Adaptive network topology dynamic reconstruction method and system based on deep reinforcement learning

Abstract

The invention provides a self-adaptive network topology dynamic reconstruction method and system based on deep reinforcement learning, and relates to the technical field of deep reinforcement learning, and the method comprises the steps: obtaining the topology state information, service flow distribution information and historical reconstruction records of a current network; extracting topological correlation characteristics among nodes through graph convolution operation, and generating fusion state representation in combination with service flow information; inputting the fusion state representation into a deep reinforcement learning model to identify bottleneck nodes and redundant links, and outputting a reconstruction action candidate set; searching and evaluating the long-term cumulative income of the candidate actions through a Monte Carlo tree, and screening an optimal reconstruction action sequence; a graph coloring algorithm is utilized to allocate time slots and process resource conflicts, and a resource-feasible topology adjustment scheme is generated; and extracting a network evolution rule through tensor decomposition, and constructing a topological optimization association mapping graph. According to the method, the network bottleneck can be intelligently identified, the network topology structure is dynamically optimized, and the network performance and the resource utilization rate are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep reinforcement learning, in particular to a deep reinforcement learning based adaptive network topology dynamic reconstruction method and system. BACKGROUND

[0002] With the rapid development of emerging technologies such as 5G, Internet of Things and cloud computing, network traffic presents characteristics such as strong dynamics, high burstiness and uneven distribution, and the traditional static network architecture has been difficult to meet the increasingly complex business requirements. Network topology reconstruction, as an effective network optimization means, can dynamically adjust the network structure according to real-time traffic changes and business requirements, improve network resource utilization, reduce network latency, and enhance network performance. In recent years, the rise of software-defined network technology provides a technical basis for flexible reconstruction of network topology, separates the network control plane from the data plane, and realizes the programmability of network configuration. At the same time, the development of artificial intelligence technology, especially deep learning and reinforcement learning, provides new ideas and methods for solving complex decision-making problems such as network topology optimization.

[0003] Existing methods mainly rely on traditional graph analysis or simple statistical features when extracting network topology features, lack the ability to mine deep relationships between network nodes, and are difficult to accurately identify bottleneck nodes and redundant links in complex network environments, resulting in insufficient accuracy of reconstruction decisions and inability to effectively cope with topology optimization requirements in large-scale dynamic network scenarios.

[0004] Existing technologies often only consider immediate benefits or short-term optimization goals when formulating topology reconstruction strategies, lack evaluation mechanisms for long-term effects of reconstruction actions, and are easily trapped in local optimal solutions, resulting in frequent network adjustments that affect network stability, and also fail to effectively balance the relationship between network performance optimization and reconstruction cost, affecting overall network operation efficiency.

[0005] Existing methods lack systematic handling of network resource constraints and action conflicts when performing topology reconstruction, multiple reconstruction actions may compete for the same network resources, resulting in resource conflicts or deadlock phenomena in the actual execution process of the reconstruction scheme, and lack in-depth analysis and utilization of network evolution rules, making it impossible to achieve continuous intelligent optimization and adaptive evolution of network architecture. SUMMARY

[0006] The embodiment of the present application provides a deep reinforcement learning based adaptive network topology dynamic reconstruction method and system, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiment of the present application provides a deep reinforcement learning based adaptive network topology dynamic reconstruction method, comprising: Topology state information, traffic distribution information and historical reconstruction records of a current network are acquired; multi-level feature extraction is performed on the topology state information, and topology correlation features between nodes are captured through graph convolution operation, and a fusion state representation is generated in combination with the traffic distribution information; The fusion state representation and the historical reconstruction records are input into a deep reinforcement learning model to identify bottleneck nodes and redundant links in the topology structure, and output a reconstruction action candidate set; long-term cumulative returns of each candidate action are evaluated through Monte Carlo tree search to screen out an optimal reconstruction action sequence; The optimal reconstruction action sequence is verified in stages, a time slot is allocated to each reconstruction action through a graph coloring algorithm, action combinations with resource conflicts are identified and processed, and a resource feasible topology adjustment scheme is generated; The resource feasible topology adjustment scheme is converted into a multi-dimensional quantized feature sequence, network evolution rules are extracted using tensor decomposition technology, and a correlation mapping atlas of topology optimization is constructed through entropy weight ordering method to realize intelligent evolution of network architecture.

[0008] The multi-level feature extraction on the topology state information includes: The topology state information is constructed into a hierarchical graph structure representation, multi-hop neighborhood features are sampled based on node degree distribution characteristics, and a local topology subgraph is constructed for each node's multi-hop neighborhood features; Based on the local topology subgraph, graph convolution operation is used to describe the correlation strength distribution of nodes at different propagation levels, and the distance decay law and information transmission capacity between nodes are combined to generate multi-level topology features representing network structure evolution rules; Based on the node interaction mode implied in the multi-level topology features, a bidirectional mapping relationship atlas of node identifiers and traffic source and sink identifiers is established, traffic demand and transmission delay requirements are converted into attribute descriptions of node feature vectors, and a fusion state representation with space-time dependence characteristics is constructed through hierarchical combination.

[0009] Based on the node interaction mode implied in the multi-level topology features, a bidirectional mapping relationship atlas of node identifiers and traffic source and sink identifiers is established, traffic demand and transmission delay requirements are converted into attribute descriptions of node feature vectors, and a fusion state representation with space-time dependence characteristics is constructed through hierarchical combination. A node interaction relationship atlas is constructed from the multi-level topology features, the dynamic coupling strength of node pairs is described based on the message passing frequency and feature similarity between nodes, and the node identifiers are divided into multiple hierarchical communities with close interaction relationships according to the distribution characteristics of the dynamic coupling strength, and a hierarchical mapping system is constructed for each community; Based on the hierarchical mapping system, a positive feature diffusion path is constructed through the collaborative transmission effect of nodes in the community, and a multi-dimensional association graph of traffic source and sink identifiers is generated using the propagation law of the positive feature diffusion path, and the hierarchical mapping system is perfected according to the multi-dimensional association graph to form a bidirectional mapping knowledge network; According to the structural characteristics of the bidirectional mapping knowledge network, the business traffic demand is converted into dynamic load distribution between nodes based on the positive feature diffusion path, the transmission time delay requirement is mapped into the time sequence response characteristics of the nodes through the multi-dimensional association graph, and the dynamic load distribution and the time sequence response characteristics are integrated into the multi-dimensional attribute expression of the nodes using feature space reconstruction technology to form an attribute description with rich semantic information.

[0010] The fusion state representation and the historical reconstruction record are input into a deep reinforcement learning model to identify bottleneck nodes and redundant links in the topology structure, and output a reconstruction action candidate set; the long-term cumulative returns of each candidate action are evaluated through Monte Carlo tree search to select an optimal reconstruction action sequence including: The fusion state representation is input into a deep reinforcement learning model, the importance distribution of each node feature vector is calculated based on an attention mechanism, resource utilization rate analysis is performed on nodes with importance distribution exceeding a dynamic threshold, nodes meeting congestion determination conditions are marked as a bottleneck node group, and a redundant link group with low link utilization rate is identified; Based on the congestion state of the bottleneck node group and the utilization rate characteristics of the redundant link group, a candidate action set of link addition and link deletion is generated, and experience actions with similarity higher than a matching threshold to the current state are extracted from historical reconstruction records, the candidate action set and the experience actions are integrated to form a reconstruction action sequence; For the reconstruction action sequence, a Monte Carlo search tree is constructed with the current fusion state representation as the root node, state transition prediction is performed for each candidate action in the sequence, the evolution trend of the successor state is evaluated through the deep reinforcement learning model, and each evolution branch is expanded for multiple steps to calculate the long-term return index of the action sequence; The candidate action corresponding to the optimal value is selected from the long-term return index as the first choice, the first choice and candidate actions with suboptimal return indexes are combined for verification, the action combination that maximizes the overall network performance is selected, and an optimal reconstruction sequence with time sequence characteristics is generated.

[0011] The optimal reconstruction action sequence is verified in stages, each reconstruction action is assigned a time slot through graph coloring algorithm, action combinations with resource conflicts are identified and processed, and a resource feasible topology adjustment scheme is generated including: mapping the optimal reconstructed action sequence as a group of nodes of a graph coloring model, constructing a conflict constraint network by analyzing resource competition and dependency characteristics among reconstructed actions, marking the competition strength of node pairs with conflict constraint connections, and determining the minimum number of colors of the graph coloring model based on the topology of the conflict constraint network; allocating corresponding color batch identifiers to the group of nodes based on the structural characteristics of the conflict constraint network in the graph coloring model, so that node pairs with a competitive relationship are divided into different execution batches, combining reconstructed actions corresponding to node pairs with the same batch identifier into parallel execution units, and assigning execution time windows to the parallel execution units in priority order; For the parallel execution units within the execution time window, construct a matching relationship between resource demand profile and network capacity distribution, identify conflict time windows where resource demand exceeds capacity constraints, and dynamically reorganize the execution units in the conflict time windows to transfer some reconstructed actions to subsequent idle time windows, forming a resource feasible topology adjustment scheme.

[0012] Convert the resource feasible topology adjustment scheme into a multi-dimensional quantized feature sequence, extract network evolution rules using tensor decomposition technology, and construct a correlation mapping atlas for topology optimization using entropy weight ordering method, including: extracting multi-dimensional features of network evolution from the resource feasible topology adjustment scheme, and reorganizing the multi-dimensional features into a multi-dimensional quantized feature sequence according to the evolution time sequence of the topology structure; constructing the multi-dimensional quantized feature sequence into a three-dimensional state expression space, revealing the internal structure of the multi-dimensional quantized feature sequence through tensor decomposition technology, decoupling the three-dimensional state expression space into a combination of core features and pattern matrices, and identifying network evolution paths from the pattern matrix in the time dimension; Based on the coupling relationship between the core features and the pattern matrices, extract the cooperative evolution law of node connectivity and link load distribution, refine the cooperative evolution law into the essential evolution mode of network structure, and construct an evolution knowledge base reflecting the dynamic characteristics of the network; According to the cooperative mode in the evolution knowledge base, perform information entropy analysis on the multi-dimensional quantized feature sequence, convert the information entropy into a quantized indicator of feature importance, establish a mapping relationship between the feature importance and the network evolution direction, and form a correlation mapping atlas with predictive value.

