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

Abstract

The application provides a deep reinforcement learning-based adaptive network topology dynamic reconstruction method and system, relates to the technical field of deep reinforcement learning, and comprises the following steps: acquiring the topology state information, service traffic distribution information and historical reconstruction records of a current network; extracting the topological correlation characteristics between nodes through graph convolution operation, combining the service traffic information to generate a fusion state representation; 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; evaluating the long-term cumulative benefits of the candidate actions through Monte Carlo tree search to select an optimal reconstruction action sequence; allocating time slots and processing resource conflicts by using a graph coloring algorithm to generate a resource feasible topology adjustment scheme; and extracting network evolution rules by using tensor decomposition to construct a topology optimization correlation mapping graph. The application can intelligently identify network bottlenecks, dynamically optimize the network topology structure, and effectively improve the network performance and resource utilization.

Description

Technical Field

[0001] This invention relates to deep reinforcement learning technology, and more particularly to an adaptive network topology dynamic reconstruction method and system based on deep reinforcement learning. Background Technology

[0002] With the rapid development of emerging technologies such as 5G, IoT, and cloud computing, network traffic is characterized by high dynamism, high burstiness, and uneven distribution. Traditional static network architectures are no longer sufficient to meet increasingly complex business demands. Network topology reconfiguration, as an effective network optimization method, can dynamically adjust the network structure based on real-time traffic changes and business needs, improving network resource utilization, reducing network latency, and enhancing network performance. In recent years, the rise of software-defined networking (SDN) technology has provided a technological foundation for flexible network topology reconfiguration, enabling the separation of the network control plane and data plane, and realizing programmable network configuration. Simultaneously, the development of artificial intelligence technologies, especially deep learning and reinforcement learning, has provided new ideas and methods for solving complex decision-making problems such as network topology optimization.

[0003] Existing methods mainly rely on traditional graph theory analysis or simple statistical features when extracting network topology features. They lack the ability to explore the deep relationships between network nodes, making it difficult to accurately identify bottleneck nodes and redundant links in complex network environments. This results in insufficient accuracy in reconstruction decisions and an inability to effectively address the topology optimization needs in large-scale dynamic network scenarios.

[0004] Existing technologies often only consider immediate benefits or short-term optimization goals when formulating topology reconfiguration strategies, lacking an evaluation mechanism for the long-term impact of reconfiguration actions. This can easily lead to local optima, resulting in frequent network adjustments that negatively affect network stability. Furthermore, they cannot effectively balance the relationship between network performance optimization and reconfiguration costs, thus impacting overall network operational efficiency.

[0005] Existing methods lack systematic handling of network resource constraints and action conflicts when performing topology reconfiguration. Multiple reconfiguration actions may compete for the same network resources simultaneously, leading to resource conflicts or deadlocks in the actual execution of the reconfiguration scheme. Furthermore, they lack in-depth analysis and utilization of network evolution patterns, making it impossible to achieve continuous intelligent optimization and adaptive evolution of the network architecture. Summary of the Invention

[0006] This invention provides an adaptive network topology dynamic reconstruction method and system based on deep reinforcement learning, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides an adaptive network topology dynamic reconstruction method based on deep reinforcement learning, comprising:

[0008] Obtain the current network topology status information, service traffic distribution information, and historical reconstruction records; perform multi-level feature extraction on the topology status information, capture the topology association features between nodes through graph convolution operation, and generate a fused status representation by combining the service traffic distribution information;

[0009] 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 and output a candidate set of reconstruction actions. The long-term cumulative benefit of each candidate action is evaluated by Monte Carlo tree search to select the optimal reconstruction action sequence.

[0010] The optimal reconstruction action sequence is verified in stages. A time slot is allocated to each reconstruction action using a graph coloring algorithm. Action combinations with resource conflicts are identified and processed to generate resource-feasible topology adjustment schemes.

[0011] The feasible topology adjustment schemes for the resources are transformed into multi-dimensional quantitative feature sequences. Tensor decomposition technology is used to extract network evolution laws, and an association mapping graph for topology optimization is constructed using entropy weight ranking method to achieve intelligent evolution of network architecture.

[0012] Multi-level feature extraction is performed on the topology state information, and topological association features between nodes are captured through graph convolution operations. Combined with the service traffic distribution information, a fused state representation is generated, including:

[0013] The topological state information is constructed into a hierarchical graph structure representation. Multi-hop neighborhood features are sampled based on the node degree distribution characteristics. A local topological subgraph is constructed for the multi-hop neighborhood features of each node.

[0014] Based on the local topological subgraph, graph convolution operation is used to characterize the distribution of the association strength of nodes at different propagation levels. Combined with the distance decay law and information transmission capability between nodes, multi-level topological features that characterize the evolution law of network structure are generated.

[0015] Based on the node interaction patterns contained in the multi-layer topological features, a bidirectional mapping relationship graph between node identifiers and traffic source and destination identifiers is established. The service traffic demand and transmission latency requirements are transformed into attribute descriptions of node feature vectors. A fusion state representation with spatiotemporal dependence is constructed through hierarchical combination.

[0016] Based on the node interaction patterns inherent in the multi-level topological features, a bidirectional mapping relationship graph between node identifiers and traffic source / destination identifiers is established. This graph transforms service traffic demand and transmission latency requirements into attribute descriptions of node feature vectors, including:

[0017] A node interaction relationship spectrum is constructed from the multi-level topological features. The dynamic coupling strength of node pairs is characterized based on the message transmission frequency and feature similarity between nodes. The node identifiers are divided into multiple hierarchical communities with close interaction relationships according to the distribution characteristics of the dynamic coupling strength. A hierarchical mapping system is constructed for each community.

[0018] Based on the hierarchical mapping system, a positive feature diffusion path is constructed through the collaborative transmission effect of nodes within the community. The propagation law of the positive feature diffusion path is used to generate a multi-dimensional association graph of traffic source and destination identifiers. The hierarchical mapping system is improved based on the multi-dimensional association graph to form a bidirectional mapping knowledge network.

[0019] Based on the structural characteristics of the bidirectional mapping knowledge network, the business traffic demand is transformed into a dynamic load distribution among nodes based on the positive feature diffusion path. The transmission delay requirement is mapped to the temporal response characteristics of the nodes through the multi-dimensional association graph. The feature space reconstruction technology is used to integrate the dynamic load distribution and the temporal response characteristics into the multi-dimensional attribute expression of the nodes, forming an attribute description with rich semantic information.

[0020] The fused 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 and output a candidate set of reconstruction actions. The long-term cumulative benefit of each candidate action is evaluated using Monte Carlo tree search to select the optimal reconstruction action sequence, which includes:

[0021] The fusion state representation is input into a deep reinforcement learning model, and the importance distribution of each node's feature vector is calculated based on the attention mechanism. Resource utilization analysis is performed on nodes whose importance distribution exceeds a dynamic threshold. Nodes that meet the congestion judgment conditions are marked as bottleneck node groups, and redundant link groups with low link utilization are identified.

[0022] Based on the congestion status of the bottleneck node group and the utilization characteristics of the redundant link group, a set of candidate actions for link addition and link deletion is generated. Experienced actions with a similarity higher than the matching threshold to the current state are extracted from historical reconstruction records. The set of candidate actions and the experienced actions are integrated to form a reconstruction action sequence.

[0023] 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 subsequent state is evaluated through the deep reinforcement learning model. Each evolution branch is expanded in multiple steps, and the long-term benefit index of the action sequence is calculated.

[0024] The candidate action corresponding to the optimal value from the long-term benefit index is selected as the preferred solution. The preferred solution is then combined with the candidate action with the second-best benefit index for verification. The action combination that maximizes the overall network performance is then selected, and the optimal reconstruction sequence with temporal characteristics is generated.

[0025] The optimal reconstruction action sequence is verified in stages. A time slot is allocated to each reconstruction action using a graph coloring algorithm. Action combinations with resource conflicts are identified and processed, and resource-feasible topology adjustment schemes are generated, including:

[0026] The optimal reconstruction action sequence is mapped to a group of nodes in a graph coloring model. A conflict constraint network is constructed by analyzing the resource competition and dependency characteristics between reconstruction actions. The competition intensity of node pairs with conflict constraint connections is marked, and the minimum number of colors in the graph coloring model is determined based on the topology of the conflict constraint network.

