimizes the routing cost of the target SDN network, deploy the routing strategy to the target SDN network, and change each link of the target SDN network according to the routing strategy. Route weight to complete routing optimization of the target SDN network.
- step S3 The specific steps of step S3 are as follows:
- Step S31 Obtain the graph adjacency matrix A and network information feature matrix H of the target SDN network;
- Step S32 Based on the trained deep graph learning model and according to the status [A, H] of the target SDN network, obtain the routing strategy that minimizes the routing cost of the target SDN network;
- Step S33 Deploy to the target SDN network according to the routing policy obtained in step S32, and change the link weights of the target SDN network according to the routing policy;
- Step S34 During the traffic transmission process, the updated weight of each link is used for traffic transmission according to the shortest path scheme.
- Embodiments of the present invention also provide a system for routing optimization methods based on graph structure characteristics.
- the target SDN network includes a control plane and a data plane, where the control plane includes an information acquisition module, a policy deployment module, and a DGL module; such that The system of the routing optimization method based on graph structure characteristics implements the routing optimization method based on graph structure characteristics.
- Each link and node of the target SDN network is deployed on the data plane.
- the information acquisition module on the control plane is used to obtain the network topology diagram of the target SDN network, generate a graph adjacency matrix and a network information feature matrix, and send them to the DGL module.
- the DGL module is based on the graph learning neural network. It takes the graph adjacency matrix and the network information feature matrix as inputs. Through the deep graph learning method, it uses the routing cost of the target SDN network in the current state as the output. Based on the gradient back propagation method, it updates the graph learning neural network. Network parameters of the network, and after a preset number of iterations, the graph learning neural network is trained to obtain a deep graph learning model that minimizes the routing overhead of the target SDN network and maximizes link utilization.