[0013] According to the cooperative mode in the evolution knowledge base, perform information entropy analysis on the multi-dimensional quantized feature sequence, convert the information entropy into a quantized indicator of feature importance, establish a mapping relationship between the feature importance and the network evolution direction, and form a correlation mapping atlas with predictive value. extract a dynamic evolution mode of a topology structure from the evolution knowledge base, evaluate evolution activity degrees of the multi-dimensional quantitative feature sequence based on coupling rules in the dynamic evolution mode, mark a feature group with an evolution activity degree exceeding a preset activity threshold as an important evolution feature, and perform preliminary evaluation on a time sequence distribution of the important evolution feature; For the time sequence distribution characteristics of the important evolution feature, an information entropy evaluation system is constructed to quantify the uncertainty of its dynamic change, the information entropy index is multi-dimensionally fused with the evolution activity degree to form a quantitative weight system reflecting the importance of the feature; According to the quantitative weight system, the important evolution features are ranked and sorted, and a feature set at the top of the ranking is extracted as a dominant feature group of network evolution, and the development trend of the dominant feature group in the dynamic evolution mode is analyzed to reveal the evolution law of the network structure; The development trend of the dominant feature group is correlated with the historical optimization experience in the evolution knowledge base to identify the path of network performance improvement, and a mapping relationship between feature importance and network evolution direction is constructed.

[0014] The second aspect of the embodiment of the application provides a deep reinforcement learning-based adaptive network topology dynamic reconstruction system, which comprises: A first unit is configured to acquire topology state information, service traffic distribution information and historical reconstruction records of a current network, perform multi-level feature extraction on the topology state information, capture topology correlation features between nodes through graph convolution operation, and generate a fusion state representation in combination with the service traffic distribution information; A second unit is configured to input the fusion state representation and the historical reconstruction records into a deep reinforcement learning model, identify bottleneck nodes and redundant links in a topology structure, and output a reconstruction action candidate set; the long-term cumulative returns of each candidate action are evaluated through Monte Carlo tree search to filter out an optimal reconstruction action sequence; A third unit is configured to perform phased verification on the optimal reconstruction action sequence, assign time slots to each reconstruction action through a graph coloring algorithm, identify and process action combinations with resource conflicts, and generate a resource feasible topology adjustment scheme; A fourth unit is configured to convert the resource feasible topology adjustment scheme into a multi-dimensional quantitative feature sequence, extract network evolution rules by using tensor decomposition technology, and construct a topology optimization correlation mapping graph by using an entropy weight ranking method to realize intelligent evolution of a network architecture.

[0015] The third aspect of the embodiment of the application provides an electronic device, which comprises: a processor; a memory for storing processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method described above.

[0016] In a fourth aspect, the application provides a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method described above.

[0017] The application has the following beneficial effects: The application extracts multi-level features from topology state information, uses graph convolution operation to deeply capture topology correlation features between nodes, and fuses the features with service traffic distribution information to generate more comprehensive and accurate network state representation.

[0018] The application uses a deep reinforcement learning model combined with Monte Carlo tree search to not only identify bottleneck nodes and redundant links in the topology structure, but also evaluate the long-term cumulative returns of each candidate action to filter out the optimal reconstruction action sequence.

[0019] The application uses a graph coloring algorithm to allocate time slots and detect resource conflicts, uses tensor decomposition technology to extract network evolution rules, and uses an entropy weight ordering method to construct a correlation mapping atlas for topology optimization to realize a complete closed loop from a theoretical scheme to actual execution. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 FIG. 1 is a flowchart of an adaptive network topology dynamic reconstruction method based on deep reinforcement learning according to an embodiment of the application; Figure 2 FIG. 2 is a multi-level topology feature and dynamic coupling mapping construction flowchart according to an embodiment of the application. DETAILED DESCRIPTION

[0021] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0022] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0023] Figure 1 The flowchart of the adaptive network topology dynamic reconstruction method based on deep reinforcement learning of the embodiments of the present application is shown in FIG. 1. Figure 1 As shown in FIG. 1, the method comprises the following steps. Obtaining the topology state information, traffic distribution information and historical reconstruction record of the current network; performing multi-level feature extraction on the topology state information, capturing the topological correlation features between nodes through graph convolution operation, and generating fusion state representation in combination with the traffic distribution information; Inputting the fusion state representation and the historical reconstruction record into a deep reinforcement learning model to identify the bottleneck nodes and redundant links in the topology structure, and outputting a reconstruction action candidate set; evaluating the long-term cumulative returns of each candidate action through Monte Carlo tree search to filter out the optimal reconstruction action sequence; Verifying the optimal reconstruction action sequence in stages, assigning time slots to each reconstruction action through graph coloring algorithm, identifying and processing the action combination with resource conflicts, and generating a resource feasible topology adjustment scheme; Converting the resource feasible topology adjustment scheme into a multi-dimensional quantized feature sequence, extracting the network evolution law by using tensor decomposition technology, and constructing a correlation mapping atlas of topology optimization by using entropy weight ordering method to realize intelligent evolution of network architecture.

[0024] In an optional embodiment, the multi-level feature extraction on the topology state information, the capturing of the topological correlation features between nodes through graph convolution operation, and the generation of the fusion state representation in combination with the traffic distribution information comprise the following steps. The topology state information is constructed into a hierarchical graph structure representation, multi-hop neighborhood features are sampled based on node degree distribution characteristics, and a local topology subgraph is constructed for the multi-hop neighborhood features of each node. Based on the local topology subgraph, the graph convolution operation is used to depict the correlation strength distribution of nodes at different propagation levels, and combined with the distance decay law and information transmission ability between nodes, a multi-level topology feature representing the evolution law of network structure is generated. Based on the node interaction mode contained in the multi-level topology feature, a bidirectional mapping relationship graph between node identification and flow source and sink identification is established, the business traffic demand and transmission delay requirement are converted into attribute description of node feature vector, and a fusion state representation with space-time dependence characteristics is constructed through hierarchical combination.

[0025] The topology state information needs to be converted into a hierarchical graph structure representation, which takes an adjacency matrix as the basic data structure. In specific operation, all nodes in the network are constructed into an adjacency matrix according to their connection relationship in the physical topology. The rows and columns of the matrix correspond to each network node respectively. When there is a physical link connection between two nodes, the element value at the corresponding position in the matrix is set to 1, otherwise it is set to 0. After completing the construction of the basic adjacency matrix, the degree value of each node needs to be calculated, that is, the number of neighbor nodes directly connected to the node. By counting the degree value distribution of all nodes, the core nodes with high degree value and the edge nodes with low degree value are identified.

[0026] When sampling the multi-hop neighborhood features for each target node, the node is expanded outward as the center. The first-hop neighborhood contains all nodes directly connected to the target node, the second-hop neighborhood contains nodes directly connected to the first-hop neighborhood nodes but not in the first-hop neighborhood, and so on. The third-hop and more distant neighborhood nodes can be obtained. In the actual sampling process, for nodes with high degree value, the number of neighborhood nodes is too large, and a random sampling strategy needs to be used for control. For example, the first-hop neighborhood of a core switch node contains 80 directly connected nodes. To avoid high feature dimension, 30 of them are randomly selected as representative neighborhood nodes. When sampling, the number of nodes in different hop neighborhoods needs to be kept relatively balanced. If 30 nodes are sampled in the first hop, 25 nodes are suggested to be sampled in the second hop, and 20 nodes are sampled in the third hop, to ensure that the hierarchical distribution of feature sampling is reasonable.

[0027] When constructing a local topology subgraph, the target node and its multi-hop neighborhood nodes are taken as the node set of the subgraph, and the original connection relationship between these nodes is retained as the edge set of the subgraph. Assuming that a local subgraph is constructed for an access layer switch node, the first-hop neighborhood of the node contains 8 access devices, the second-hop neighborhood contains 3 aggregation layer switches, and the third-hop neighborhood contains 1 core router. The local subgraph contains 13 nodes, and the number of edges is determined according to the actual existing physical links between these nodes. Each local topology subgraph is independently stored as a graph data structure, including a node list, an edge list, and node attribute information.

[0028] The graph convolution operation performs feature propagation and aggregation operation on the local topological subgraph. In the first layer of graph convolution processing, each node collects the feature information of its direct neighbor nodes, and performs weighted summation on these neighbor features, and the weight is determined by the connection strength of the edge. If the link bandwidth between two nodes is 10 Gbps, the weight of the edge can be set to 10, and if the link bandwidth is 1 Gbps, the weight is set to 1. For a certain aggregation layer node, it is directly connected to 5 access layer nodes and 2 core layer nodes. In the first layer of graph convolution, the node will collect the feature vectors of the 7 neighbor nodes, and perform weighted average according to the respective link weights to obtain the aggregated features of the node in the first propagation level. The second layer of graph convolution operation continues to propagate information based on the output features of the first layer, and at this time the node can perceive the topological information within the two-hop range, and the third layer of graph convolution extends to the three-hop range.