[0027] Based on the structural features of the conflict constraint network in the graph coloring model, corresponding color batch identifiers are assigned to the node group, so that nodes with competing relationships are divided into different execution batches. The reconstruction actions corresponding to nodes with the same batch identifier are combined into parallel execution units, and the execution time windows are allocated to the parallel execution units according to priority order.

[0028] For parallel execution units within the execution time window, a matching relationship between resource demand profiles and network capacity distribution is constructed. Conflicting time windows where resource demand exceeds capacity constraints are identified. Execution units within the conflicting time windows are dynamically reorganized, and some reorganization actions are transferred to subsequent idle time windows to form a resource-feasible topology adjustment scheme.

[0029] The feasible topology adjustment schemes for the aforementioned resources are transformed into multi-dimensional quantitative feature sequences. Tensor decomposition techniques are used to extract network evolution patterns, and an association mapping graph for topology optimization is constructed using the entropy weight ranking method.

[0030] Extract multi-dimensional features of network evolution from the feasible topology adjustment schemes of the resources, and reorganize the multi-dimensional features into a multi-dimensional quantified feature sequence according to the evolution time sequence of the topology structure;

[0031] The multi-dimensional quantized feature sequence is constructed into a three-dimensional state expression space. The internal structure of the multi-dimensional quantized feature sequence is revealed by tensor decomposition technology. The three-dimensional state expression space is decoupled into a combination of core features and pattern matrix. The network evolution path is identified from the pattern matrix in the time dimension.

[0032] Based on the coupling relationship between the core features and the pattern matrix, the co-evolution law of node connectivity and link load distribution is extracted, and the co-evolution law is refined into the essential evolution pattern of network structure to construct an evolutionary knowledge base that reflects the dynamic characteristics of the network.

[0033] Based on the collaborative patterns in the evolutionary knowledge base, information entropy analysis is performed on the multi-dimensional quantitative feature sequences, and the information entropy is transformed into a quantitative index of feature importance. A mapping relationship between the feature importance and the network evolution direction is established to form a correlation mapping map with predictive value.

[0034] Based on the collaborative patterns in the evolutionary knowledge base, information entropy analysis is performed on the multi-dimensional quantified feature sequences. This information entropy is then transformed into a quantitative index of feature importance. The mapping relationship between feature importance and network evolution direction is established, including:

[0035] The dynamic evolution pattern of the topology is extracted from the evolutionary knowledge base. The evolutionary activity of the multi-dimensional quantized feature sequence is evaluated based on the coupling rules in the dynamic evolution pattern. Feature groups whose evolutionary activity exceeds a preset activity threshold are marked as important evolutionary features. The temporal distribution of the important evolutionary features is preliminarily evaluated.

[0036] Based on the temporal distribution characteristics of the important evolutionary features, an information entropy evaluation system is constructed to quantify the uncertainty of their dynamic changes. The information entropy index is integrated with the degree of evolutionary activity in multiple dimensions to form a quantitative weighting system that reflects the importance of the features.

[0037] The important evolutionary features are ranked and sorted according to the quantized weight system. The top-ranked feature set is extracted as the dominant feature group of network evolution. By analyzing the development trend of the dominant feature group in the dynamic evolution mode, the evolution law of network structure is revealed.

[0038] By establishing a correspondence between the development trend of the dominant feature group and the historical optimization experience in the evolutionary knowledge base, the path to improve network performance is identified, and a mapping relationship between feature importance and network evolution direction is constructed.

[0039] A second aspect of the present invention provides an adaptive network topology dynamic reconstruction system based on deep reinforcement learning, comprising:

[0040] The first unit is used to acquire the current network topology status information, service traffic distribution information and historical reconstruction records; to perform multi-level feature extraction on the topology status information, to capture the topology association features between nodes through graph convolution operation, and to generate a fused status representation in combination with the service traffic distribution information.

[0041] The second unit is used to input the fusion state representation and the historical reconstruction record into a deep reinforcement learning model, identify bottleneck nodes and redundant links in the topology, and output a candidate set of reconstruction actions; evaluate the long-term cumulative benefit of each candidate action through Monte Carlo tree search, and select the optimal reconstruction action sequence.

[0042] The third unit is used to perform phased verification of the optimal reconstruction action sequence. It allocates time slots to each reconstruction action using a graph coloring algorithm, identifies and processes action combinations with resource conflicts, and generates resource-feasible topology adjustment schemes.

[0043] The fourth unit is used to transform the feasible topology adjustment schemes of the resources into multi-dimensional quantitative feature sequences, extract network evolution laws using tensor decomposition technology, and construct a correlation mapping map for topology optimization using entropy weight sorting method to realize intelligent evolution of network architecture.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0049] The beneficial effects of this application are as follows:

[0050] This invention extracts features from topology state information at multiple levels, uses graph convolution operations to deeply capture topological association features between nodes, and fuses these features with service traffic distribution information to generate a more comprehensive and accurate network state representation. This multi-source information fusion approach can more accurately reflect the real-time operating status of the network, providing a high-quality data foundation for subsequent topology reconfiguration decisions and significantly improving the accuracy and comprehensiveness of network state perception.

[0051] This invention employs a deep reinforcement learning model combined with Monte Carlo tree search, which not only identifies bottleneck nodes and redundant links in the topology but also evaluates the long-term cumulative benefits of each candidate action, thereby selecting the optimal reconstruction action sequence. This method overcomes the limitation of traditional topology optimization methods that only focus on short-term gains, enabling decision-making from a global and long-term perspective, ensuring the scientific nature and effectiveness of network topology reconstruction, and improving network resource utilization efficiency.

[0052] This invention employs a graph coloring algorithm for time slot allocation and resource conflict detection, utilizes tensor decomposition technology to extract network evolution patterns, and combines entropy weight ranking to construct an association mapping graph for topology optimization, achieving a complete closed loop from theoretical solution to practical execution. This phased verification and intelligent evolution mechanism effectively avoids resource conflicts and network interruptions during the reconstruction process, ensures a smooth transition in topology adjustments, and improves the reliability and intelligence level of network operation and maintenance. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the adaptive network topology dynamic reconstruction method based on deep reinforcement learning, as described in an embodiment of the present invention.

[0054] Figure 2 This is a flowchart illustrating the construction process of multi-level topological features and dynamic coupling mapping in an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0057] Figure 1 This is a flowchart illustrating the adaptive network topology dynamic reconstruction method based on deep reinforcement learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] Obtain the current network topology status information, service traffic distribution information, and historical reconstruction records; perform multi-level feature extraction on the topology status information, capture the topology association features between nodes through graph convolution operation, and generate a fused status representation by combining the service traffic distribution information;

[0059] 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 and output a candidate set of reconstruction actions. The long-term cumulative benefit of each candidate action is evaluated by Monte Carlo tree search to select the optimal reconstruction action sequence.

[0060] The optimal reconstruction action sequence is verified in stages. A time slot is allocated to each reconstruction action using a graph coloring algorithm. Action combinations with resource conflicts are identified and processed to generate resource-feasible topology adjustment schemes.

[0061] The feasible topology adjustment schemes for the resources are transformed into multi-dimensional quantitative feature sequences. Tensor decomposition technology is used to extract network evolution laws, and an association mapping graph for topology optimization is constructed using entropy weight ranking method to achieve intelligent evolution of network architecture.

[0062] In one optional implementation, multi-level feature extraction is performed on the topology state information, capturing the topological association features between nodes through graph convolution operations, and generating a fused state representation by combining the service traffic distribution information, including:

[0063] The topological state information is constructed into a hierarchical graph structure representation. Multi-hop neighborhood features are sampled based on the node degree distribution characteristics. A local topological subgraph is constructed for the multi-hop neighborhood features of each node.

[0064] Based on the local topological subgraph, graph convolution operation is used to characterize the distribution of the association strength of nodes at different propagation levels. Combined with the distance decay law and information transmission capability between nodes, multi-level topological features that characterize the evolution law of network structure are generated.

[0065] Based on the node interaction patterns contained in the multi-layer topological features, a bidirectional mapping relationship graph between node identifiers and traffic source and destination identifiers is established. The service traffic demand and transmission latency requirements are transformed into attribute descriptions of node feature vectors. A fusion state representation with spatiotemporal dependence is constructed through hierarchical combination.

[0066] Topology state information needs to be transformed into a hierarchical graph structure representation, with an adjacency matrix as the basic data structure. Specifically, an adjacency matrix is ​​constructed for all nodes in the network according to their connections in the physical topology. The rows and columns of the matrix correspond to individual network nodes. When two nodes have a physical link connection, the element at the corresponding position in the matrix is ​​set to 1; otherwise, it is set to 0. After constructing the basic adjacency matrix, the degree value of each node needs to be calculated, i.e., the number of neighboring nodes directly connected to that node. By statistically analyzing the degree distribution of all nodes, core nodes with high degree values ​​and edge nodes with low degree values ​​are identified.