- the policy deployment module on the control plane is used to obtain the routing strategy that minimizes the routing cost of the target SDN network based on the trained deep graph learning model obtained by the DGL module and based on the status of the target SDN network, and combine the routing strategy with the target SDN network Routing overhead is sent to the data plane.
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
Description
V={V实,V虚}
V实={vs1,vs2,...,vsM}
V虚={vx1,vx2,...,vxN}
E={e1,e2,...,e2N}
hi=[Bwi,Thi,Lpi,Dti]
Policy=H2×K
ft=U(Bw,Th,Lp,Dt)×Kf
θ′=τθ+(1-τ)θ′
ω′=τω+(1-τ)ω′
V={V实,V虚}
V实={vs1,vs2,...,vsM}
V虚={vx1,vx2,...,vxN}
E={e1,e2,...,e2N}
hi=[Bwi,Thi,Lpi,Dti]
Policy=H2×K
ft=U(Bw,Th,Lp,Dt)×Kf
θ′=τθ+(1-τ)θ′
ω′=τω+(1-τ)ω′
Claims (7)
- 一种基于图结构特征的路由优化方法,其特征在于,针对目标SDN网络,执行以下步骤S1-步骤S3,获得目标SDN网络中各条链路的路由开销,调整各条链路的权重,完成目标SDN网络的路由优化:步骤S1:针对目标SDN网络,基于南向接口协议,获取目标SDN网络的网络拓扑图,根据网络拓扑图中目标SDN网络的各链路上各节点之间的连接关系,构建图邻接矩阵,分别针对目标SDN网络的各链路上各节点,根据各节点的链路带宽、流量、丢包率、传输时延,构建各节点的信息特征向量,并基于各节点的信息特征向量,构建目标SDN网络的网络信息特征矩阵;步骤S2:以图邻接矩阵、网络信息特征矩阵为目标SDN网络的状态,基于图学习神经网络,以图邻接矩阵、网络信息特征矩阵为输入,通过深度图学习方法,以当前状态下目标SDN网络的路由策略、路由开销为输出,基于梯度反向传播方法,更新图学习神经网络的网络参数,并经过预设次数的迭代,对图学习神经网络进行训练,获得使目标SDN网络路由开销最小、链路利用率最大的深度图学习模型;步骤S3:根据训练好的深度图学习模型,基于目标SDN网络的状态,获得使目标SDN网络路由开销最小的路由策略,将路由策略部署至目标SDN网络,根据路由策略改变目标SDN网络的各链路权重,完成目标SDN网络的路由优化。
- 根据权利要求1所述的一种基于图结构特征的路由优化方法,其特征在于,步骤S1的具体步骤如下:步骤S1.1:针对目标SDN网络,基于南向接口协议,获取目标SDN网络的网络拓扑结构,其中网络拓扑结构包含M个路由器、N条链路;步骤S1.2:针对目标SDN网络的网络拓扑结构,每个路由器对应一个实节点,每条链路对应一条边,在每条链路所对应的边上插入虚节点,将目标SDN网络的网络拓扑结构表示为M个实节点、N个虚节点、2N条边的网络拓扑图G(V,E),其中,V表示节点集合,E表示边集合,具体如下式:
V={V实,V虚}其中,V实表示实节点集合,V虚表示虚节点集合;V实={vs1,vs2,...,vsM}其中,vs1,vs2,...,vsM表示M个实节点;
V虚={vx1,vx2,...,vxN}其中,vx1,vx2,...,vxN表示N个虚节点;
E={e1,e2,...,e2N}其中,e1,e2,...,e2N表示2N条边;步骤S1.3:令x=M+N,x表示节点总数,节点包括M个实节点、N个虚节点,基于目标SDN网络的网络拓扑图,构建x阶的图邻接矩阵A如下式:
其中,图邻接矩阵A中的元素aij如下式:
步骤S1.4:针对目标SDN网络的任一节点i,根据节点i的链路带宽、流量、丢包率、传输时延,构建节点i的信息特征向量hi如下式:
hi=[Bwi,Thi,Lpi,Dti]式中,Bwi为节点i的链路带宽,Thi为节点i的流量,Lpi为节点i的丢包率,Dti为节点i的传输时延;基于各节点的信息特征向量,构建目标SDN网络的网络信息特征矩阵H如下式:
式中,h1,h2,...,hi,...,hx为各节点的信息特征向量。 - 根据权利要求2所述的一种基于图结构特征的路由优化方法,其特征在于,步骤S1.4中 所述的节点i,若节点i为虚节点,则节点i的流量Thi、丢包率Lpi、传输时延Dti为0,若节点i为实节点,则节点i的链路带宽Bwi为0。
- 根据权利要求2所述的一种基于图结构特征的路由优化方法,其特征在于,步骤S2中所述深度图学习方法包括四个图学习神经网络和一个经验池,四个图学习神经网络分别为在线图策略网络、在线图价值网络、目标图策略网络、目标图价值网络,四个图学习神经网络分别均包括一个输入层、两个隐藏层、一个输出层;在线图策略网络、目标图策略网络的输入层以图邻接矩阵A、网络信息特征矩阵H为输入,在线图策略网络、目标图策略网络的输出分别作为在线图价值网络、目标图价值网络的输入,其中,各图学习神经网络的输入层到隐藏层、以及隐藏层之间的传播公式相同,将输入层记为第0层,第一个隐藏层记为第1层,第二个隐藏层记为第2层,则传播公式如下式:
式中,σ(·)表示将括号内部的公式进行归一化,Hl为第l层的网络信息特征矩阵,Wl+1为第l+1层的权重矩阵,其中,H0=H,I为x阶单位矩阵,为的度矩阵,如下式:
其中,如下式:
其中,在线图策略网络、目标图策略网络中,W1是一个4×4的矩阵,W2是一个4×1的矩阵,输出层为全连接层,其输出值为x×1矩阵,记为路由策略Policy,具体如下式:
Policy=H2×K式中,K为在线图策略网络、目标图策略网络输出层的权重矩阵,H2为第2层的网络信息特征矩阵;在线图价值网络、目标图价值网络中,W1和W2均为1×1的矩阵,输出层为聚合层,其输出值为1×1矩阵,记为Value,具体如下式:
式中,Q为输出层的权重值,为第2层的网络信息特征矩阵H2中的第i个值;根据在线图策略网络输出的路由策略Policy,更新目标SDN网络中各条链路的路由开销。 - 根据权利要求4所述的一种基于图结构特征的路由优化方法,其特征在于,步骤S2的具体步骤如下:步骤S2.1:对在线图策略网络、在线图价值网络、目标策略网络、目标图价值网络的权重矩阵初始化,其中,在线图策略网络的权重矩阵为Wθ,在线图价值网络的权重矩阵为Wθ′,目标图策略网络的权重矩阵为Wω,目标图价值网络的权重矩阵为Wω′;步骤S2.2:对经验池进行初始化,具体步骤如下:步骤S2.2.1:以图邻接矩阵A、网络信息特征矩阵H作为目标SDN网络的状态S,定义S=[A,H],st表示t时刻目标SDN网络的状态,st=[At,Ht],At表示t时刻目标SDN网络的图邻接矩阵,Ht表示t时刻目标SDN网络的网络信息特征矩阵;步骤S2.2.2:定义分别为在线图策略网络、目标图策略网络、在线图价值网络、目标图价值网络的输出层在t时刻的输出;根据下式计算在线图策略网络输出路由策略所获得的环境反馈ft:
ft=U(Bw,Th,Lp,Dt)×Kf式中,U(Bw,Th,Lp,Dt)为链路利用率,Bw、Th、Lp、Dt分别为目标SDN网络的链路带宽、流量、丢包率、传输时延,Kf为比例系数;构建目标SDN网络链路利用率最大化的目标函数为Umax(Bw,Th,Lp,Dt);步骤S2.2.3:定义经验池R如下式:
式中,st+1表示t+1时刻目标SDN网络的状态,即在线图策略网络输出路由策略所获得目标SDN网络的状态;步骤S2.3:针对目标SDN网络,进行预设次数的迭代,其中预设迭代次数为T,具体步骤如 下:步骤S2.3.1:令t=1,获取目标SDN网络的初始状态s1;步骤S2.3.2:在线图策略网络根据t时刻目标SDN网络的状态st,输出路由策略过程记为其中,θ为在线图策略网络的网络参数;步骤S2.3.3:根据路由策略更新目标SDN网络中各条链路的路由开销;步骤S2.3.4:获取根据路由策略更新后的目标SDN网络的状态st+1,同时获取环境反馈ft;步骤S2.3.5:将作为一组历史记录存入经验池R中;步骤S2.3.6:从经验池R中随机抽取Y组历史记录其中,下标m表示经验池R中任意一组历史记录;步骤S2.3.7:根据步骤S2.3.6所抽取的历史记录计算目标图价值网络所对应的输出如下式:
式中,表示目标图策略网络根据目标SDN网络的状态sm+1所选择的路由策略,θ′所选择的路由策略,ω′为目标图价值网络的网络参数,表示目标图价值网络基于目标SDN网络的状态sm+1且网络参数为ω′时,目标图策略网络所选取的路由策略π′(sm+1|θ′)的期望值,γ为折扣因子,是一个常数,且γ∈(0,1);步骤S2.3.8:根据下式计算在线图价值网络输出值的损失Lossogvn:
式中,表示网络参数ω的在线图价值网络在目标SDN网络的状态sm下,在线图策略网络输出的路由策略为π(sm|θ)时,在线图价值 网络输出的价值;步骤S2.3.9:根据在线图价值网络输出值的损失Lossogvn,基于梯度反向传播方法,更新在线图价值网络的网络参数ω;步骤S2.3.10:计算梯度值根据梯度值基于梯度反向传播方法,更新在线图策略网络的网络参数θ,其中表示对括号内公式求梯度;步骤S2.3.11:分别根据下式,更新目标图策略网络的网络参数θ′、目标图价值网络的网络参数ω′:
θ′=τθ+(1-τ)θ′