[0029] The description of the correlation strength distribution needs to consider the distance decay law between nodes. The correlation strength of directly connected nodes is the highest, which is set to the reference value 1.0. The correlation strength of nodes with two-hop distance decays to 0.6, and the correlation strength of nodes with three-hop distance decays to 0.3. The correlation strength of nodes with four-hop or more distance decays to less than 0.1. The path length from a certain edge access node to a core routing node is 3 hops, and the correlation strength between the edge node and the core node is 0.3, which means that the influence of the state change of the edge node on the core node is relatively weak. The information transmission capacity is quantified by the forwarding capacity of the node. The forwarding capacity of the core router is 500 Gbps, and the information transmission capacity coefficient is set to 5.0. The forwarding capacity of the aggregation switch is 100 Gbps, and the information transmission capacity coefficient is 2.0. The forwarding capacity of the access switch is 20 Gbps, and the information transmission capacity coefficient is 0.5.

[0030] The multi-level topological features are composed of the output features of the three-layer graph convolution. The feature dimension of the first layer convolution output is a 64-dimensional vector, the second layer output is a 48-dimensional vector, and the third layer output is a 32-dimensional vector. The three-layer feature vectors are connected in dimension order to form a 144-dimensional multi-level topological feature vector. The feature vector encodes the topological structure information of the node in different receptive field ranges, and can reflect the local characteristics and global evolution law of the network structure.

[0031] When establishing the bidirectional mapping relationship between node identifiers and traffic source and sink identifiers, a source-sink correspondence table needs to be constructed, which records the correspondence between the source node identifier and the destination node identifier of the service traffic and the actual network node. The source identifier of a certain service traffic is server group A, and the destination identifier is server group B. By querying the mapping table, it is determined that the node ID corresponding to server group A in the network is node 15, and the node ID corresponding to server group B is node 42, thereby establishing the mapping relationship from the service traffic identifier to the topology node identifier. The reverse mapping relationship records the source and sink roles of the service traffic carried by each network node. Node 15 is the source node of 12 service flows and the destination node of 8 service flows.

[0032] When the traffic demand is converted into a node feature vector, the total traffic demand of the node as a source node or a destination node is accumulated and counted. The total traffic demand of a certain node as a source node is 150 Gbps, and the total traffic demand as a destination node is 200 Gbps. These two values are used as two-dimensional components of the node feature vector. The transmission delay requirement is also converted into a feature vector attribute. The average delay requirement of the related service flow of the node is counted. If 60% of the service flows related to the node require a delay of less than 10 milliseconds, 30% require a delay of less than 50 milliseconds, and 10% require a delay of less than 100 milliseconds, the time delay sensitivity feature value of the node is set to 0.75, indicating that the node has high time delay sensitivity.

[0033] When constructing the fusion state representation by layering and combining, the topology feature vector and the traffic feature vector are dimensionally spliced. The topology feature is a 144-dimensional vector, the traffic demand feature is a 2-dimensional vector, and the time delay feature is a 1-dimensional vector. After splicing, a 147-dimensional basic fusion feature is formed. The time-space dependence characteristic is realized by introducing a time window history feature. On the basis of the current time feature, the feature vectors of the previous time and the two previous times are superimposed. The feature vectors of the three times are combined by weighting to form a time-space fusion representation. The current time weight is set to 0.6, the previous time weight is set to 0.3, and the two previous time weights are set to 0.1. The weighted feature vectors of the three times are summed to obtain the final fusion state representation vector. This vector contains not only the topology structure information but also the service traffic distribution characteristics and the time sequence evolution law.

[0034] In an optional implementation, based on the node interaction mode implied in the multi-level topology feature, a bidirectional mapping relationship graph between node identifiers and traffic source and sink identifiers is established, and the traffic demand and the transmission delay requirement are converted into attribute descriptions of a node feature vector, including: construct a node interaction relationship spectrum from the multi-level topological features, depict the dynamic coupling strength of a node pair based on the message passing frequency and feature similarity between nodes, and divide the node identities into a plurality of hierarchical communities with close interaction relationships according to the distribution characteristics of the dynamic coupling strength, and construct a hierarchical mapping system for each community; construct a positive feature diffusion path based on the collaborative transmission effect of nodes in a community through the hierarchical mapping system, and generate a multi-dimensional association graph of traffic source and sink identities using the propagation law of the positive feature diffusion path, and perfect the hierarchical mapping system according to the multi-dimensional association graph to form a bidirectional mapping knowledge network; According to the structural characteristics of the bidirectional mapping knowledge network, the business traffic demand is converted into the dynamic load distribution between nodes based on the positive feature diffusion path, the transmission delay requirement is mapped into the timing response characteristics of the nodes through the multi-dimensional association graph, and the dynamic load distribution and the timing response characteristics are integrated into the multi-dimensional attribute expression of the nodes using feature space reconstruction technology, forming an attribute description with rich semantic information.

[0035] As shown in Figure 2 the method comprises: In the process of constructing the node interaction relationship spectrum, the connection state information and communication record data between nodes are extracted from the multi-level topological features. For any two nodes in the network, the system counts the number of message passing within a preset time window, which is set to 600 seconds. In a network instance containing 150 nodes, node A and node B pass messages 42 times within the statistical period, and node A and node C pass messages 8 times. The system also calculates the similarity of the node feature vectors, which include the processing capacity, storage capacity, link bandwidth and other 12-dimensional numerical values of the nodes. The components of the feature vector of node A are 8-core processor, 16GB memory, 100Mbps bandwidth and other parameters, and the feature vector of node B is 8-core processor, 32GB memory, 100Mbps bandwidth. By comparing the differences of the numerical values in each dimension of the two feature vectors and calculating the comprehensive deviation, the feature similarity of node A and node B is 0.87, and the feature similarity of node A and node C is 0.34.

[0036] The calculation of the dynamic coupling strength comprehensively considers the message passing frequency and the characteristic similarity. The coupling strength between node A and node B is obtained by normalizing the message passing frequency 42 to the interval of 0 to 1 to obtain 0.91, and then performing weighted average with the characteristic similarity 0.87, and the weight coefficients are 0.6 and 0.4 respectively, and finally the coupling strength value is 0.894. After the system calculates the coupling strength of all node pairs in the network, a 150 by 150 coupling strength matrix is formed. By analyzing the numerical distribution of the matrix, the system identifies that there are a total of 238 node pairs with coupling strength greater than 0.75. Cluster analysis is performed on these tightly coupled node pairs, and the system divides the 150 nodes into 5 hierarchical communities, and each community contains 38, 45, 32, 21, and 14 nodes respectively.

[0037] When constructing the hierarchical mapping system for each community, a three-layer mapping structure is established for the first community containing 38 nodes. The top layer is the community representative node layer, and the 3 nodes with the highest average coupling strength with other nodes in the community are selected as representatives. The average coupling strength of these 3 nodes with nodes outside the community is 0.82, 0.79, and 0.81 respectively. The middle layer is the functional grouping layer, and the 38 nodes are divided into data processing group, data storage group, and data forwarding group according to the business type attribute of the nodes, and each group contains 15, 12, and 11 nodes respectively. The bottom layer is the specific node layer, which records the detailed identification information of each node including IP address, MAC address, device serial number, etc. Cross-layer index table is established in the mapping system to record the mapping relationship from the top layer representative node to the middle layer group and from the middle layer group to the bottom layer specific node.

[0038] When constructing the forward feature diffusion path, the first community is taken as an example for processing, and the source node S in the community is selected, which needs to transmit feature information to the target node T. The system determines the diffusion path according to the dynamic coupling strength matrix, and the path selection rule is to select the node with the highest coupling strength and not yet visited at each hop. The feature vector of the source node S contains 12 initial values, and when transmitting to the first hop node N1, the values of each component of the feature vector remain unchanged but the path weight identifier 0.89 is added. After receiving the feature vector, node N1 fuses its own feature vector with the received feature vector, and the fusion method is to mix the corresponding dimension values according to the proportion of 0.5. The fused feature vector continues to transmit to the second hop node N2, and the weight identifier is updated to 0.82. Through multi-hop transmission, the feature information reaches the target node T from the source node S through 5 intermediate nodes, forming a complete diffusion path.

[0039] In the process of generating the multi-dimensional association graph of traffic source and sink identification, 1000 historical communication records in the network are collected. Each record contains source node identification, destination node identification, traffic size, transmission delay and other information. The system analyzes all records of source node M as traffic source, and finds that it has sent data to 32 different destination nodes. The 32 destination nodes are sorted by traffic size, and the top 10 destination nodes receive 76% of the total traffic. The system constructs an association graph for source node M containing geographic location dimension, business type dimension, traffic size dimension, and delay sensitivity dimension. The geographic location dimension identifies the network distance hop count between the source node and each destination node, the business type dimension identifies the application protocol type involved in the communication, there are 8 types, the traffic size dimension records the cumulative received traffic of each destination node, which ranges from 2GB to 450GB, and the delay sensitivity dimension identifies the maximum delay required by the destination node, which ranges from 50 milliseconds to 500 milliseconds.

[0040] When perfecting the hierarchical mapping system to form a bidirectional mapping knowledge network, the information of the multi-dimensional association graph is injected into the hierarchical mapping system in reverse. In the mapping system of the first community, the static attribute information of the node is originally recorded, and now the dynamic traffic association information is added. The representative node R originally only has identification information in the top layer, and now the list of commonly used destination nodes when it is a traffic source contains 18 node identifications. The data forwarding group in the middle layer increases the description of the traffic forwarding mode, which identifies that the group of nodes mainly handles cross-community traffic forwarding, with daily forwarding traffic reaching 2.3TB. The bottom layer node increases the bidirectional traffic statistics information, node X not only sends out daily average 340GB of traffic as a traffic source, but also receives daily average 280GB of traffic as a traffic sink.