[0067] When performing multi-hop neighborhood feature sampling for each target node, the sampling expands outward from that node. The first-hop neighborhood includes all nodes directly connected to the target node, the second-hop neighborhood includes nodes directly connected to nodes in the first-hop neighborhood but not within the first-hop neighborhood, and so on, to obtain the third-hop and more distant neighboring nodes. In actual sampling, for nodes with high degree values, the number of neighboring nodes is too large, requiring a random sampling strategy for control. For example, if the first-hop neighborhood of a core exchange node contains 80 directly connected nodes, to avoid excessively high feature dimensionality, 30 nodes are randomly selected as representative neighbors. During sampling, it is necessary to maintain a relatively balanced proportion of nodes in neighborhoods with different hop numbers. If 30 nodes are sampled in the first hop, it is recommended to sample 25 nodes in the second hop and 20 nodes in the third hop to ensure a reasonable hierarchical distribution of feature sampling.

[0068] When constructing a local topology subgraph, the target node and its multi-hop neighbor nodes are used as the node set of the subgraph, while the original connections between these nodes are preserved as the edge set of the subgraph. For example, assuming a local subgraph is constructed for a certain access layer switch node, where the first-hop neighbor contains 8 access devices, the second-hop neighbor contains 3 aggregation layer switches, and the third-hop neighbor contains 1 core router, then this local subgraph contains 13 nodes. The number of edges is determined by the actual physical links between these nodes. Each local topology subgraph is stored independently as a graph data structure, containing a node list, an edge list, and node attribute information.

[0069] Graph convolution operations perform feature propagation and aggregation on local topological subgraphs. In the first layer of graph convolution, each node collects feature information from its direct neighbors and sums these neighbor features using a weighted average. The weights are determined by the connection strength of the edges. If the link bandwidth between two nodes is 10Gbps, the weight of the edge can be set to 10; if the link bandwidth is 1Gbps, the weight is set to 1. For a given aggregation layer node that directly connects to 5 access layer nodes and 2 core layer nodes, in the first layer of graph convolution, this node collects the feature vectors of these 7 neighbor nodes and averages them according to their respective link weights to obtain the aggregated features of this node at the first propagation layer. The second layer of graph convolution continues information propagation based on the output features of the first layer. At this point, the node can perceive topological information within a two-hop range. The third layer of graph convolution extends to a three-hop range.

[0070] Characterizing the association strength distribution requires considering the distance decay law between nodes. Directly connected nodes have the highest association strength, set as a baseline value of 1.0. The association strength decays to 0.6 for nodes with a two-hop distance, to 0.3 for nodes with a three-hop distance, and to below 0.1 for nodes with a four-hop or longer distance. If the path length from an edge access node to a core routing node is 3 hops, the association strength between this edge node and the core node is 0.3, meaning that changes in the state of the edge node have a relatively weak impact on the core node. Information transmission capacity is quantified by the forwarding capacity of nodes. The core router has a forwarding capacity of 500Gbps, with an information transmission capacity coefficient set to 5.0; the aggregation switch has a forwarding capacity of 100Gbps, with an information transmission capacity coefficient of 2.0; and the access switch has a forwarding capacity of 20Gbps, with an information transmission capacity coefficient of 0.5.

[0071] The multi-level topological feature is constructed by concatenating the output features of three layers of graph convolution. The first layer outputs a 64-dimensional vector, the second layer outputs a 48-dimensional vector, and the third layer outputs a 32-dimensional vector. These three feature vectors are concatenated in dimensional order to form a 144-dimensional multi-level topological feature vector. This feature vector encodes the topological structure information of nodes within different receptive fields, reflecting both the local characteristics and global evolutionary patterns of the network structure.

[0072] When establishing a bidirectional mapping relationship between node identifiers and traffic source / destination identifiers, a source / destination mapping table needs to be constructed. This table records the correspondence between the source node identifier and destination node identifier of a service traffic and the actual network nodes. For example, if the source identifier of a service traffic is server group A and the destination identifier is server group B, querying the mapping table determines 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, thus establishing a mapping relationship from service traffic identifiers to topology node identifiers. The reverse mapping relationship records the source / destination role of each network node in carrying service traffic. Node 15 serves as the source node for 12 service flows and the destination node for 8 service flows.

[0073] When traffic demand is converted into a node feature vector, the traffic demand of all traffic sources or destination nodes of that node is summed up. For example, if a node's total traffic demand as a source node is 150Gbps and its total traffic demand as a destination node is 200Gbps, these two values ​​are used as the two dimensions of the node's feature vector. Similarly, transmission latency requirements are converted into feature vector attributes. The average latency requirement of the relevant service flows of that node is calculated. If 60% of the service flows involving that node require a latency of less than 10 milliseconds, 30% require a latency of less than 50 milliseconds, and 10% require a latency of less than 100 milliseconds, then the latency sensitivity feature value of that node is set to 0.75, indicating that the node has high latency sensitivity.

[0074] When constructing the fusion state representation through layered combination, the topology feature vector and the traffic feature vector are concatenated dimensionally. The topology feature is a 144-dimensional vector, the traffic demand feature is a 2-dimensional vector, and the latency feature is a 1-dimensional vector, resulting in a 147-dimensional basic fusion feature. Spatiotemporal dependency is achieved by introducing historical features from a time window. Based on the current moment's features, the feature vectors from the previous moment and the two moments before that are superimposed. The feature vectors from these three moments are weighted and combined to form a spatiotemporal fusion representation. The weight of the current moment is set to 0.6, the weight of the previous moment is 0.3, and the weight of the two moments before that is 0.1. The sum of the weighted feature vectors from the three moments yields the final fusion state representation vector, which includes topology information, business traffic distribution characteristics, and temporal evolution patterns.

[0075] In one optional implementation, based on the node interaction patterns implied in the multi-level topology features, a bidirectional mapping relationship graph between node identifiers and traffic source / destination identifiers is established, transforming service traffic demand and transmission latency requirements into attribute descriptions of node feature vectors, including:

[0076] A node interaction relationship spectrum is constructed from the multi-level topological features. The dynamic coupling strength of node pairs is characterized based on the message transmission frequency and feature similarity between nodes. The node identifiers are divided into multiple hierarchical communities with close interaction relationships according to the distribution characteristics of the dynamic coupling strength. A hierarchical mapping system is constructed for each community.

[0077] Based on the hierarchical mapping system, a positive feature diffusion path is constructed through the collaborative transmission effect of nodes within the community. The propagation law of the positive feature diffusion path is used to generate a multi-dimensional association graph of traffic source and destination identifiers. The hierarchical mapping system is improved based on the multi-dimensional association graph to form a bidirectional mapping knowledge network.

[0078] Based on the structural characteristics of the bidirectional mapping knowledge network, the business traffic demand is transformed into a dynamic load distribution among nodes based on the positive feature diffusion path. The transmission delay requirement is mapped to the temporal response characteristics of the nodes through the multi-dimensional association graph. The feature space reconstruction technology is used to integrate the dynamic load distribution and the temporal response characteristics into the multi-dimensional attribute expression of the nodes, forming an attribute description with rich semantic information.

[0079] like Figure 2 As shown, the method includes:

[0080] In constructing the node interaction relationship spectrum, connection status information and communication record data between nodes are extracted from multi-level topological features. For any two nodes in the network, the system counts the number of message transmissions within a preset time window, which is set to 600 seconds. In a network instance containing 150 nodes, node A and node B transmitted messages 42 times within the statistical period, and node A and node C transmitted messages 8 times. The system also calculates the similarity of node feature vectors, which contain values ​​of 12 dimensions such as node processing power, storage capacity, and link bandwidth. The feature vector of node A contains parameters such as 8-core processor, 16GB memory, and 100Mbps bandwidth, while the feature vector of node B contains parameters such as 8-core processor, 32GB memory, and 100Mbps bandwidth. By comparing the differences in the values ​​of each dimension in the two feature vectors and calculating the comprehensive deviation, the feature similarity between node A and node B is 0.87, and the feature similarity between node A and node C is 0.34.