ω′=τω+(1-τ)ω′式中,τ为常数,且τ∈(0,1);步骤S2.3.12:重复S2.3.2至步骤S2.3.11,直至迭代次数达到预设次数T,获得使目标SDN网络路由开销最小的路由策略。 - 根据权利要求5所述的一种基于图结构特征的路由优化方法,其特征在于,步骤S3的具体步骤如下:步骤S31:获取目标SDN网络的图邻接矩阵A、网络信息特征矩阵H;步骤S32:基于训练好的深度图学习模型,根据目标SDN网络的状态[A,H],获得使目标SDN网络路由开销最小的路由策略;步骤S33:根据步骤S32所获得的路由策略,部署至目标SDN网络,根据路由策略改变目标SDN网络的各链路权重;步骤S34:在流量传输过程中,根据最短路径方案,采用更新后的各链路权重进行流量传输。
- 一种基于图结构特征的路由优化方法的系统,其特征在于,目标SDN网络包括控制平面、数据平面,其中,控制平面包括信息获取模块、策略部署模块、DGL模块;使得所述基于图结构特征的路由优化方法的系统实现如权利要求1-6中任一项所述的基于图结构特征的路由优化方法;目标SDN网络的各链路及各节点部署于数据平面,控制平面上的信息获取模块用于获取目标SDN网络的网络拓扑图,生成图邻接矩阵、网络信息特征矩阵,发送至DGL模块;DGL模块基于图学习神经网络,以图邻接矩阵、网络信息特征矩阵为输入,通过深度图学 习方法,以当前状态下目标SDN网络的路由开销为输出,基于梯度反向传播方法,更新图学习神经网络的网络参数,并经过预设次数的迭代,对图学习神经网络进行训练,获得使目标SDN网络路由开销最小、链路利用率最大的深度图学习模型;控制平面上的策略部署模块用于根据DGL模块所获得的训练好的深度图学习模型,基于目标SDN网络的状态,获得使目标SDN网络路由开销最小的路由策略,并将路由策略及目标SDN网络路由开销发送到数据平面。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2025508521A JP7755361B2 (ja) | 2022-08-15 | 2023-06-07 | グラフ構造特徴に基づくルーティング最適化方法、およびシステム |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210974378.6 | 2022-08-15 | ||
| CN202210974378.6A CN115225561B (zh) | 2022-08-15 | 2022-08-15 | 一种基于图结构特征的路由优化方法与系统 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2024037136A1 true WO2024037136A1 (zh) | 2024-02-22 |
Family
ID=83615692
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2023/098735 Ceased WO2024037136A1 (zh) | 2022-08-15 | 2023-06-07 | 一种基于图结构特征的路由优化方法与系统 |
Country Status (3)
| Country | Link |
|---|---|
| JP (1) | JP7755361B2 (zh) |
| CN (1) | CN115225561B (zh) |
| WO (1) | WO2024037136A1 (zh) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118228514A (zh) * | 2024-05-23 | 2024-06-21 | 中国海洋大学 | 一种浮式海洋平台推进器推力优化分配方法 |
| CN118337269A (zh) * | 2024-05-15 | 2024-07-12 | 中国人民解放军陆军工程大学 | 一种卫星星座路由方法、装置、可读存储介质和设备 |
| CN118400314A (zh) * | 2024-04-29 | 2024-07-26 | 广州市杰青计算机有限公司 | 信息化机房监控及管理系统 |
| CN118590292A (zh) * | 2024-06-21 | 2024-09-03 | 威海众达信息科技有限公司 | 基于人工智能的企业数据安全管理方法和系统 |
| CN119011463A (zh) * | 2024-10-23 | 2024-11-22 | 南京信息工程大学 | 一种基于链路状态感知增强的路径智能优选方法及系统 |
| CN119449738A (zh) * | 2025-01-08 | 2025-02-14 | 长沙超创电子科技有限公司 | 一种基于优级排序的图像传输方法及系统 |
| CN120296542A (zh) * | 2025-06-06 | 2025-07-11 | 国网山西省电力公司晋城供电公司 | 一种基于深度强化学习的配电网单相接地故障线路选拉决策方法、系统及相关设备 |
| CN120321167A (zh) * | 2025-06-16 | 2025-07-15 | 南京信息工程大学 | 基于图注意力机制和动态优先级回放的sdn路由方法及架构 |