[0041] When converting the business traffic demand into dynamic load distribution between nodes, a typical business scenario is processed: the source node needs to transmit a total of 800GB of data traffic to the destination node. According to the 6 feasible paths identified by the forward feature diffusion path, the system evaluates the carrying capacity of each path. The first path contains 4 intermediate nodes, and the current load rates of each node are 45%, 52%, 38%, and 61% respectively, and the remaining bandwidth capacities are 55Mbps, 48Mbps, 62Mbps, and 39Mbps respectively. The system calculates that the bottleneck bandwidth of this path is 39Mbps, and it takes about 47 hours to transmit 800GB of data in theory. By analyzing the carrying capacity of all paths, the system decomposes 800GB of traffic into 240GB, 310GB, 180GB, and 70GB, which are transmitted on four different paths respectively. The load distribution of each node on the path is dynamically adjusted, and the load rates of the nodes on the first path are updated to 58%, 66%, 51%, and 74%, forming a new load distribution state.

[0042] When the transmission delay requirement is mapped to the timing response characteristics of the nodes through the multi-dimensional correlation graph, the processing delay requirement is 100 milliseconds. According to the historical data of the delay sensitivity dimension in the correlation graph, the system identifies that the node set satisfying the delay requirement is a total of 67. Analyze the timing response characteristics of these nodes, including processing delay, queuing delay, and transmission delay. The average value of the processing delay of node P is 8 milliseconds, and the standard deviation is 3 milliseconds; the queuing delay is 2 milliseconds at light load and up to 35 milliseconds at heavy load; the transmission delay depends on the link bandwidth and the packet size, and the transmission delay of a 1500 byte packet on a 100 Mbps link is 0.12 milliseconds. The system maps the delay requirement to the timing constraint conditions that the nodes must satisfy, that is, the comprehensive requirements that the processing delay does not exceed 10 milliseconds, the queuing delay does not exceed 30 milliseconds, and the transmission delay does not exceed 60 milliseconds.

[0043] When the dynamic load distribution and the timing response characteristics are integrated into the multi-dimensional attribute expression of the node by using the feature space reconstruction technology, the original feature vector of node Q is expanded. The original feature vector contains 12 dimensions of static attributes, and now 6 dimensions of dynamic attributes are added. The newly added dynamic attributes include the current load rate value 68%, the remaining processing capacity value 3200 MIPS, the average queuing delay value 18 milliseconds, the traffic throughput rate value 78 Mbps, the connection number value 145, and the cache occupancy rate value 52%. The expanded 18-dimensional feature vector comprehensively describes the static capability and dynamic state of the node, providing rich semantic information support for subsequent network optimization decisions. The system completes the feature vector expansion for all 150 nodes in the network, forming a unified attribute description specification, and the attribute description of each node not only retains the original device configuration information, but also reflects the traffic transmission and delay constraint business demand characteristics in real time.

[0044] In an optional implementation, the fusion state representation and the historical reconstruction record are input into a deep reinforcement learning model to identify bottleneck nodes and redundant links in the topology structure, and output a reconstruction action candidate set; the long-term cumulative returns of each candidate action are evaluated through Monte Carlo tree search to filter out an optimal reconstruction action sequence including: The fusion state representation is input into a deep reinforcement learning model, the importance distribution of each node feature vector is calculated based on an attention mechanism, resource utilization analysis is performed on nodes whose importance distribution exceeds a dynamic threshold, nodes that meet the congestion determination condition are marked as a bottleneck node group, and a redundant link group with low link utilization is identified; Based on the congestion state of the bottleneck node group and the utilization characteristics of the redundant link group, a candidate action set of link addition and link deletion is generated, and experience actions with a similarity higher than a matching threshold to the current state are extracted from the historical reconstruction record. The candidate action set and the experience actions are integrated to form a reconstruction action sequence. For the reconstructed action sequence, a Monte Carlo search tree is constructed with the current fusion state representation as the root node, state transition prediction is performed for each candidate action in the sequence, the evolution trend of the successor state is evaluated through the deep reinforcement learning model, each evolution branch is expanded for multiple steps, and the long-term reward indicator of the action sequence is calculated; The candidate action corresponding to the optimal value is selected from the long-term reward indicator as the preferred solution, and the preferred solution and the candidate action with a suboptimal reward indicator are combined for verification to select the action combination that maximizes the overall network performance and generate the optimal reconstruction sequence with timing characteristics.

[0045] The fusion state representation and historical reconstruction records are input to the deep reinforcement learning model through a standardized interface. The model uses an Actor-Critic architecture, including two sub-modules: a policy network and a value network. The fusion state representation is a 512-dimensional floating-point vector with a value range of 0 to 1, and the historical reconstruction records are stored in time series format, including action type, execution time, affected nodes, and effect evaluation. The input layer of the model is set to 640 dimensions, of which 512 dimensions receive the fusion state representation and 128 dimensions receive the encoded vector of the historical records. The historical record encoding uses position embedding, and the time decay weight is calculated as an exponential function with a decay factor of 0.92 and a maximum backtracking length of 50 historical actions.

[0046] The attention mechanism is embedded in the second layer of the model, using a multi-head self-attention structure with 16 heads and a dimension of 32 per head. In the attention weight calculation process, the node feature vector generates a query matrix, a key-value matrix, and a numerical matrix through three independent linear transformation layers, and the weight matrix of the transformation layer has a dimension of 512x512. The calculation of scaled dot-product attention uses a softmax activation function, and the temperature parameter is set to 0.1 to enhance the discriminability of the weight distribution. The importance distribution output is a probability vector with a length equal to the number of network nodes, and the sum of the vector elements is 1, with a value range of 0 to 1 for each element.

[0047] The dynamic threshold uses an adaptive adjustment strategy, with an initial threshold set to the 75th percentile of the network node importance distribution, and an adjustment period of every 10 minutes. The threshold update uses an exponential moving average method with a smoothing coefficient of 0.8 and upper and lower limit restrictions of 0.3 and 0.9, respectively. When the node importance distribution exceeds the dynamic threshold, the resource utilization analysis module is triggered, which obtains the CPU usage, memory occupancy, network interface bandwidth utilization, and queue length of the node from the monitoring system in real time. The resource utilization data collection interval is 30 seconds, and the analysis window length is the latest 300 sampling points, about 150 minutes of historical data.

[0048] The congestion determination condition is based on a multi-dimensional threshold combination, and the specific rules are that the CPU usage rate is continuously above 80% for 5 minutes, the memory occupancy rate is above 85%, or the network interface bandwidth utilization rate is above 95% in any 1 minute. The duration requirement of the congestion state is that the time meeting the determination condition is accumulated to more than 8 minutes. The formation of the bottleneck node group adopts a graph clustering algorithm, and the nodes meeting the congestion condition within 2 hops are grouped into the same group, and the maximum number of nodes in a single group is limited to 8. The identification of the redundant link group is based on statistical analysis of link utilization, and the statistical window is the past 72 hours. When the average utilization of the link is less than 15%, the 95th percentile utilization is less than 35%, and the peak utilization is less than 50%, it is marked as a redundant link. Adjacent redundant links are merged into a redundant link group through connectivity analysis, and the number of links in the group is limited to between 3 and 10.

[0049] The candidate action set generation module constructs link addition actions and link deletion actions according to the characteristics of the bottleneck node group and the redundant link group. The link addition strategy includes three types of near connection, center node connection and load dispersion connection. The near connection selects the node with the lightest load among the 2-hop neighbors of the bottleneck node to establish a link, the center node connection selects the top 10% nodes as the connection target based on the betweenness centrality of the network topology, and the load dispersion connection connects the bottleneck node to the node with a load of less than 30%. The link deletion strategy preferentially deletes the redundant link with the least impact on network connectivity, and the impact of deletion is evaluated by calculating the edge connectivity and the critical path coverage rate. The edge connectivity threshold is set to 3, and the critical path coverage rate threshold is set to 5%.

[0050] The experience action extraction in the history reconstruction record adopts a state space similarity matching algorithm, and the similarity calculation is based on the cosine similarity of the fused state representation, and the calculation complexity is O(n), where n is the dimension of the state vector. The matching process first arranges the historical records in reverse chronological order, and then calculates the similarity between the current state and the historical state one by one. When the similarity is greater than 0.75, the corresponding reconstruction action is extracted. The effectiveness verification of the experience action is based on the effect evaluation after execution, and only the action with an effect score greater than 7.0 is included in the candidate set. The integration of the candidate action set and the experience action adopts a probability weighting method, and the selection probability of the newly generated action is 0.6, and the selection probability of the experience action is 0.4. The length of the integrated action sequence is controlled to be within 15.

[0051] The Monte Carlo search tree is constructed with the current fusion state representation as the root node state. The tree node data structure includes six attributes: state vector, parent node pointer, child node list, visit count, cumulative reward, and average reward. The maximum depth of the search tree is limited to 6 layers, and the upper limit of the number of child nodes expanded at each layer is 8. The total number of nodes is controlled within 5000. State transition prediction is achieved through forward inference of a deep reinforcement learning model. The input is the concatenation of the current state vector and the action encoding, which is in one-hot format with a dimension equal to the size of the action space. The batch size for model inference is set to 32, and the single inference time is controlled within 50 milliseconds.

[0052] The evolution trend evaluation uses a multi-step guidance method, with each evolution branch independently performing 3-step state transition prediction and selecting actions through a deterministic policy. The value function evaluation is based on a neural network model trained using the time difference learning algorithm. The network structure consists of 3 fully connected layers with 256, 128, and 64 neurons in the hidden layers, respectively. During the multi-step expansion process, the uncertainty of action execution is considered, with a success probability of 0.9 for each action. In the case of failure, the state remains unchanged but a negative reward is generated.

[0053] The long-term revenue indicator integrates network performance, implementation cost, and risk assessment in three dimensions. Network performance indicators include end-to-end delay, network throughput, load balancing, and reliability, with weights of 0.3, 0.3, 0.2, and 0.2, respectively. Delay calculation is based on the shortest path algorithm, throughput is estimated by the maximum flow algorithm, load balancing is measured by the Gini coefficient, and reliability is evaluated by network connectivity. Implementation costs include equipment procurement costs, installation and deployment costs, and operation and maintenance costs. The cost model is based on a linear combination of link type, transmission distance, and bandwidth capacity. Risk assessment considers service interruption risk, configuration error risk, and rollback complexity during implementation, and uses a combination of expert scoring and historical statistics for quantification. The revenue indicator uses the net present value calculation method, with a discount rate of 8% and an evaluation period of 24 months.