[0081] The calculation of dynamic coupling strength comprehensively considers two factors: message passing frequency and feature similarity. The coupling strength between node A and node B is obtained by normalizing the message passing frequency of 42 to the interval of 0 to 1, resulting in 0.91. This is then weighted and averaged with the feature similarity of 0.87, with weighting coefficients of 0.6 and 0.4, respectively, ultimately yielding a coupling strength value of 0.894. After calculating the coupling strength of all node pairs in the network, the system forms a 150x150 coupling strength matrix. By analyzing the numerical distribution of this matrix, the system identifies a total of 238 node pairs with a coupling strength greater than 0.75. Cluster analysis is then performed on these tightly coupled node pairs, dividing the 150 nodes into 5 hierarchical communities, with each community containing 38, 45, 32, 21, and 14 nodes, respectively.

[0082] When constructing a hierarchical mapping system for each community, a three-layer mapping structure was established for the first community, which contains 38 nodes. The top layer is the community representative node layer, selecting the three nodes with the highest average coupling strength with other nodes within the community as representatives. The average coupling strengths of these three nodes with nodes outside the community are 0.82, 0.79, and 0.81, respectively. The middle layer is the functional grouping layer, dividing the 38 nodes into data processing, data storage, and data forwarding groups based on their business type attributes. Each group contains 15, 12, and 11 nodes, respectively. The bottom layer is the specific node layer, recording detailed identification information for each node, including IP address, MAC address, and device serial number. A cross-layer index table is established in the mapping system to record the mapping relationship from the top-level representative nodes to the middle-layer groups and the mapping relationship from the middle-layer groups to the bottom-level specific nodes.

[0083] When constructing a forward feature diffusion path, taking the first community as an example, a source node S is selected within the community. This node needs to transmit feature information to the target node T. The system determines the diffusion path based on the dynamic coupling strength matrix. The path selection rule is to select the node with the highest coupling strength to the current node and which has not yet been visited for each hop. The feature vector of the source node S contains initial values ​​in 12 dimensions. When transmitting to the first-hop node N1, the values ​​of each component of the feature vector remain unchanged, but an additional path weight identifier of 0.89 is added. After receiving the feature vector, node N1 fuses its own feature vector with the received feature vector, mixing the corresponding dimension values ​​at a ratio of 0.5. The fused feature vector continues to be transmitted to the second-hop node N2, with the weight identifier updated to 0.82. Through multi-hop transmission, the feature information travels from the source node S through 5 intermediate nodes to the target node T, forming a complete diffusion path.

[0084] In generating a multi-dimensional association graph of traffic source and destination identifiers, 1000 historical communication records were collected from the network. Each record contains information such as source node identifier, destination node identifier, traffic volume, and transmission latency. The system analyzed all records of source node M as a traffic source and found that it had sent data to 32 different destination nodes. These 32 destination nodes were sorted by traffic volume, and the top 10 destination nodes received 76% of the total traffic. The system constructed an association graph for source node M that included dimensions of geographical location, service type, traffic volume, and latency sensitivity. The geographical location dimension indicates the network distance hops between the source node and each destination node; the service type dimension indicates the types of application protocols involved in the communication, totaling 8; the traffic volume dimension records the cumulative traffic received by each destination node, ranging from 2GB to 450GB; and the latency sensitivity dimension indicates that the maximum latency required by the destination node is distributed from 50 milliseconds to 500 milliseconds.

[0085] When perfecting the hierarchical mapping system to form a bidirectional mapping knowledge network, information from the multi-dimensional association graph is injected back into the hierarchical mapping system. In the mapping system of the first community, which originally only recorded static attribute information of nodes, dynamic traffic association information is now added. For example, node R, which originally only had identification information at the top level, now has a list of commonly used destination nodes associated with it when acting as a traffic source, containing 18 node identifiers. The data forwarding group in the middle layer adds a description of the traffic forwarding mode, indicating that the nodes in this group mainly handle cross-community traffic forwarding, with an average daily forwarding traffic of 2.3TB. The bottom-level nodes add bidirectional traffic statistics information; node X acts as both a traffic source sending an average of 340GB of traffic per day and a traffic sink receiving an average of 280GB of traffic per day.

[0086] When converting business traffic demands into dynamic load distribution among nodes, consider a typical business scenario: the source node needs to transmit a total of 800GB of data traffic to the destination node. Based on six feasible paths identified through forward feature diffusion paths, the system evaluates the carrying capacity of each path. The first path contains four intermediate nodes with current load rates of 45%, 52%, 38%, and 61%, and remaining bandwidth capacities of 55Mbps, 48Mbps, 62Mbps, and 39Mbps, respectively. The system calculates the bottleneck bandwidth of this path to be 39Mbps, and theoretically, transmitting 800GB of data would take approximately 47 hours. By analyzing the carrying capacity of all paths, the system decomposes the 800GB traffic into four parts: 240GB, 310GB, 180GB, and 70GB, which are then transmitted on four different paths. The load distribution of each node on the path is dynamically adjusted; the node load rates on the first path are updated to 58%, 66%, 51%, and 74%, forming a new load distribution state.

[0087] When the transmission latency requirement is mapped to the timing response characteristics of nodes through a multi-dimensional correlation graph, the service requirement is a processing latency of 100 milliseconds. Based on historical data of the latency sensitivity dimension in the correlation graph, the system identified a set of 67 nodes that meet this latency requirement. The timing response characteristics of these nodes are analyzed, including three components: processing latency, queuing latency, and transmission latency. The average processing latency of node P is 8 milliseconds, with a standard deviation of 3 milliseconds; the queuing latency is 2 milliseconds under light load and can reach 35 milliseconds under heavy load; the transmission latency depends on the link bandwidth and the data packet size, with a 1500-byte data packet having a transmission latency of 0.12 milliseconds on a 100Mbps link. The system maps the latency requirement to timing constraints that nodes must meet, namely, a comprehensive requirement that the processing latency does not exceed 10 milliseconds, the queuing latency does not exceed 30 milliseconds, and the transmission latency does not exceed 60 milliseconds.

[0088] When incorporating dynamic load distribution and timing response characteristics into the multi-dimensional attribute representation of nodes using feature space reconstruction technology, the original feature vector of node Q is expanded. The original feature vector contains 12 static attributes, and now 6 dynamic attributes are added. The newly added dynamic attributes include the current load rate of 68%, the remaining processing capacity of 3200 MIPS, the average queuing latency of 18 milliseconds, the traffic throughput rate of 78 Mbps, the number of connections of 145, and the cache utilization rate of 52%. The expanded 18-dimensional feature vector comprehensively describes the static capabilities and dynamic states of the node, providing rich semantic information support for subsequent network optimization decisions. The system completes feature vector expansion for all 150 nodes in the network, forming a unified attribute description specification. The attribute description of each node retains the original device configuration information while reflecting the service requirements characteristics of traffic transmission and latency constraints in real time.

[0089] In one optional implementation, the fused 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 and output a candidate set of reconstruction actions. The long-term cumulative benefit of each candidate action is evaluated using Monte Carlo tree search to select the optimal reconstruction action sequence, including:

[0090] The fusion state representation is input into a deep reinforcement learning model, and the importance distribution of each node's feature vector is calculated based on the attention mechanism. Resource utilization analysis is performed on nodes whose importance distribution exceeds a dynamic threshold. Nodes that meet the congestion judgment conditions are marked as bottleneck node groups, and redundant link groups with low link utilization are identified.

[0091] Based on the congestion status of the bottleneck node group and the utilization characteristics of the redundant link group, a set of candidate actions for link addition and link deletion is generated. Experienced actions with a similarity higher than the matching threshold to the current state are extracted from historical reconstruction records. The set of candidate actions and the experienced actions are integrated to form a reconstruction action sequence.

[0092] 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 subsequent state is evaluated through the deep reinforcement learning model. Each evolution branch is expanded in multiple steps, and the long-term benefit index of the action sequence is calculated.

[0093] The candidate action corresponding to the optimal value from the long-term benefit index is selected as the preferred solution. The preferred solution is then combined with the candidate action with the second-best benefit index for verification. The action combination that maximizes the overall network performance is then selected, and the optimal reconstruction sequence with temporal characteristics is generated.

[0094] The fused state representation and historical reconstruction records are input into a deep reinforcement learning model via a standardized interface. This model employs an Actor-Critic architecture, comprising two sub-modules: a policy network and a value network. The fused state representation is a 512-dimensional floating-point vector with values ​​ranging from 0 to 1. The historical reconstruction records are stored in time-series format, containing four fields: action type, execution time, influencing node, and effect evaluation. The model's input layer is set to 640 dimensions, with 512 dimensions receiving the fused state representation and 128 dimensions receiving the encoded vectors of historical records. Historical record encoding uses positional embedding, with time decay weights calculated exponentially using a decay factor of 0.92, and a maximum backtracking length of 50 historical actions.