| CN121089769A (zh) * | 2025-11-11 | 2025-12-09 | 成都信息工程大学 | 一种针对交通不确定性拓扑网络的导航方法 |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115225561B (zh) * | 2022-08-15 | 2022-12-06 | 南京邮电大学 | 一种基于图结构特征的路由优化方法与系统 |
| CN116055378B (zh) * | 2023-01-10 | 2024-05-28 | 中国联合网络通信集团有限公司 | 流量调度策略生成模型的训练方法及装置 |
| CN116346641B (zh) * | 2023-03-17 | 2026-04-03 | 东南大学 | 一种基于gnn故障模型学习的控制平面策略验证方法 |
| CN116366529B (zh) * | 2023-04-20 | 2024-07-19 | 哈尔滨工业大学 | 一种sdn背景下基于深度强化学习的自适应路由方法 |
| CN116527565B (zh) * | 2023-05-11 | 2026-03-31 | 清华大学 | 基于图卷积神经网络的互联网路由优化方法及装置 |
| CN116847425A (zh) * | 2023-06-20 | 2023-10-03 | 云南省通信产业服务有限公司 | 一种基于高维数据联合优化的多资源路由优化方法 |
| CN120434167A (zh) * | 2025-06-09 | 2025-08-05 | 天翼云科技有限公司 | 路由优化方法、装置、计算机设备、可读存储介质和程序产品 |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111245718A (zh) * | 2019-12-30 | 2020-06-05 | 浙江工商大学 | 一种基于sdn情景感知的路由优化方法 |
| CN111314171A (zh) * | 2020-01-17 | 2020-06-19 | 深圳供电局有限公司 | 一种sdn路由性能预测和优化的方法、设备及介质 |
| CN113194034A (zh) * | 2021-04-22 | 2021-07-30 | 华中科技大学 | 基于图神经网络和深度强化学习的路由优化方法及系统 |
| WO2022116957A1 (zh) * | 2020-12-02 | 2022-06-09 | 中兴通讯股份有限公司 | 算法模型及路径的确定方法、电子设备、sdn控制器和介质 |
| CN114697229A (zh) * | 2022-03-11 | 2022-07-01 | 华中科技大学 | 一种分布式路由规划模型的构建方法及应用 |
| CN115225561A (zh) * | 2022-08-15 | 2022-10-21 | 南京邮电大学 | 一种基于图结构特征的路由优化方法与系统 |
Family Cites Families (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103281247B (zh) * | 2013-05-09 | 2016-06-15 | 北京交通大学 | 一种数据中心网络的通用路由方法及系统 |
| US20190184561A1 (en) * | 2017-12-15 | 2019-06-20 | The Regents Of The University Of California | Machine Learning based Fixed-Time Optimal Path Generation |
| CN110275437B (zh) * | 2019-06-06 | 2022-11-15 | 江苏大学 | Sdn网络流量优势监控节点动态选择系统及其方法 |
| CN110611619B (zh) * | 2019-09-12 | 2020-10-09 | 西安电子科技大学 | 一种基于ddpg强化学习算法的智能化路由决策方法 |
| CN113556281A (zh) * | 2020-04-23 | 2021-10-26 | 中兴通讯股份有限公司 | 一种重新路由方法及装置、电子设备、计算机可读介质 |
| CN111862579B (zh) * | 2020-06-10 | 2021-07-13 | 深圳大学 | 一种基于深度强化学习的出租车调度方法及系统 |
| CN113036772B (zh) * | 2021-05-11 | 2022-07-19 | 国网江苏省电力有限公司南京供电分公司 | 一种基于深度强化学习的配电网拓扑电压调节方法 |
| CN113285831B (zh) * | 2021-05-24 | 2022-08-02 | 广州大学 | 网络行为知识智能学习方法、装置、计算机设备及存储介质 |
| CN114286413B (zh) * | 2021-11-02 | 2023-09-19 | 北京邮电大学 | Tsn网络联合路由选择与流分配方法及相关设备 |