[0054] The optimal value selection is achieved through revenue indicator sorting, with the first choice corresponding to the maximum revenue indicator value, and the requirement that the revenue indicator be greater than the current benchmark value by 20%. The candidate actions for the suboptimal revenue indicator are defined as the set of actions whose revenue indicators are in the top 30% and differ from the optimal value by less than 15%. Combination verification uses a discrete event simulation method, with the simulation environment based on the current network topology and traffic pattern. The simulation time step is 1 minute, and the total simulation time is 7 days. During the verification process, the dependency between actions and resource competition are considered, with the number of simultaneously executed actions limited to 3 and the minimum time interval between actions set to 15 minutes. The overall network performance maximization is achieved through a multi-objective optimization algorithm, which uses a weighted aggregation method to combine multiple performance indicators into a single objective function.

[0055] The time sequence feature modeling adopts a long short-term memory network, which includes two layers of LSTM units, with a hidden state dimension of 128 and a sequence length of the number of actions. The time-dependent relationship is represented by an attention weight matrix, and the calculation of the weight matrix takes into account the action type similarity and time interval influence. The optimal reconstruction sequence generation adopts a greedy search strategy, starting with the action with the highest return, and then selecting subsequent actions that are compatible with the selected actions and have higher returns. The sequence length is controlled between 3 and 8 actions, and the total execution time does not exceed 4 hours. Resource conflict detection is achieved through a graph coloring algorithm.

[0056] In an alternative embodiment, the optimal reconstruction action sequence is verified in stages, and a time slot is assigned to each reconstruction action through a graph coloring algorithm. The action combination with resource conflict is identified and processed, and a resource feasible topology adjustment scheme is generated, including: Mapping the optimal reconstruction action sequence to a node group of the graph coloring model, constructing a conflict constraint network by analyzing the resource competition and dependency characteristics between reconstruction actions, marking the competition intensity of node pairs with conflict constraint connections, and determining the minimum number of colors of the graph coloring model based on the topology structure of the conflict constraint network; Based on the structural characteristics of the conflict constraint network in the graph coloring model, assign corresponding color batch identifiers to the node group, so that node pairs with competition relationship are divided into different execution batches, combine reconstruction action pairs corresponding to the same batch identifier into parallel execution units, and assign execution time windows to the parallel execution units in priority order; For the parallel execution units in the execution time window, construct a matching relationship between resource demand profile and network capacity distribution, identify conflict time windows where resource demand exceeds capacity constraints, and dynamically reorganize the execution units in the conflict time windows, transfer part of the reconstruction actions to the subsequent idle time windows, and form a resource feasible topology adjustment scheme.

[0057] The optimal reconstruction action sequence is mapped to a node group of the graph coloring model through an action identifier, and each reconstruction action corresponds to a node in the graph model. The node attributes include four fields: action type, resource demand, execution duration, and priority. The action type is divided into four categories: link addition, link deletion, node configuration modification, and route adjustment, and is stored using enumeration value encoding. The resource demand field records the computing resources, network bandwidth, and storage space required for action execution. The computing resources are represented by CPU core number, the network bandwidth is in units of Mbps, and the storage space is in units of GB. The execution duration is estimated through historical statistical data, and the duration distribution is modeled using a lognormal distribution with a mean parameter of 5 minutes and a standard deviation parameter of 2 minutes.

[0058] The resource competition characteristics between the reconstructed actions are determined by resource occupancy matrix analysis, and the matrix dimension is the square of the number of actions. The matrix elements represent the resource overlap between two actions. The resource overlap is calculated based on the intersection ratio involving the device, and a competition relationship occurs when two actions operate the same device or adjacent devices. The dependency characteristic analysis considers the logical dependencies and timing constraints between actions. Logical dependencies include pre-condition dependencies and configuration consistency dependencies. Timing constraints require that certain actions must be executed in a specific order. The conflict constraint network is stored in an undirected graph structure. The edges in the graph represent the conflict constraints between two actions, and the edge weights represent the conflict strength, which is divided into high, medium, and low levels, corresponding to weight values of 3, 2, and 1.

[0059] The node pair marking of the conflict constraint connection is achieved by adjacency matrix traversal. During the traversal process, it is detected whether there is a direct conflict or an indirect conflict between each pair of nodes. A direct conflict refers to two actions operating the same resource, and an indirect conflict refers to a cascading influence relationship between actions. The competition intensity evaluation is based on three dimensions: resource conflict degree, influence range, and recovery difficulty. Each dimension uses a scoring standard of 0 to 10. The resource conflict degree is calculated based on the proportion of shared resources, the influence range is measured by the number of affected nodes, and the recovery difficulty is evaluated based on configuration complexity and rollback time. The competition intensity is finally taken as the weighted average of the three dimensions, with weights of 0.5, 0.3, and 0.2.

[0060] The minimum number of colors in the graph coloring model is determined by combining the greedy coloring algorithm and the exact algorithm. The greedy algorithm uses a degree sorting strategy, which prioritizes assigning colors to nodes with the highest degree. If the degrees are the same, the actions are sorted by priority. The algorithm maintains a color usage state table during the iteration process, which records the set of nodes that have been assigned each color. The exact algorithm is implemented based on backtracking search, with a search depth limit of the total number of nodes. The backtracking pruning condition is that the number of currently used colors exceeds the known upper bound. The theoretical lower bound of the minimum number of colors is estimated by the maximum clique algorithm. In the actual coloring process, the number of colors is controlled within 1.5 times the theoretical lower bound.

[0061] The color batch identification of the node group is assigned using a sequential coloring strategy, and nodes with the same color are grouped into the same execution batch. Load balancing constraints are considered during the coloring process, with the number of actions in a single batch limited to 8 or less, and the total resource demand not exceeding 80% of the available resources. Color allocation prioritizes meeting the needs of high-priority actions, with priority scores calculated based on the expected improvement in network performance and the urgency of the action. Node pairs with competition relationships are identified through edge connections in the conflict constraint network. Nodes with edge connections must be assigned different colors to ensure that they are not executed in the same time window.

[0062] Parallel execution units are formed by the combination of reconfiguration actions with the same batch identifier, and actions within a unit are executed in parallel. The parallelism degree is dynamically adjusted based on resource availability and execution complexity, and the upper limit of parallel actions in a single execution unit is 6. The data structure of an execution unit contains four attributes: action list, resource requirement summary, estimated execution time, and dependency relationship. The resource requirement summary is calculated by vector addition, the estimated execution time is taken from the execution time of the longest action in the unit, and the dependency relationship records the precedence constraints with other execution units.

[0063] The allocation of execution time windows uses a time slice round-robin scheduling algorithm, with a time slice size of 15 minutes and a window interval of 5 minutes of buffer time. The priority order is determined based on the comprehensive score of the execution unit, which includes performance improvement potential, implementation urgency, and resource efficiency. Performance improvement potential is predicted by network simulation, implementation urgency is evaluated based on fault risk and business impact, and resource efficiency is calculated by input-output ratio. Maintenance time constraints are considered in the time window allocation process to avoid executing reconfiguration actions during peak periods, which are defined as periods when network traffic exceeds 150% of the average.

[0064] Resource requirement profiling is represented by a multi-dimensional feature vector, with dimensions including CPU usage, memory consumption, network bandwidth occupation, storage IO demand, and concurrent connection number. The profile is generated based on statistical analysis of historical execution data, using a sliding average method to handle data fluctuations with a window length of 20 historical execution records. Network capacity distribution is obtained through real-time monitoring data, with monitoring indicators including device resource usage, link bandwidth utilization, and cache hit rate. Capacity data is updated every 2 minutes, with a data accuracy guarantee within a 5% error range.

[0065] Matching relationships are established through comparative analysis of resource requirements and available capacity, using an approximate solution method for the multi-dimensional bin packing problem. The algorithm considers the heterogeneity and substitutability of resource types, allowing compensation for insufficient dimensions with surplus dimensions. The matching success threshold is set to 90% of the available capacity, with a safety margin of more than 10%. When matching fails, resource scheduling optimization is triggered to alleviate capacity bottlenecks through dynamic resource allocation and load migration.

[0066] Conflict time window identification is based on a resource over-limit detection algorithm, with a detection period of 3 minutes before the start of each time window. Over-limit conditions include demand exceeding available capacity in any resource dimension, or a comprehensive load index exceeding 0.95 in multiple dimensions. The comprehensive load index is calculated by weighted average, with weights set according to resource importance: CPU and memory weights are 0.3, network bandwidth weight is 0.25, and storage and connection number weights are each 0.075. Detection results are stored in a conflict event table, recording conflict type, involved resources, duration, and impact degree.

[0067] The dynamic reconfiguration of execution units employs a heuristic optimization algorithm, with the objective of minimizing total execution time and resource waste. The reconfiguration strategies include action splitting, batch merging, and time window adjustment. Action splitting decomposes complex actions into multiple sub-actions, which are distributed across different time windows for execution, with the order of execution guaranteed by dependency relationships between sub-actions. Batch merging combines execution units with complementary resource requirements, improving resource utilization efficiency. Time window adjustment is achieved through a sliding window technique, with the window size dynamically adjustable within a range of 10 minutes to 60 minutes.

[0068] The transfer of reconfigured actions to subsequent idle time windows is achieved through a time scheduling algorithm, which maintains a global schedule and a resource occupancy table. Idle window identification is based on a resource utilization threshold, with any window marked as idle when the resource utilization within it is below 50%. The transfer strategy prioritizes the nearest idle window, while considering the dependency constraints between actions and the requirement for business continuity. During the transfer process, the execution plan and resource reservation for the actions are updated to ensure that the adjusted solution still meets the performance targets.