[0095] The attention mechanism is embedded in the second layer of the model, employing a multi-head self-attention structure with 16 heads, each with a dimension of 32. During attention weight calculation, node feature vectors are transformed into query, key, and numerical matrices through three independent linear transformation layers. The weight matrix of each transformation layer has a dimension of 512×512. The scaled dot product attention calculation uses the softmax activation function, with a temperature parameter set to 0.1 to enhance the discriminative power of the weight distribution. The importance distribution output is a probability vector of length equal to the number of network nodes, with an element sum of 1 and individual elements ranging from 0 to 1.

[0096] The dynamic threshold employs an adaptive adjustment strategy, with the initial threshold set at the 75th quantile of the network node importance distribution, and an adjustment cycle of every 10 minutes. Threshold updates utilize an exponential moving average method with a smoothing coefficient of 0.8 and upper and lower bounds of 0.3 and 0.9, respectively. When the node importance distribution exceeds the dynamic threshold, the resource utilization analysis module is triggered. This module acquires four metrics in real-time from the monitoring system: CPU utilization, memory usage, network interface bandwidth utilization, and queue length. Resource utilization data is collected every 30 seconds, and the analysis window contains the most recent 300 sampling points, approximately 150 minutes of historical data.

[0097] Congestion determination criteria are based on a combination of multi-dimensional thresholds. Specifically, CPU utilization exceeds 80% for 5 consecutive minutes, memory usage exceeds 85%, or network interface bandwidth utilization exceeds 95% within any one minute. The congestion condition must persist for a cumulative period of at least 8 minutes. Bottleneck node groups are formed using a graph clustering algorithm, grouping nodes less than two hops apart that all meet the congestion criteria into the same group. The maximum number of nodes in a single group is limited to 8. Redundant link group identification is based on link utilization statistics, with a statistical window spanning the past 72 hours. Links are marked as redundant when their average utilization is below 15%, their 95th percentile utilization is below 35%, and their peak utilization is below 50%. Adjacent redundant links are merged into redundant link groups through connectivity analysis, with the number of links within each group limited to between 3 and 10.

[0098] The candidate action set generation module constructs link addition and deletion actions based on the characteristics of bottleneck node groups and redundant link groups, respectively. Link addition strategies include three types: proximity connection, central node connection, and load-distributed connection. Proximity connection establishes a link between the bottleneck node and its two-hop neighbors, selecting the lightest-loaded node. Central node connection selects the top 10% of nodes based on the betweenness centrality of the network topology as connection targets. Load-distributed connection connects the bottleneck node to nodes with a current load below 30%. The link deletion strategy prioritizes deleting redundant links with the least impact on network connectivity. The impact of deletion is evaluated by calculating edge connectivity and critical path coverage, with an edge connectivity threshold of 3 and a critical path coverage threshold of 5%.

[0099] The extraction of experiential actions from historical reconstruction records employs a state-space similarity matching algorithm. Similarity calculation is based on the cosine similarity of the fused state representations, with a computational complexity of O(n), where n is the dimension of the state vector. The matching process first arranges the historical records in reverse chronological order, then calculates the similarity between the current state and each historical state. When the similarity exceeds 0.75, the corresponding reconstruction action is extracted. The validity of experiential actions is verified based on post-execution performance evaluation; only actions with a performance score greater than 7.0 are included in the candidate set. The integration of the candidate action set and experiential actions uses a probability-weighted approach, with a selection probability of 0.6 for newly generated actions and 0.4 for experiential actions. The length of the integrated action sequence is controlled to be within 15 elements.

[0100] The Monte Carlo search tree is constructed with the current fusion state 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, with a maximum of 8 child nodes per layer, and the total number of nodes is controlled to within 5000. State transition prediction is achieved through forward inference of a deep reinforcement learning model. The input is a concatenated vector of the current state vector and action encoding. The action encoding uses a one-hot format with a dimension equal to the action space size. The batch size for model inference is set to 32, and the single inference time is controlled to within 50 milliseconds.

[0101] Evolutionary trend assessment employs a multi-step guided approach, with each evolutionary branch independently executing a 3-step state transition prediction, and intermediate states selecting actions through deterministic policies. The value function evaluation is based on a neural network model trained using a time-difference learning algorithm, with a network structure of 3 fully connected layers and 256, 128, and 64 hidden neurons, respectively. The uncertainty of action execution is considered during the multi-step unfolding process; the success probability of each action is set to 0.9, and in the event of failure, the state remains unchanged but a negative reward is generated.

[0102] The long-term benefit metric integrates three dimensions: network performance, implementation cost, and risk assessment. Network performance metrics include end-to-end latency, network throughput, load balancing, and reliability, with weights of 0.3, 0.3, 0.2, and 0.2, respectively. Latency is calculated using the shortest path algorithm, throughput is estimated using the maximum flow algorithm, load balancing is measured using the Gini coefficient, and reliability is assessed using network connectivity. Implementation cost includes equipment procurement cost, installation and deployment cost, and maintenance cost; 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, quantifying them using a combination of expert scoring and historical statistics. The benefit metric uses the net present value (NPV) method, with a discount rate of 8% and an assessment period of 24 months.

[0103] The optimal value is selected by ranking the benefit indicators. The preferred solution corresponds to the maximum benefit indicator, which must be greater than 20% of the current benchmark value. Candidate actions for the second-best benefit indicators are defined as the set of actions whose benefit indicators rank in the top 30% and are less than 15% different from the optimal value. Combinatorial verification uses a discrete event simulation method. The simulation environment is built based on the current network topology and traffic patterns, with a simulation time step of 1 minute and a total simulation duration of 7 days. During verification, dependencies and resource contention between actions are considered. The number of actions executed simultaneously is limited to 3, and the minimum time interval between actions is 15 minutes. Maximizing overall network performance is achieved through a multi-objective optimization algorithm, using a weighted aggregation method to merge multiple performance indicators into a single objective function.

[0104] Temporal feature modeling employs a Long Short-Term Memory (LSTM) network, consisting of two LSTM layers with a hidden state dimension of 128. The sequence length is equal to the number of actions. Temporal dependencies are represented by an attention weight matrix, the calculation of which considers the similarity of action types and the influence of time intervals. The optimal reconstructed sequence is generated using a greedy search strategy, starting with the action with the highest reward and sequentially selecting subsequent actions that are compatible with the selected actions and have high rewards. The sequence length is controlled between 3 and 8 actions, and the total execution time does not exceed 4 hours. Resource conflict detection is implemented using a graph coloring algorithm.

[0105] In one optional implementation, the optimal reconstruction action sequence is verified in stages. A graph coloring algorithm is used to allocate time slots to each reconstruction action, identify and process action combinations with resource conflicts, and generate resource-feasible topology adjustment schemes, including:

[0106] The optimal reconstruction action sequence is mapped to a group of nodes in a graph coloring model. A conflict constraint network is constructed by analyzing the resource competition and dependency characteristics between reconstruction actions. The competition intensity of node pairs with conflict constraint connections is marked, and the minimum number of colors in the graph coloring model is determined based on the topology of the conflict constraint network.

[0107] Based on the structural features of the conflict constraint network in the graph coloring model, corresponding color batch identifiers are assigned to the node group, so that nodes with competing relationships are divided into different execution batches. The reconstruction actions corresponding to nodes with the same batch identifier are combined into parallel execution units, and the execution time windows are allocated to the parallel execution units according to priority order.

[0108] For parallel execution units within the execution time window, a matching relationship between resource demand profiles and network capacity distribution is constructed. Conflicting time windows where resource demand exceeds capacity constraints are identified. Execution units within the conflicting time windows are dynamically reorganized, and some reorganization actions are transferred to subsequent idle time windows to form a resource-feasible topology adjustment scheme.

[0109] The optimal refactoring action sequence is mapped to a group of nodes in a graph coloring model through action identifiers. Each refactoring action corresponds to a node in the graph model, and the node attributes include four fields: action type, resource requirements, execution duration, and priority. Action types are divided into four categories: link addition, link deletion, node configuration modification, and route adjustment, and are stored using enumerated values. The resource requirements field records the computing resources, network bandwidth, and storage space required for action execution. Computing resources are represented by the number of CPU cores, network bandwidth by Mbps, and storage space by GB. The execution duration is estimated using historical statistical data, and the duration distribution is modeled using a log-normal distribution with a mean of 5 minutes and a standard deviation of 2 minutes.

