Cooperative adaptive routing method applied to communication network
By introducing a cooperative adaptive routing method into the communication network, multi-protocol cooperative selection and parameter optimization are achieved, solving the problem of insufficient protocol adaptability in the existing technology and improving network performance and adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2026-03-02
- Publication Date
- 2026-04-17
AI Technical Summary
Existing communication network routing protocols lack a multi-protocol collaborative selection mechanism when facing dynamic topologies and changing service loads. They rely on large-scale information collection, resulting in high overhead. Fixed thresholds lack adaptability, and reinforcement learning algorithms are computationally complex and unstable, making it difficult to meet the QoS requirements of various service types.
A cooperative adaptive routing method is adopted, which deploys a protocol switching module, a threshold adaptive optimization module, and a parameter tuning module through cluster head nodes. By combining network topology changes and service load information, it realizes multi-protocol collaborative selection and parameter optimization, reduces control overhead, and improves adaptability and robustness.
It achieves multi-protocol collaborative optimization in dynamic network environments, reduces computational and control overhead, improves the timeliness and adaptability of routing decisions, and meets different QoS requirements.
Smart Images
Figure CN121887701A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of network communication technology, and specifically relates to a cooperative adaptive routing method applied to communication networks. Background Technology
[0002] The rapid development of electronic and communication technologies has greatly expanded the application scope of communication networks, including but not limited to smart cities, industrial automation, and the Internet of Things (IoT). As the core support for information transmission, communication networks face increasingly complex business demands and network environment challenges. Especially in network environments with highly dynamic topologies and varying service loads, traditional routing protocols often struggle to balance factors such as latency, throughput, reliability, and overhead.
[0003] Frequent changes in network topology, unstable link quality, and uneven node load can all significantly impact the performance of communication networks. Especially in complex application scenarios, such as real-time communication, data transmission, and high-priority task processing, single or static routing strategies are insufficient to effectively meet the Quality of Service (QoS) requirements of various service types. Therefore, adaptive adjustment of routing protocols for communication networks has become one of the key issues in improving network performance and adapting to different needs.
[0004] Currently, improvements to communication networks mainly focus on optimizing single protocols or adaptively adjusting protocols through machine learning and other intelligent algorithms. However, existing technologies have several shortcomings: 1) They lack multi-protocol collaborative selection mechanisms that consider multi-dimensional characteristics such as network size, topology dynamics, and node load, typically performing only local optimization within a single protocol; 2) They lack a refined perception and low-overhead decision-making framework that combines source-destination local topology features, leading to routing decisions relying on extensive information collection, increasing control overhead, and reducing decision-making efficiency; 3) Existing protocol switching mechanisms either use fixed thresholds to determine network status, lacking adaptability, or employ reinforcement learning algorithms, but these have long convergence times and high computational costs; 4) Reinforcement learning algorithms have high computational complexity when adjusting many parameters and may lead to network performance instability. Therefore, when facing dynamic network states and business requirements, current methods often exhibit poor adaptability and require significant computational and control overhead. Summary of the Invention
[0005] The purpose of this invention is to address the following issues in existing technologies: 1) lack of a joint mechanism for multi-protocol collaborative selection and protocol parameter adjustment; 2) reliance on large-scale information collection for routing decisions; 3) lack of adaptive capability when switching protocols or using fixed thresholds, or the lack of sufficient consideration of convergence time and computational overhead when using machine learning methods for adaptive adjustment; and 4) difficulty in deploying high-frequency sensing and high-dimensional action evaluation on resource-constrained nodes. This invention provides a cooperative adaptive routing method for communication networks to achieve joint optimization of multiple routing protocols and their internal parameters, thereby improving the overall communication performance and adaptability of the network while ensuring controllable computational overhead.
[0006] The technical solution adopted in this invention is:
[0007] A cooperative adaptive routing method for communication networks defines that nodes in the communication network have the same data forwarding capability, and all nodes form a cluster, including:
[0008] A protocol switching module, a threshold adaptive optimization module, and a parameter tuning module are deployed on the cluster head node. The protocol switching module adaptively switches between multiple protocols in each sensing period based on the communication sensing information of network topology changes and service load, and in combination with dynamic thresholds and elastic sensing periods. The switching control message is broadcast to the other nodes, and the other nodes execute the corresponding protocol switching strategy according to the received switching control message.
[0009] Cluster head nodes can be configured to collect multi-dimensional QoS metrics in real time to quantify network performance, including packet delivery rate (PDR), average end-to-end latency (E2ED), control overhead (OH), and average throughput (TH), and perform statistics within a set evaluation window, such as the sensing period.