[0069] The generation of resource-feasible topology adjustment schemes is achieved through constraint satisfaction problem solving, with constraint conditions including resource capacity constraints, timing dependency constraints, and performance target constraints. The solving algorithm employs a branch-and-bound method, with search space pruning based on constraint propagation and heuristic estimation. The feasibility of the schemes is verified through simulation execution, with the simulation environment replicating the resource state and load characteristics of the real network environment. The output of the verified schemes includes three parts: execution schedule, resource allocation plan, and risk assessment report.

[0070] In an alternative embodiment, the resource-feasible topology adjustment schemes are converted into a multi-dimensional quantified feature sequence, the network evolution law is extracted using tensor decomposition technology, and the correlation mapping atlas for topology optimization is constructed through entropy weight ordering method, including: extracting multi-dimensional features of network evolution from the resource-feasible topology adjustment schemes, and reorganizing the multi-dimensional features into a multi-dimensional quantified feature sequence according to the evolution time sequence of the topology structure; constructing the multi-dimensional quantified feature sequence into a three-dimensional state expression space, revealing the internal structure of the multi-dimensional quantified feature sequence through tensor decomposition technology, decoupling the three-dimensional state expression space into a combination of core features and pattern matrices, and identifying the network evolution path from the pattern matrix in the time dimension; based on the coupling relationship between the core features and the pattern matrices, extracting the cooperative evolution law of node connectivity and link load distribution, refining the cooperative evolution law into the essential evolution mode of network structure, and constructing an evolution knowledge base reflecting the dynamic characteristics of the network; According to the coordination mode in the evolution knowledge base, information entropy analysis is performed on the multi-dimensional quantified feature sequence, information entropy is converted into a quantified index of feature importance, a mapping relationship between the feature importance and the network evolution direction is established, and an associated mapping atlas with predictive value is formed.

[0071] The resource feasible topology adjustment scheme is converted into multi-dimensional features of network evolution by a feature extraction module. The module uses a sliding window technique to scan each execution time point in the adjustment scheme. The feature extraction window is set at 5-minute intervals, covering the time span of the entire topology adjustment process. The multi-dimensional features include network topology features, traffic distribution features, resource utilization features, and performance indicator features. The network topology features are obtained by eigenvalue decomposition of the adjacency matrix, including four indicators of connectivity, clustering coefficient, average path length, and degree distribution entropy. The traffic distribution features are based on traffic matrix analysis, extracting four parameters of traffic gravity center, load balancing index, peak-to-valley ratio, and spatial correlation. The resource utilization features cover the statistical features of CPU usage, memory occupancy, bandwidth utilization, and cache hit rate, including mean, variance, skewness, and kurtosis. The performance indicator features include the time series statistics of end-to-end delay, packet loss rate, throughput, and jitter.

[0072] The multi-dimensional features are reorganized into a quantified feature sequence according to the evolution time sequence of the topology structure. The reorganization process uses time alignment and interpolation techniques to handle feature data with different sampling frequencies. The time alignment benchmark is set to the minimum sampling interval, and the interpolation method uses cubic spline interpolation to ensure data continuity. The data structure of the quantified feature sequence is represented by a three-dimensional tensor. The first dimension corresponds to the time sequence, with a length equal to the time steps of the evolution process. The second dimension corresponds to the network node or link identifier, and the third dimension corresponds to the feature type. Feature value normalization uses the maximum and minimum value normalization method to ensure the comparability of features with different dimensions. The sequence length is dynamically determined according to the complexity of the adjustment scheme, with 10 time steps corresponding to a single adjustment action, and the total sequence length controlled within 200 time steps.

[0073] The three-dimensional state expression space is constructed through tensor reshaping operation, which reorganizes the multi-dimensional quantified feature sequence into a three-order tensor with time dimension, space dimension, and attribute dimension. The length of the time dimension is equal to the number of evolution time steps, the length of the space dimension is equal to the total number of nodes and links in the network, and the length of the attribute dimension is equal to the number of extracted feature types. The data type of tensor elements is 64-bit floating point number, with a value range limited between 0 and 1. Tensor storage uses compressed sparse format to reduce memory occupancy by taking advantage of the sparsity of network topology, with a compression ratio usually exceeding 70%. The three-dimensional state expression space supports slicing and aggregation operations, facilitating subsequent tensor decomposition processing.

[0074] The tensor decomposition technique adopts CANDECOMP or PARAFAC decomposition method to decompose the three-dimensional state representation space into a linear combination of several rank-one tensors. The decomposition algorithm uses the alternating least squares method to solve, and the maximum number of iterations is set to 1000 times, and the convergence tolerance is set to ten to the minus sixth power. The rank number of decomposition is determined by cross-validation, and the candidate rank number range is 5 to 50, and the rank number that minimizes the reconstruction error is selected as the final parameter. The algorithm implementation adopts multi-thread parallel computing, and the number of threads is set to half of the CPU core number, and the memory requirement is about 3 times the size of the original tensor.

[0075] The decomposition result is represented as a combination of core features and mode matrices. The core feature is a one-dimensional vector with a length equal to the decomposition rank, representing the importance weight of each mode component. The mode matrix includes a time mode matrix, a space mode matrix, and an attribute mode matrix. The time mode matrix has a row number equal to the time dimension length and a column number equal to the decomposition rank. The matrix elements represent the activation degree of each time point in different modes. The space mode matrix and the attribute mode matrix have similar structures and meanings. The normalization of the mode matrix uses column normalization method to ensure that the norm of each mode component is 1.

[0076] The network evolution path in the time dimension mode matrix is identified by trajectory clustering algorithm. The clustering algorithm uses the DBSCAN method based on density. The clustering parameters include a neighborhood radius of 0.1 and a minimum sample number of 5. The distance metric uses Euclidean distance. The evolution path is represented as a continuous trajectory in the time mode space. Trajectory smoothing uses a Gaussian filter with a window size of 7 time steps. Path features include start state, end state, path length, curvature, and speed change. Path classification is based on shape similarity measurement, with a similarity threshold of 0.8, forming a class library of typical evolution patterns.

[0077] The coupling relationship between core features and mode matrices is quantified by correlation analysis. The correlation calculation uses the Pearson correlation coefficient. The coupling strength threshold is set to 0.3, and the features with coupling strength exceeding the threshold are marked as strong coupling. The collaborative evolution law of node connectivity and link load distribution is extracted by multivariate time series analysis. The analysis window length is 20 time steps, and the sliding step is 5 time steps. Collaborative pattern recognition is based on Granger causality test with a test significance level of 0.05. The evolution law is quantified as a state transition probability matrix with a dimension of 10x10.

[0078] The essential evolution pattern of network structure is extracted by a pattern mining algorithm, which adopts the Apriori method of frequent sequence mining. The minimum support is set to 0.1, the minimum confidence is set to 0.7, and the maximum sequence length is limited to 8 state changes. The evolution pattern is represented as a directed graph structure, where the nodes represent network states, the edges represent state transitions, and the edge weights represent transition probabilities. The pattern evaluation is based on two indicators: information gain and prediction accuracy. The information gain threshold is set to 0.2, and the prediction accuracy threshold is set to 0.75.

[0079] The evolution knowledge base is constructed using a graph database to store evolution patterns, supporting complex graph queries and pattern matching operations. The knowledge base graph includes four entity types: state nodes, transition edges, pattern labels, and attribute fields. State nodes store feature vectors and timestamps of network states, transition edges store transition probabilities and trigger conditions, pattern labels are used for classification and retrieval, and attribute fields record relevant quantitative indicators. The knowledge base is updated using incremental learning, with new evolution data integrated into the existing knowledge structure through pattern matching and similarity calculation.

[0080] The information entropy analysis of multi-dimensional quantitative feature sequences uses the Shannon entropy calculation method, with entropy values calculated based on the probability distribution of feature values. The probability distribution is obtained through a histogram estimate, with the number of histogram bins set to the square root of the feature sequence length. The calculation window for information entropy is set to 50 time steps, with a sliding step size of 10 time steps, ensuring the time resolution of the entropy value sequence. The entropy value is normalized using the Shannon entropy with a base of 2, with a value range of 0 to the logarithmic feature space dimension.

[0081] The conversion of information entropy into a feature importance quantitative indicator is achieved through a weight assignment function, which uses the softmax function to ensure that the weights sum to 1. The feature importance is inversely proportional to the information entropy, with lower entropy values indicating more important features. The weight calculation formula is the exponential normalization of the negative entropy value. The importance threshold is set to 1.5 times the average importance, and features exceeding this threshold are marked as key features. The stability of key features is verified through guided program sampling, with 1000 sampling times. The stability indicator is the Kendall correlation coefficient of the importance ranking, and the stability threshold is set to 0.8.

[0082] The mapping relationship between feature importance and network evolution direction is established through regression analysis, with the regression model using a multivariate linear regression method. The evolution direction is quantified as a change vector of network performance indicators, including time delay change, throughput change, and reliability change. The input of the regression model is the feature importance vector, and the output is the evolution direction vector. The model training uses the least squares method to solve. The model verification uses the cross-validation method, with 5 validation folds. The performance evaluation indicators are the root mean square error and the correlation coefficient.

[0083] The correlation mapping graph adopts a weighted directed graph representation, in which nodes represent features or evolution states, edges represent correlation relationships, and edge weights represent correlation strengths. The graph construction algorithm is based on correlation analysis of feature importance and evolution direction, with a correlation threshold of 0.4. The visualization of the graph uses a force-directed layout algorithm, with node size representing feature importance, edge thickness representing correlation strength, and color coding representing evolution direction categories. The graph update adopts an online learning method, with new data integrated into the existing structure through graph embedding technology, and the embedding dimension set to 128 dimensions.