[0110] The resource contention characteristics between reconstructed actions are determined through resource occupancy matrix analysis. The matrix dimension is the square of the number of actions, and the matrix elements represent the resource overlap between two actions. Resource overlap is calculated based on the intersection ratio of the involved devices; a competition relationship arises when two actions operate on the same or adjacent devices. Dependency analysis considers logical dependencies and temporal constraints between actions. Logical dependencies include precondition dependencies and configuration consistency dependencies, while temporal constraints require certain actions to be executed in a specific order. The conflict constraint network is stored using an undirected graph structure. Edges in the graph represent conflict constraints between two actions, and edge weights represent conflict intensity. Conflict intensity is divided into three levels: high, medium, and low, corresponding to weight values ​​of 3, 2, and 1, respectively.

[0111] Conflict-constrained node pair marking is achieved through adjacency matrix traversal, during which direct or indirect conflicts are detected between each pair of nodes. Direct conflict refers to two actions operating on the same resource, while indirect conflict refers to actions with a cascading effect. Competition intensity is assessed based on three dimensions: resource conflict degree, impact scope, and recovery difficulty, each using a scoring scale from 0 to 10. Resource conflict degree is calculated based on the proportion of shared resources, impact scope is measured by the number of affected nodes, and recovery difficulty is assessed based on configuration complexity and rollback time. The final competition intensity is the weighted average of the three dimensions, with weights set to 0.5, 0.3, and 0.2, respectively.

[0112] The minimum number of colors for the graph coloring model is determined by combining a greedy coloring algorithm and an exact coloring algorithm. The greedy algorithm uses a degree-based sorting strategy, prioritizing color allocation to nodes with the highest degree, and sorting by action priority when degrees are the same. During the algorithm iteration, a color usage state table is maintained, recording the set of nodes assigned to each color. The exact algorithm is implemented based on backtracking search, with the search depth limited to the total number of nodes. The backtracking pruning condition is that the current number of colors used exceeds a known upper bound. The theoretical lower bound for the minimum number of colors is estimated using the maximum clique algorithm, and the actual coloring process keeps the number of colors within 1.5 times the theoretical lower bound.

[0113] The color batch assignment for node groups employs a sequential coloring strategy, grouping nodes of the same color into the same execution batch. Load balancing constraints are considered during the coloring process, limiting the number of actions within a single batch to eight and ensuring that total resource requirements do not exceed 80% of available resources. Color assignment prioritizes high-priority actions, with priority scores calculated based on the expected improvement in network performance and the urgency of the action. Competing node pairs are identified through conflict-constrained edge connections in the network; nodes with edge connections must be assigned different colors to ensure they do not execute within the same time window.

[0114] Parallel execution units are formed by combining refactoring actions with the same batch identifier. Actions within a unit are executed using parallel scheduling. Parallelism control is dynamically adjusted based on resource availability and execution complexity, with a maximum of 6 parallel actions per execution unit. The data structure of an execution unit includes four attributes: an action list, a summary of resource requirements, estimated execution time, and dependencies. The summary of resource requirements is calculated using vector addition, the estimated execution time is the execution time of the longest action within the unit, and the dependencies record the sequential constraints with other execution units.

[0115] The execution time window allocation employs a time-slice round-robin scheduling algorithm, with a time slice size of 15 minutes and a 5-minute buffer interval between windows. Priority is determined based on a comprehensive score of the execution unit, with evaluation metrics including performance improvement potential, implementation urgency, and resource efficiency. Performance improvement potential is predicted through network simulation; implementation urgency is based on fault risk and business impact assessments; and resource efficiency is calculated using the return on investment (ROI). Maintenance time constraints are considered during time window allocation to avoid performing refactoring actions during peak business periods, defined as times when network traffic exceeds 150% of the average.

[0116] Resource demand profiles are constructed using multi-dimensional feature vectors, encompassing five dimensions: CPU usage, memory consumption, network bandwidth usage, storage I / O requirements, and concurrent connections. Profile generation is based on statistical analysis of historical execution data, employing a moving 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 metrics including device resource utilization, link bandwidth utilization, and cache hit rate. Capacity data is updated every 2 minutes, with data accuracy guaranteed within a 5% error range.

[0117] Matching relationships are established through a comparative analysis of resource demand and available capacity. The matching algorithm employs an approximate solution method for the multidimensional bin packing problem. The algorithm considers the heterogeneity and substitutability of resource types, allowing for compensation for insufficient capacity in dimensions with surplus capacity. The threshold for successful matching is set at no more than 90% of available capacity for resource demand, with a safety margin maintained above 10%. When matching fails, resource scheduling optimization is triggered, alleviating capacity bottlenecks through dynamic resource allocation and load migration.

[0118] Conflict time window identification is based on a resource overload detection algorithm, with a detection period of 3 minutes before the start of each time window. Overload conditions are defined as demand exceeding available capacity in any resource dimension, or a combined load index exceeding 0.95 across multiple dimensions. The combined load index is calculated using a weighted average, with each dimension weighted according to resource importance: CPU and memory have a weight of 0.3, network bandwidth has a weight of 0.25, and storage and connection count each have a weight of 0.075. Detection results are stored in a conflict event table, recording the conflict type, involved resources, duration, and impact.

[0119] The dynamic reorganization of execution units employs a heuristic optimization algorithm, aiming to minimize total execution time and resource waste. The reorganization strategy includes three methods: action splitting, batch merging, and time window adjustment. Action splitting decomposes complex actions into multiple sub-actions, distributed across different time windows for execution, with dependencies ensuring the execution order between sub-actions. Batch merging combines execution units with complementary resource needs, improving resource utilization efficiency. Time window adjustment is achieved using a sliding window technique; the window size can be dynamically adjusted, ranging from 10 to 60 minutes.

[0120] The refactoring process is shifted to subsequent idle time windows using a time scheduling algorithm that maintains a global timetable and resource occupancy table. Idle windows are identified based on resource utilization thresholds; any resource with utilization below 50% within a window is marked as idle. The shifting strategy prioritizes the nearest idle window while considering dependencies between actions and business continuity requirements. During the shift, the execution plan and resource reservations are updated to ensure the adjusted solution still meets performance targets.

[0121] The generation of resource-feasible topology adjustment schemes is achieved through constraint satisfaction problem solving, with constraints including resource capacity constraints, temporal dependency constraints, and performance target constraints. The solution algorithm employs the branch and bound method, and search space pruning is based on constraint propagation and heuristic estimation. Scheme feasibility verification is completed through simulation execution, with the simulated environment replicating the resource status and load characteristics of the real network environment. The output of a verified scheme includes three parts: an execution schedule, a resource allocation plan, and a risk assessment report.

[0122] In one optional implementation, the feasible topology adjustment schemes for the resources are transformed into multi-dimensional quantified feature sequences, the network evolution patterns are extracted using tensor decomposition techniques, and an association mapping graph for topology optimization is constructed using the entropy weight ranking method, including:

[0123] Extract multi-dimensional features of network evolution from the feasible topology adjustment schemes of the resources, and reorganize the multi-dimensional features into a multi-dimensional quantified feature sequence according to the evolution time sequence of the topology structure;

[0124] The multi-dimensional quantized feature sequence is constructed into a three-dimensional state expression space. The internal structure of the multi-dimensional quantized feature sequence is revealed by tensor decomposition technology. The three-dimensional state expression space is decoupled into a combination of core features and pattern matrix. The network evolution path is identified from the pattern matrix in the time dimension.

[0125] Based on the coupling relationship between the core features and the pattern matrix, the co-evolution law of node connectivity and link load distribution is extracted, and the co-evolution law is refined into the essential evolution pattern of network structure to construct an evolutionary knowledge base that reflects the dynamic characteristics of the network.

[0126] Based on the collaborative patterns in the evolutionary knowledge base, information entropy analysis is performed on the multi-dimensional quantitative feature sequences, and the information entropy is transformed into a quantitative index of feature importance. A mapping relationship between the feature importance and the network evolution direction is established to form a correlation mapping map with predictive value.