| CN114500360B (zh) * | 2022-01-27 | 2022-11-11 | 河海大学 | 一种基于深度强化学习的网络流量调度方法以及系统 |
| CN114629543B (zh) * | 2022-01-28 | 2024-03-29 | 航天东方红卫星有限公司 | 一种基于深度监督学习的卫星网络自适应流量调度方法 |
| CN114741191B (zh) * | 2022-03-30 | 2024-09-06 | 西安电子科技大学 | 一种面向计算密集型任务关联性的多资源分配方法 |
| CN114859719B (zh) * | 2022-05-05 | 2025-01-21 | 电子科技大学长三角研究院(衢州) | 一种基于图神经网络的强化学习集群蜂拥控制方法 |
| CN114817578B (zh) * | 2022-06-29 | 2022-09-09 | 北京邮电大学 | 科技论文引用关系表示学习方法、系统及存储介质 |
-
2022
- 2022-08-15 CN CN202210974378.6A patent/CN115225561B/zh active Active
-
2023
- 2023-06-07 WO PCT/CN2023/098735 patent/WO2024037136A1/zh not_active Ceased
- 2023-06-07 JP JP2025508521A patent/JP7755361B2/ja active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111245718A (zh) * | 2019-12-30 | 2020-06-05 | 浙江工商大学 | 一种基于sdn情景感知的路由优化方法 |
| CN111314171A (zh) * | 2020-01-17 | 2020-06-19 | 深圳供电局有限公司 | 一种sdn路由性能预测和优化的方法、设备及介质 |
| WO2022116957A1 (zh) * | 2020-12-02 | 2022-06-09 | 中兴通讯股份有限公司 | 算法模型及路径的确定方法、电子设备、sdn控制器和介质 |
| CN113194034A (zh) * | 2021-04-22 | 2021-07-30 | 华中科技大学 | 基于图神经网络和深度强化学习的路由优化方法及系统 |
| CN114697229A (zh) * | 2022-03-11 | 2022-07-01 | 华中科技大学 | 一种分布式路由规划模型的构建方法及应用 |
| CN115225561A (zh) * | 2022-08-15 | 2022-10-21 | 南京邮电大学 | 一种基于图结构特征的路由优化方法与系统 |
Non-Patent Citations (1)
| Title |
|---|
| CHE XIANG-BEI, KANG WEN-QIAN, DENG BING, YANG KE-HAN, LI JIAN: "A Prediction Model of SDN Routing Performance Based on Graph Neural Network", ACTA ELECTRONICA SINICA, vol. 49, no. 3, 1 March 2021 (2021-03-01), pages 484 - 491, XP093140194, DOI: 10.12263/DZXB.20200120 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118400314A (zh) * | 2024-04-29 | 2024-07-26 | 广州市杰青计算机有限公司 | 信息化机房监控及管理系统 |
| CN118337269A (zh) * | 2024-05-15 | 2024-07-12 | 中国人民解放军陆军工程大学 | 一种卫星星座路由方法、装置、可读存储介质和设备 |
| CN118228514A (zh) * | 2024-05-23 | 2024-06-21 | 中国海洋大学 | 一种浮式海洋平台推进器推力优化分配方法 |
| CN118590292A (zh) * | 2024-06-21 | 2024-09-03 | 威海众达信息科技有限公司 | 基于人工智能的企业数据安全管理方法和系统 |
| CN118590292B (zh) * | 2024-06-21 | 2025-04-18 | 威海众达信息科技有限公司 | 基于人工智能的企业数据安全管理方法和系统 |
| CN119011463A (zh) * | 2024-10-23 | 2024-11-22 | 南京信息工程大学 | 一种基于链路状态感知增强的路径智能优选方法及系统 |
| CN119449738A (zh) * | 2025-01-08 | 2025-02-14 | 长沙超创电子科技有限公司 | 一种基于优级排序的图像传输方法及系统 |