[0084] The prediction value evaluation is verified by a time series prediction task, and the prediction model is implemented based on a graph neural network of the correlation mapping graph. The network structure includes two graph convolution layers and one fully connected layer, with a hidden layer dimension of 64 and a ReLU activation function. The model training uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 200 training rounds. The prediction accuracy is evaluated using the mean absolute percentage error, with a target accuracy of within 15%.

[0085] In an alternative embodiment, according to the coordination mode in the evolution knowledge base, the information entropy of the multi-dimensional quantitative feature sequence is analyzed, the information entropy is converted into a quantitative indicator of feature importance, and the mapping relationship between the feature importance and the network evolution direction is established, including: Extracting the dynamic evolution mode of the topological structure from the evolution knowledge base, evaluating the evolution activity degree of the multi-dimensional quantitative feature sequence based on the coupling law in the dynamic evolution mode, marking the feature group with an evolution activity degree exceeding a preset activity threshold as an important evolution feature, and preliminarily evaluating the time series distribution of the important evolution feature; Constructing an information entropy evaluation system to quantify the uncertainty of dynamic changes of the important evolution feature according to the time series distribution characteristics of the important evolution feature, and forming a quantitative weight system reflecting the importance of the feature by multi-dimensionally integrating the information entropy indicator and the evolution activity degree; According to the quantitative weight system, the important evolution features are ranked and sorted, and the top-ranked feature set is extracted as the dominant feature group of network evolution, and the development trend of the dominant feature group in the dynamic evolution mode is analyzed to reveal the evolution law of the network structure; The development trend of the dominant feature group is correlated with the historical optimization experience in the evolution knowledge base to identify the path of network performance improvement, and the mapping relationship between feature importance and network evolution direction is constructed.

[0086] When extracting the dynamic evolution pattern of the topology structure from the evolution knowledge base, the system identifies the change path of the node connection relationship by traversing the historical network state data stored in the knowledge base. Specifically, the system reads at least 1000 groups of historical topology snapshot data contained in the knowledge base, and each group of snapshot records the connection state of each node, the edge weight and the node attribute in the network at a specific time. By comparing and analyzing the snapshots of adjacent time, the system extracts the basic evolution events such as the increase and decrease of nodes, the creation and deletion of edges, and the change amplitude of weight. In the extraction process, the system organizes these basic events into an evolution sequence in chronological order, and calculates the occurrence frequency of different evolution events by using the sliding time window method. The time window is set to 10 time units, and the window slides by 2 time units as a step, and the numerical fluctuation range of the node degree change, the clustering coefficient change, the path length change and other characteristics is calculated in each window.

[0087] When evaluating the evolution activity degree of the multi-dimensional quantitative feature sequence based on the extracted dynamic evolution pattern, a coupling rule evaluation mechanism is established. This mechanism reflects the dynamic change intensity of the feature by calculating the difference value of the numerical values at adjacent time points in the feature sequence. For a feature sequence containing 50 time points, the system calculates 49 difference values, and calculates the absolute value sum of these difference values as the change intensity indicator. At the same time, the system analyzes the cooperative change relationship between the feature sequence and other feature sequences, and quantifies the coupling strength by calculating the correlation degree of the difference value sequences of two feature sequences. When the coupling strength of a certain feature sequence with at least three other feature sequences exceeds 0.7, the feature is identified as having significant cooperative evolution characteristics. The evolution activity is obtained by comprehensively considering the change intensity of the feature itself and the coupling strength with other features. Specifically, the change intensity is normalized to the interval of 0 to 1, and then weighted average is performed with the average coupling strength, and the weights are set to 0.6 and 0.4 respectively.

[0088] The preset active threshold is set to 0.55, and the features with evolution activity exceeding the threshold are marked as important evolution features. In an actual case containing 200 original features, 43 important evolution features are screened out by activity evaluation. When preliminarily evaluating the time sequence distribution of these important evolution features, the system analyzes the distribution density and aggregation characteristics of the feature values on the time axis. The specific method is to divide the entire time range into 10 equal time periods, and count the number of changes and the change amplitude of the feature values in each time period. When the number of changes in a certain time period exceeds 20% of the total number of changes, the time period is marked as an active period of feature evolution. In the actual case, it is found that the important evolution features show obvious aggregation changes in the 3rd and 7th time periods, and the cumulative change amplitude of these two time periods accounts for 65% of the total change amplitude.

[0089] When constructing the information entropy evaluation system for the time sequence distribution characteristics of important evolutionary features, the distribution state of the feature value is taken as the analysis object. The system divides the value range of the feature value into 20 equal intervals, and counts the frequency of the feature value falling into each interval. The probability distribution of the feature value is obtained by calculating the proportion of the frequency of each interval in the total frequency. The information entropy is calculated by taking the negative value of the sum of the logarithm of the probability value multiplied by the probability value. In specific implementation, for a feature with a value range of 0 to 100, it is divided into 20 intervals with a width of 5, and the sample numbers of each interval are obtained as 8, 12, 15, 20, 18, etc. After calculating the probability of each interval, the information entropy value of the feature is obtained as 2.35 through logarithmic operation and weighted summation. A higher information entropy value indicates that the feature value distribution is more uniform, and the uncertainty of change is higher, reflecting that the feature has strong randomness and complexity in the evolution process.

[0090] When the information entropy index is fused with the evolution activity in multiple dimensions, a comprehensive quantitative weight calculation mechanism is established. This mechanism normalizes the information entropy value and multiplies it with the evolution activity to obtain the preliminary fusion weight. Considering the contribution difference of information entropy and activity in different scenarios, the system introduces an adaptive adjustment factor. The adjustment factor is dynamically set according to the stationarity of the feature sequence. When the standard deviation of the feature sequence is less than 10% of the mean, the adjustment factor tilts towards information entropy and is set to 1.2. When the standard deviation is greater than 30% of the mean, the adjustment factor tilts towards activity and is set to 0.8. In actual cases, the normalized value of a certain feature's information entropy is 0.78, the evolution activity is 0.62, and the ratio of standard deviation to mean is 0.15. The corresponding adjustment factor is 1.05. The final quantitative weight of the feature is calculated as the product of 0.78 and 0.62 multiplied by the adjustment factor, which is 0.51.

[0091] When the important evolutionary features are ranked and sorted according to the quantitative weight system, the weights are arranged in descending order from large to small. In a case involving 43 important evolutionary features, the top 30% of the feature set with the highest weights are selected as the dominant feature group, resulting in 13 dominant features. These dominant features include the change rate of the network's average path length, the degree distribution concentration of the core nodes, and the variance change of the edge weight. The system further analyzes the development trend of these dominant features in the dynamic evolution mode and identifies the trend type by fitting the change curve of the feature value over time. For features showing a continuous upward trend, the system determines that the network is developing in the direction represented by the feature. For features showing periodic fluctuations, the system extracts the fluctuation period and amplitude parameters as the description index of the evolution law. In the case, the average path length change rate shows a continuous downward trend with a decrease rate of 0.03 per time unit, indicating that the network structure is tending towards a more tightly connected state.

[0092] When the development trend of the dominant feature group is established in correspondence with the historical optimization experience in the evolution knowledge base, the historical optimization cases recorded in the knowledge base are retrieved. The knowledge base stores more than 300 historical network optimization instances, each instance records the change trajectory of the dominant feature before and after optimization and the final performance improvement effect achieved. The system identifies the matching optimization experience by calculating the similarity between the current dominant feature development trend and the feature change trajectory in the historical cases. The similarity is quantified by comparing the shape characteristics of the trend curve and the position deviation of the key turning points, and when the similarity exceeds 0.8, it is considered as a high match.

[0093] In actual cases, the system identifies 15 highly matched historical optimization cases, which show that when the average path length continues to decline and the core node degree distribution concentration improves synchronously, the network transmission efficiency can be improved by 40% by increasing the cross-community connection edge. Based on these matched optimization experiences, the system establishes a mapping relationship between feature importance and network evolution direction, takes the weight value of the dominant feature as the quantitative indicator of the mapping strength, takes the operation type and parameter adjustment range in the optimization experience as the target output of the mapping, and forms a complete mapping function from the feature space to the evolution direction space.

[0094] The second aspect of the embodiment of the application provides a deep reinforcement learning-based adaptive network topology dynamic reconstruction system, which comprises: A first unit is configured to acquire topology state information, service traffic distribution information and historical reconstruction records of a current network, perform multi-level feature extraction on the topology state information, capture topological correlation features between nodes through graph convolution operation, and generate a fusion state representation in combination with the service traffic distribution information; A second unit is configured to input the fusion state representation and the historical reconstruction records into a deep reinforcement learning model, identify bottleneck nodes and redundant links in a topology structure, and output a reconstruction action candidate set; and the second unit is further configured to evaluate long-term cumulative returns of each candidate action through Monte Carlo tree search, and filter out an optimal reconstruction action sequence; A third unit is configured to perform staged verification on the optimal reconstruction action sequence, assign time slots to each reconstruction action through a graph coloring algorithm, identify and process action combinations with resource conflicts, and generate a resource feasible topology adjustment scheme; A fourth unit is configured to convert the resource feasible topology adjustment scheme into a multi-dimensional quantitative feature sequence, extract network evolution rules by using tensor decomposition technology, and construct a correlation mapping atlas of topology optimization by using an entropy weight ordering method, so as to realize intelligent evolution of a network architecture.

[0095] The third aspect of the embodiment of the application provides an electronic device, which comprises: a processor; a memory for storing processor-executable instructions; The processor is configured to invoke instructions stored in the memory to perform the method described above.

[0096] In a fourth aspect, the present application provides a computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method described above.

[0097] The present application can be a method, apparatus, system, and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which, when executed by a processor, perform various aspects of the present application.