[0127] Resource-feasible topology adjustment schemes are transformed into multi-dimensional features of network evolution through a feature extraction module. This module uses a sliding window technique to scan each execution time point in the adjustment scheme. The feature extraction window is set to a 5-minute interval, covering the entire time span of the topology adjustment process. The multi-dimensional features include four categories: network topology features, traffic distribution features, resource utilization features, and performance indicator features. Network topology features are obtained through eigenvalue decomposition of the adjacency matrix, including four indicators: connectivity, clustering coefficient, average path length, and degree distribution entropy. Traffic distribution features are based on traffic matrix analysis, extracting four parameters: traffic centroid, load balancing index, peak-to-valley ratio, and spatial correlation. Resource utilization features cover statistical characteristics of CPU utilization, memory usage, bandwidth utilization, and cache hit rate, including mean, variance, skewness, and kurtosis. Performance indicator features include time-series statistics of end-to-end latency, packet loss rate, throughput, and jitter.

[0128] Multi-dimensional features are recombined into a quantized feature sequence according to the evolutionary timeline of the topology. The recombination process employs time alignment and interpolation techniques to handle feature data with different sampling frequencies. The time alignment benchmark is set as the minimum sampling interval, and cubic spline interpolation is used to ensure data continuity. The data structure of the quantized feature sequence is represented by a three-dimensional tensor. The first dimension corresponds to the time series, with a length equal to the number of time steps in the evolutionary 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 a maximum-minimum normalization method to ensure the comparability of features with different dimensions. The sequence length is dynamically determined based on the complexity of the adjustment scheme; a single adjustment action corresponds to 10 time steps, and the total sequence length is controlled within 200 time steps.

[0129] The three-dimensional state representation space is constructed through tensor reshaping operations, reorganizing multi-dimensional quantized feature sequences into third-order tensors with temporal, spatial, and attribute dimensions. The length of the temporal dimension equals the number of evolutionary steps, the length of the spatial dimension equals the total number of nodes and links in the network, and the length of the attribute dimension equals the number of extracted feature types. Tensor elements are 64-bit floating-point numbers, with values ​​limited to 0 to 1. Tensor storage employs a compressed sparse format, leveraging the sparsity of the network topology to reduce memory usage; compression ratios typically exceed 70%. The three-dimensional state representation space supports slicing and aggregation operations, facilitating subsequent tensor decomposition processing.

[0130] Tensor decomposition techniques employ the CANDECOMP or PARAFAC decomposition methods to decompose the three-dimensional state representation space into a linear combination of several rank tensors. The decomposition algorithm uses alternating least squares, with a maximum of 1000 iterations and a convergence tolerance of 10 to the power of -6. The ranks of the decomposed tensors are determined through cross-validation, with candidate ranks ranging from 5 to 50. The rank that minimizes the reconstruction error is selected as the final parameter. The algorithm is implemented using multi-threaded parallel computation, with the number of threads set to half the number of CPU cores, and the memory requirement approximately three times the size of the original tensor.

[0131] The decomposition result is represented as a combination of core features and a pattern matrix. The core features are one-dimensional vectors with a length equal to the decomposition rank, representing the importance weight of each pattern component. The pattern matrix consists of three parts: a temporal pattern matrix, a spatial pattern matrix, and an attribute pattern matrix. The number of rows in the temporal pattern matrix equals the length of the time dimension, and the number of columns equals the decomposition rank. Matrix elements represent the activation level of different patterns at each time point. The spatial pattern matrix and the attribute pattern matrix have similar structures and meanings. The pattern matrix is ​​normalized using column normalization to ensure that the norm of each pattern component is 1.

[0132] Network evolution paths in the temporal pattern matrix are identified using a trajectory clustering algorithm, employing the density-based DBSCAN method. The clustering parameters include a neighborhood radius of 0.1, a minimum sample size of 5, and Euclidean distance as the distance metric. Evolutionary paths are represented as continuous trajectories in the temporal pattern space. Trajectory smoothing utilizes a Gaussian filter with a filtering window size of 7 time steps. Path features include five parameters: initial state, final state, path length, curvature, and velocity change. Path classification is based on shape similarity, with a similarity threshold of 0.8, forming a category library of typical evolutionary patterns.

[0133] The coupling relationship between core features and the pattern matrix was quantified through correlation analysis, using the Pearson correlation coefficient. A coupling strength threshold of 0.3 was set; feature pairs exceeding this threshold were marked as strongly coupled. The co-evolutionary relationship between node connectivity and link load distribution was extracted through multivariate time series analysis, with an analysis window length of 20 time steps and a sliding step size of 5 time steps. Co-evolutionary pattern recognition was based on the Granger causality test, with a significance level set to 0.05. The evolutionary relationship was quantified as a state transition probability matrix, with the matrix dimension being the discrete level of the network state space, typically a 10x10 matrix.

[0134] The essential evolutionary pattern of the network structure was extracted using a pattern mining algorithm, specifically the Apriori method for frequent sequence mining. The minimum support was set to 0.1, the minimum confidence to 0.7, and the maximum sequence length was limited to 8 state transitions. The evolutionary pattern was represented as a directed graph structure, where nodes represent network states, edges represent state transitions, and edge weights represent transition probabilities. Pattern evaluation was based on two metrics: information gain and prediction accuracy. The information gain threshold was set to 0.2, and the prediction accuracy threshold was set to 0.75.

[0135] The evolutionary knowledge base is constructed using a graph database to store evolutionary patterns, supporting complex graph queries and pattern matching operations. The knowledge base graph contains four entity types: state nodes, transition edges, pattern labels, and attribute fields. State nodes store the feature vectors and timestamps of network states, transition edges store transition probabilities and triggering conditions, pattern labels are used for classification and retrieval, and attribute fields record relevant quantitative indicators. Knowledge base updates employ an incremental learning approach, integrating new evolutionary data into the existing knowledge structure through pattern matching and similarity calculations.

[0136] The information entropy analysis of the multidimensional quantified feature sequence employs the Shannon entropy calculation method, with entropy values ​​calculated based on the probability distribution of the feature values. The probability distribution is obtained through histogram estimation, with the number of bins set to the integer part of the square root of the feature sequence length. The information entropy calculation window is set to 50 time steps, with a sliding step size of 10 time steps to ensure the temporal resolution of the entropy value sequence. Entropy normalization uses Shannon entropy with a logarithmic base of 2, ranging from 0 to the logarithmic dimension of the feature space.

[0137] Information entropy is converted into a feature importance quantification index through a weighting function, with the softmax function ensuring the weights sum to 1. Feature importance is inversely proportional to information entropy; lower entropy values ​​indicate greater feature importance. The weighting formula is an exponential normalization of negative entropy values. The importance threshold is set to 1.5 times the average importance; features exceeding this threshold are marked as key features. The stability of key features is validated through guided sampling, with 1000 sampling iterations. The stability index is the Kendall correlation coefficient, ranking importance, with a stability threshold set to 0.8.

[0138] The mapping relationship between feature importance and network evolution direction was established through regression analysis, using a multiple linear regression model. The evolution direction was quantified as a vector representing changes in network performance metrics, including changes in latency, throughput, and reliability. The input to the regression model was the feature importance vector, and the output was the evolution direction vector. Model training employed the least squares method. Model validation used cross-validation with a validation fold of 5, and the performance evaluation metrics were root mean square error and correlation coefficient.

[0139] The association mapping graph is represented by a weighted directed graph, where nodes represent features or evolutionary states, edges represent associations, and edge weights represent association strength. The graph construction algorithm is based on correlation analysis of feature importance and evolutionary direction, with a correlation threshold set to 0.4. Graph visualization employs a force-directed layout algorithm, where node size represents feature importance, edge thickness represents association strength, and color coding represents evolutionary direction category. Graph updates utilize an online learning approach, integrating new data into the existing structure through graph embedding technology with a 128-dimensional embedding dimension.

[0140] Predictive value was evaluated and validated using a time series prediction task. The prediction model was implemented based on a graph neural network with an association mapping graph. The network structure consisted of two graph convolutional layers and one fully connected layer, with a hidden layer dimension of 64 and the ReLU activation function. The model was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 200 training epochs. Prediction accuracy was evaluated using the mean absolute percentage error (MAE), with a target accuracy within 15%.

[0141] In one optional implementation, based on the collaborative patterns in the evolutionary knowledge base, information entropy analysis is performed on the multi-dimensional quantified feature sequences to convert the information entropy into a quantitative index of feature importance. Establishing a mapping relationship between the feature importance and the network evolution direction includes:

[0142] The dynamic evolution pattern of the topology is extracted from the evolutionary knowledge base. The evolutionary activity of the multi-dimensional quantized feature sequence is evaluated based on the coupling rules in the dynamic evolution pattern. Feature groups whose evolutionary activity exceeds a preset activity threshold are marked as important evolutionary features. The temporal distribution of the important evolutionary features is preliminarily evaluated.