| CN120296542A (zh) * | 2025-06-06 | 2025-07-11 | 国网山西省电力公司晋城供电公司 | 一种基于深度强化学习的配电网单相接地故障线路选拉决策方法、系统及相关设备 |
| CN120321167A (zh) * | 2025-06-16 | 2025-07-15 | 南京信息工程大学 | 基于图注意力机制和动态优先级回放的sdn路由方法及架构 |
| CN121089769A (zh) * | 2025-11-11 | 2025-12-09 | 成都信息工程大学 | 一种针对交通不确定性拓扑网络的导航方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115225561A (zh) | 2022-10-21 |
| CN115225561B (zh) | 2022-12-06 |
| JP7755361B2 (ja) | 2025-10-16 |
| JP2025526143A (ja) | 2025-08-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115225561B (zh) | 一种基于图结构特征的路由优化方法与系统 | |
| Mai et al. | Packet routing with graph attention multi-agent reinforcement learning | |
| CN112437020B (zh) | 一种基于深度强化学习的数据中心网络负载均衡方法 | |
| CN113010305B (zh) | 部署在边缘计算网络中的联邦学习系统及其学习方法 | |
| CN116527567B (zh) | 一种基于深度强化学习的智能网络路径优选方法与系统 | |
| CN114697229B (zh) | 一种分布式路由规划模型的构建方法及应用 | |
| CN111988225B (zh) | 基于强化学习和迁移学习的多路径路由方法 | |
| CN110012516B (zh) | 一种基于深度强化学习架构的低轨卫星路由策略方法 | |
| CN115396366B (zh) | 基于图注意力网络的分布式智能路由方法 | |
| CN107547457A (zh) | 一种基于改进粒子群优化bp神经网络的盲信道均衡方法 | |
| CN111898728A (zh) | 一种基于多Agent强化学习的团队机器人决策方法 | |
| Lei et al. | Congestion control in SDN-based networks via multi-task deep reinforcement learning | |
| CN114143264B (zh) | 一种SRv6网络下基于强化学习的流量调度方法 | |
| CN119011463B (zh) | 一种基于链路状态感知增强的路径智能优选方法及系统 | |
| CN109039942A (zh) | 一种基于深度强化学习的网络负载均衡系统及均衡方法 | |
| CN113612692B (zh) | 基于dqn算法的集中式光片上网络自适应路由规划方法 | |
| CN110825116B (zh) | 基于时变网络拓扑的无人机编队方法 | |
| CN113395207A (zh) | 一种sdn架构下基于深度强化学习的路由优化架构及方法 | |
| CN116055324B (zh) | 一种用于数据中心网络自优化的数字孪生方法 | |
| CN111917642B (zh) | 分布式深度强化学习的sdn网络智慧路由数据传输方法 | |
| CN109947131A (zh) | 一种基于强化学习的多水下机器人编队控制方法 | |
| CN118282918A (zh) | 基于知识定义网络和图强化学习的智能路由系统及方法 | |
| CN120301812A (zh) | 一种基于gnn和多智能体drl的自适应路径选择系统 | |
| CN111885493A (zh) | 一种基于改进布谷鸟搜索算法的微云部署方法 | |
| CN115150335A (zh) | 一种基于深度强化学习的最优流量分割的方法和系统 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23854042 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2025508521 Country of ref document: JP |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 23854042 Country of ref document: EP Kind code of ref document: A1 |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 23854042 Country of ref document: EP Kind code of ref document: A1 |