[0098] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An adaptive network topology dynamic reconfiguration method based on deep reinforcement learning, characterized in that, The method comprises the following steps: acquiring topology state information, traffic distribution information and historical reconstruction records of a current network; performing multi-level feature extraction on the topology state information, capturing topology correlation features between nodes through graph convolution operation, and generating fusion state representation in combination with the traffic distribution information; inputting the fusion state representation and the historical reconstruction records into a deep reinforcement learning model to identify bottleneck nodes and redundant links in the topology structure, and outputting a reconstruction action candidate set; evaluating the long-term cumulative returns of each candidate action through Monte Carlo tree search to filter out an optimal reconstruction action sequence; performing phased verification on the optimal reconstruction action sequence, assigning time slots to each reconstruction action through graph coloring algorithm, identifying and processing action combinations with resource conflicts, and generating a resource feasible topology adjustment scheme; transforming the resource feasible topology adjustment scheme into a multi-dimensional quantized feature sequence, extracting network evolution rules by using tensor decomposition technology, and constructing a correlation mapping atlas of topology optimization by using entropy weight ordering method to realize intelligent evolution of network architecture.

2. The method of claim 1, wherein, The multi-level feature extraction on the topology state information, the capturing of topology correlation features between nodes through graph convolution operation, and the generation of fusion state representation in combination with the traffic distribution information comprise: constructing the topology state information into a hierarchical graph structure representation, sampling multi-hop neighborhood features based on node degree distribution characteristics, and constructing a local topology subgraph for each node's multi-hop neighborhood features; using graph convolution operation based on the local topology subgraph to depict the correlation strength distribution of nodes at different propagation levels, combining the distance decay law and information transmission capacity between nodes, and generating multi-level topology features representing network structure evolution rules; based on the node interaction mode implied in the multi-level topology features, establishing a bidirectional mapping relationship atlas of node identifiers and traffic source and sink identifiers, converting traffic demand and transmission delay requirements into attribute descriptions of node feature vectors, and constructing fusion state representation with spatiotemporal dependence characteristics through hierarchical combination.

3. The method of claim 2, wherein, Based on the node interaction mode implied in the multi-level topology features, establishing a bidirectional mapping relationship atlas of node identifiers and traffic source and sink identifiers, converting traffic demand and transmission delay requirements into attribute descriptions of node feature vectors, comprises: constructing a node interaction relationship atlas from the multi-level topology features, depicting the dynamic coupling strength of node pairs based on the message passing frequency and feature similarity between nodes, and dividing node identifiers into multiple hierarchical communities with close interaction relationships according to the distribution characteristics of the dynamic coupling strength, and constructing a hierarchical mapping system for each community; constructing a forward feature diffusion path based on the hierarchical mapping system through the cooperative transmission effect of nodes within the community, and generating a multi-dimensional correlation atlas of traffic source and sink identifiers using the propagation law of the forward feature diffusion path, improving the hierarchical mapping system according to the multi-dimensional correlation atlas, and forming a bidirectional mapping knowledge network; According to the structural characteristics of the bidirectional mapping knowledge network, the business traffic demand is converted into a dynamic load distribution between nodes based on the forward feature diffusion path, the transmission time delay requirement is mapped to the time sequence response characteristics of the nodes through the multi-dimensional correlation graph, and the dynamic load distribution and the time sequence response characteristics are integrated into the multi-dimensional attribute expression of the nodes by using a feature space reconstruction technology, so as to form an attribute description with rich semantic information.

4. The method of claim 1, wherein, The fusion state representation and the historical reconstruction record are input into a deep reinforcement learning model to identify bottleneck nodes and redundant links in the topology structure, and a reconstruction action candidate set is output; The long-term cumulative benefits of each candidate action are evaluated by Monte Carlo tree search to filter out an optimal reconstruction action sequence including: The fusion state representation is input into a deep reinforcement learning model, the importance distribution of each node feature vector is calculated based on an attention mechanism, resource utilization analysis is performed on nodes with an importance distribution exceeding a dynamic threshold, nodes meeting congestion determination conditions are marked as a bottleneck node group, and a redundant link group with low link utilization is identified; Based on the congestion state of the bottleneck node group and the utilization rate characteristics of the redundant link group, a candidate action set of link addition and link deletion is generated, and an experienced action with a similarity higher than a matching threshold to the current state is extracted from the historical reconstruction record, the candidate action set and the experienced action are integrated to form a reconstruction action sequence; For the reconstruction action sequence, a Monte Carlo search tree is constructed with the current fusion state representation as a root node, state transition prediction is performed on each candidate action in the sequence, the evolution trend of the successor state is evaluated by the deep reinforcement learning model, each evolution branch is expanded for multiple steps, and a long-term benefit index of the action sequence is calculated; The candidate action corresponding to the optimal value is selected from the long-term benefit index as a preferred solution, the preferred solution and candidate actions with suboptimal benefit indexes are verified, an action combination maximizing the overall network performance is filtered out, and an optimal reconstruction sequence with time sequence characteristics is generated.

5. The method of claim 1, wherein, The optimal reconstruction action sequence is verified in stages, a time slot is allocated to each reconstruction action by a graph coloring algorithm, action combinations with resource conflicts are identified and processed, and a resource feasible topology adjustment scheme is generated including: The optimal reconstruction action sequence is mapped to a node group of a graph coloring model, a conflict constraint network is constructed by analyzing the resource competition and dependence characteristics between reconstruction actions, the competition strength of node pairs with conflict constraint connections is marked, and the minimum number of colors of the graph coloring model is determined based on the topology structure of the conflict constraint network; Based on the structural characteristics of the conflict constraint network in the graph coloring model, corresponding color batch identifiers are allocated to the node group, so that node pairs with a competitive relationship are divided into different execution batches, reconstruction action combinations corresponding to the same batch identifier are combined into parallel execution units, and the execution time window is allocated to the parallel execution units in priority order; A matching relationship between resource requirement profiles and network capacity distribution is constructed for parallel execution units within the execution time window, conflict time windows in which resource requirements exceed capacity constraints are identified, execution units in the conflict time windows are dynamically reorganized, part of the reconstruction actions are transferred to subsequent idle time windows, and a resource feasible topology adjustment scheme is formed.

6. The method of claim 1, wherein, The resource feasible topology adjustment scheme is converted into a multi-dimensional quantified feature sequence, network evolution rules are extracted using tensor decomposition technology, and a correlation mapping atlas for topology optimization is constructed using an entropy weight ordering method, including: Multi-dimensional features of network evolution are extracted from the resource feasible topology adjustment scheme, and the multi-dimensional features are reorganized into a multi-dimensional quantified feature sequence according to the evolution time sequence of the topology structure. The multi-dimensional quantified feature sequence is constructed into a three-dimensional state expression space, the internal structure of the multi-dimensional quantified feature sequence is revealed through tensor decomposition technology, the three-dimensional state expression space is decoupled into a combination of core features and pattern matrices, and network evolution paths are identified from the pattern matrix in the time dimension. Based on the coupling relationship between the core features and the pattern matrices, the coordinated evolution rules of node connectivity and link load distribution are extracted, the coordinated evolution rules are refined into the essential evolution mode of network structure, and an evolution knowledge base reflecting the dynamic characteristics of the network is constructed. According to the coordinated mode in the evolution knowledge base, information entropy analysis is performed on the multi-dimensional quantified feature sequence, the information entropy is converted into a quantified index of feature importance, a mapping relationship between the feature importance and the network evolution direction is established, and a correlation mapping atlas with predictive value is formed.

7. The method of claim 6, wherein, According to the coordinated mode in the evolution knowledge base, information entropy analysis is performed on the multi-dimensional quantified feature sequence, the information entropy is converted into a quantified index of feature importance, a mapping relationship between the feature importance and the network evolution direction is established, and a correlation mapping atlas with predictive value is formed. The dynamic evolution mode of the topology structure is extracted from the evolution knowledge base, the evolution activity level of the multi-dimensional quantified feature sequence is evaluated based on the coupling rules in the dynamic evolution mode, the feature group with an evolution activity level exceeding a preset activity threshold is marked as an important evolution feature, and the time sequence distribution of the important evolution feature is preliminarily evaluated. An information entropy evaluation system is constructed for the time sequence distribution characteristics of the important evolution feature to quantify the uncertainty of its dynamic change, the information entropy index and the evolution activity level are multi-dimensionally fused to form a quantified weight system reflecting the importance of the feature, The important evolution features are hierarchically ordered according to the quantified weight system, a feature set with a high ranking is extracted as a dominant feature group of network evolution, the development trend of the dominant feature group in the dynamic evolution mode is analyzed to reveal the evolution rules of the network structure, The development trend of the dominant feature group and the historical optimization experience in the evolution knowledge base are correlated to identify the path of network performance improvement, and a mapping relationship between feature importance and network evolution direction is constructed.

8. An adaptive network topology dynamic reconfiguration system based on deep reinforcement learning for implementing the method of any of the preceding claims 1-7, characterized in that, including: The first unit is configured to acquire topology state information, traffic distribution information and historical reconstruction records of a current network, perform multi-level feature extraction on the topology state information, capture topology correlation features between nodes through graph convolution operation, and generate fusion state representation in combination with the traffic distribution information; The second unit is configured to input the fusion state representation and the historical reconstruction records into a deep reinforcement learning model, identify bottleneck nodes and redundant links in a topology structure, and output a reconstruction action candidate set; and evaluate long-term cumulative returns of each candidate action through Monte Carlo tree search to filter out an optimal reconstruction action sequence. The third unit is configured to perform staged verification on the optimal reconstruction action sequence, assign time slots to each reconstruction action through a graph coloring algorithm, identify and process action combinations with resource conflicts, and generate a resource feasible topology adjustment scheme. The fourth unit is configured to convert the resource feasible topology adjustment scheme into a multi-dimensional quantified feature sequence, extract network evolution rules by using tensor decomposition technology, construct a correlation mapping atlas of topology optimization by using an entropy weight ordering method, and realize intelligent evolution of a network architecture.

9. An electronic device, comprising: comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7.

Images