[0143] Based on the temporal distribution characteristics of the important evolutionary features, an information entropy evaluation system is constructed to quantify the uncertainty of their dynamic changes. The information entropy index is integrated with the degree of evolutionary activity in multiple dimensions to form a quantitative weighting system that reflects the importance of the features.

[0144] The important evolutionary features are ranked and sorted according to the quantized weight system. The top-ranked feature set is extracted as the dominant feature group of network evolution. By analyzing the development trend of the dominant feature group in the dynamic evolution mode, the evolution law of network structure is revealed.

[0145] By establishing a correspondence between the development trend of the dominant feature group and the historical optimization experience in the evolutionary knowledge base, the path to improve network performance is identified, and a mapping relationship between feature importance and network evolution direction is constructed.

[0146] When extracting dynamic evolution patterns of topology from an evolutionary knowledge base, the system identifies the changing paths of node connectivity by traversing historical network state data stored in the knowledge base. Specifically, the system reads at least 1000 sets of historical topology snapshots from the knowledge base. Each snapshot records the connectivity status, edge weights, and node attributes of each node in the network at a specific moment. By comparing and analyzing snapshots from adjacent moments, the system extracts basic evolutionary events such as node additions and deletions, edge creation and deletions, and weight changes. During the extraction process, the system organizes these basic events into an evolutionary sequence in chronological order and uses a sliding time window method to statistically analyze the frequency of different evolutionary events. The time window is set to 10 time units, and the window slides in steps of 2 time units. Within each window, the system statistically analyzes the numerical fluctuation range of features such as changes in node degree, clustering coefficient, and path length.

[0147] When evaluating the evolutionary activity of multi-dimensional quantitative feature sequences based on extracted dynamic evolutionary patterns, a coupling law evaluation mechanism is established. This mechanism reflects the intensity of dynamic change of a feature by calculating the difference between values ​​at adjacent time points in the feature sequence. For a feature sequence containing 50 time points, the system calculates 49 difference values ​​and uses the sum of the absolute values ​​of these difference values ​​as an indicator of change intensity. Simultaneously, the system analyzes the co-evolutionary relationship between the feature sequence and other feature sequences, quantifying the coupling strength by calculating the correlation between the difference values ​​of two feature sequences. When the coupling strength between a feature sequence and at least three other feature sequences exceeds 0.7, the feature is considered to have significant co-evolutionary characteristics. Evolutionary activity is derived by comprehensively considering the feature's own change intensity and its coupling strength with other features. Specifically, the change intensity is normalized to the 0-1 interval and then weighted and averaged with the average coupling strength, with weights set to 0.6 and 0.4, respectively.

[0148] A preset activity threshold of 0.55 was set, and features with evolutionary activity exceeding this threshold were marked as important evolutionary features. In a real-world case with 200 original features, 43 important evolutionary features were selected through activity assessment. When initially evaluating the temporal distribution of these important evolutionary features, the system analyzed the distribution density and clustering characteristics of feature values ​​along the time axis. Specifically, the entire time range was divided into 10 equal-length time periods, and the number and magnitude of feature value changes were counted within each time period. When the number of changes in a certain time period exceeded 20% of the total number of changes, that time period was marked as an active period for feature evolution. In the real-world case, it was found that important evolutionary features exhibited significant clustering changes in time periods 3 and 7, with the cumulative change magnitude in these two time periods accounting for 65% of the total change magnitude.

[0149] When constructing an information entropy evaluation system for the temporal distribution characteristics of important evolutionary features, the distribution of feature values ​​is taken as the analysis object. The system divides the range of feature values ​​into 20 equally spaced intervals and counts the frequency of feature values ​​falling into each interval. By calculating the proportion of each interval's frequency to the total frequency, the probability distribution of the feature values ​​is obtained. Information entropy is calculated by multiplying the logarithm of the probability value of each interval by the probability value and then summing the results to obtain the negative 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 size of each interval is obtained as 8, 12, 15, 20, 18, etc. After calculating the probability of each interval, the information entropy value of the feature is obtained by logarithmic operation and weighted summation, resulting in a value of 2.35. A higher information entropy value indicates that the feature value distribution is relatively uniform, the uncertainty of the change is high, reflecting that the feature has strong randomness and complexity in the evolution process.

[0150] When integrating information entropy and evolutionary activity across multiple dimensions, a comprehensive quantification weighting mechanism is established. This mechanism normalizes the information entropy value and multiplies it by the evolutionary activity to obtain an initial fusion weight. Considering the differences in the contributions of information entropy and activity in different scenarios, the system introduces an adaptive adjustment factor. The adjustment factor is dynamically set based on 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 a practical case, the normalized information entropy value of a certain feature is 0.78, the evolutionary activity is 0.62, and the ratio of standard deviation to mean is 0.15, corresponding to an adjustment factor of 1.05. Finally, the quantification weight of this feature is calculated as the product of 0.78 and 0.62, multiplied by the adjustment factor, resulting in 0.51.

[0151] When ranking important evolutionary features according to the quantified weighting system, they are arranged in descending order of weight value. In the case containing 43 important evolutionary features, the feature set with the top 30% weight values ​​after ranking is extracted as the dominant feature group, resulting in 13 dominant features. These dominant features include the average path length change rate of the network, the degree distribution concentration of core nodes, and the variance change of edge weights. The system further analyzes the development trend of these dominant features in the dynamic evolution mode, identifying the trend type by fitting the feature value change curve over time. For features showing a continuous upward trend, the system determines that the network is developing in the direction represented by that feature; for features showing periodic fluctuations, the system extracts the fluctuation period and amplitude parameters as descriptive indicators of the evolution law. In the case, the average path length change rate shows a continuous downward trend, with a decrease of 0.03 per time unit, indicating that the network structure is trending towards a more compact connection state.

[0152] When establishing a correspondence between the development trend of the dominant feature group and historical optimization experiences in the evolutionary knowledge base, the system retrieves historical optimization cases recorded in the knowledge base. The knowledge base stores over 300 historical network optimization instances, each recording the trajectory of the dominant feature changes before and after optimization and the final performance improvement achieved. The system identifies matching optimization experiences by calculating the similarity between the current dominant feature development trend and the feature change trajectory in historical cases. Similarity is quantified by comparing the shape features of the trend curve and the positional deviation of key inflection points; a similarity exceeding 0.8 is considered a high match.

[0153] In real-world cases, the system identified 15 highly matched historical optimization cases. These cases showed that when the average path length continuously decreased and the concentration of core node degree distribution increased simultaneously, increasing cross-community connection edges could improve network transmission efficiency by 40%. Based on these matching optimization experiences, the system established a mapping relationship between feature importance and network evolution direction, using the weight values ​​of dominant features as a quantitative indicator of mapping strength, and the operation types and parameter adjustment ranges in the optimization experiences as the target output of the mapping, forming a complete mapping function from the feature space to the evolution direction space.

[0154] A second aspect of the present invention provides an adaptive network topology dynamic reconstruction system based on deep reinforcement learning, comprising:

[0155] The first unit is used to acquire the current network topology status information, service traffic distribution information and historical reconstruction records; to perform multi-level feature extraction on the topology status information, to capture the topology association features between nodes through graph convolution operation, and to generate a fused status representation in combination with the service traffic distribution information.

[0156] The second unit is used to input the fusion state representation and the historical reconstruction record into a deep reinforcement learning model, identify bottleneck nodes and redundant links in the topology, and output a candidate set of reconstruction actions; evaluate the long-term cumulative benefit of each candidate action through Monte Carlo tree search, and select the optimal reconstruction action sequence.

[0157] The third unit is used to perform phased verification of the optimal reconstruction action sequence. It allocates time slots to each reconstruction action using a graph coloring algorithm, identifies and processes action combinations with resource conflicts, and generates resource-feasible topology adjustment schemes.

[0158] The fourth unit is used to transform the feasible topology adjustment schemes of the resources into multi-dimensional quantitative feature sequences, extract network evolution laws using tensor decomposition technology, and construct a correlation mapping map for topology optimization using entropy weight sorting method to realize intelligent evolution of network architecture.

[0159] A third aspect of the present invention provides an electronic device, comprising:

[0160] processor;

[0161] Memory used to store processor-executable instructions;

[0162] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0163] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0164] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

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