[0010] The method for obtaining communication-aware information based on network topology changes and service load is as follows:
[0011] Construct containing domains In each perception cycle Before starting, the source node is established in the network by the currently running routing protocol. To the destination node Working path, source node Based on the local routing table and the path record information provided by the routing protocol, determine the node sequence on the working path s–d at time t. , For the first in the path , 1 node , The path endpoint sequence number is used to construct a service path registration message carrying the source node identifier, destination node identifier, service flow identifier, and the node sequence. This message is reported to the cluster head node via the intra-cluster control channel. The cluster head node establishes a mapping relationship between service flows and working paths locally. Map to working path , Indicates the current service flow identifier; cluster head node uses Centered on each node, in the current network topology graph Perform a k-hop neighborhood search to obtain the corresponding set of k-hop reachable nodes. where k is a positive integer. The set of nodes in the current topology. Given the set of edges in the current topology; take the union of the neighborhood sets of all nodes to obtain the set of nodes in the source-destination relational subgraph. :
[0012] ,
[0013] Will be The set of edges formed by the physical links between nodes is denoted as . Thus, a subgraph containing the domain is obtained. ;
[0014] right Internal nodes and links collect communication-aware and load-aware information: Nodes in Obtain your own geographical location information Current speed information At the same time, the nodes are synchronized via a clock synchronization mechanism. and all neighboring nodes The location and velocity information are time-aligned; then, the obtained location and velocity information are used to calculate the node values respectively. with neighboring nodes relative velocity between ,node and neighboring nodes relative distance between ,node and neighboring nodes Link quality between Based on changes in relative node speed and link quality, nodes Calculate the node variation factor of the network topology , used to quantify topological changes between nodes in a network;
[0015] In each perception cycle ,node The current queue occupancy status can be read by calling the queue query interface provided by the MAC layer. ;
[0016] The result and Encapsulated as a status reporting message, it is sent to the cluster head node. The cluster head node will then receive the message. and After summing and normalizing, we obtain the current network topology dynamic factor R and service load factor L:
[0017] ,
[0018] ,
[0019] in, The current number of nodes, based on the threshold. Determine whether to include it in the handover decision. For network topology dynamic factor threshold, The business load factor threshold is specifically when... and When, the active protocol is selected as the target protocol, when and In this case, select the on-demand protocol as the target protocol; otherwise, leave it unchanged.
[0020] The method for broadcasting the handover control message to the other nodes is as follows: the protocol handover module of the cluster head node constructs and broadcasts the protocol handover control message. , This includes the target protocol identifier and the perception cycle. Non-cluster head node receives Then, execute the protocol corresponding to the target protocol identifier, according to... Adjust the next sensing cycle;
[0021] The threshold adaptive optimization module is used to optimize the threshold. , Perception cycle The optimization involves implementing adaptive threshold optimization based on the differential evolution algorithm, including:
[0022] The threshold to be optimized , Perception cycle The parameter vector encoded as a differential evolution algorithm is denoted as:
[0023] ,
[0024] The range of values is limited to:
[0025] ,
[0026] in, , , , These are all set thresholds;
[0027] Construct a fitness function to measure the fitness of a given parameter vector. Performance advantages and disadvantages of routing switching strategies:
[0028] ,
[0029] in, The weights are non-negative and satisfy the following conditions: ;
[0030] The threshold adaptive optimization module randomly generates parameters within a preset parameter space. The candidate solutions serve as the initial population, with each individual corresponding to a different set of solutions. The value is denoted as:
[0031] ,
[0032] Each component is initialized as follows:
[0033] ,
[0034] ,
[0035] ,
[0036] in, Represents a uniformly random number in the interval [0,1].
[0037] The threshold adaptive optimization module substitutes the network performance metrics obtained by the cluster head nodes based on the elastic sensing period time window into the fitness function to calculate the initial fitness value of all individuals. ;
[0038] Perform differential mutation operation: at the first In the iteration, for each target individual in the population The threshold adaptive optimization module randomly selects three distinct individuals from the population. ,and Construct the mutation vector:
[0039] ,
[0040] in, This is the differential variation scaling factor, used to control the variable asynchronous length;
[0041] Perform cross operations for each target individual Its corresponding mutation vector Perform crossover operations to generate experimental individuals. ; for the parameter vector of the first One portion, Generate random numbers ;like or Then let Otherwise, let ;in, For crossover probability, The index is randomly selected from {1,2,3} to ensure that each experimental individual has at least one component from the mutation vector;
[0042] Execution Selection and Iterative Update: The Threshold Adaptive Optimization Module in Subsequent Experimental Individuals The corresponding parameter vector is used to configure the protocol switching strategy, and network performance is statistically analyzed and its fitness value is calculated. Then, the optimal selection step is performed:
[0043] ,
[0044] Repeat the differential mutation, crossover, selection, and iterative update processes until the preset maximum number of iterations is reached. The iteration stops when the optimal fitness improvement is less than a threshold for several consecutive generations; after the iteration ends, the individual with the highest fitness value is selected from the final population. As the current optimization result, its parameter components are:
[0045] ,
[0046] in, , , These are the parameters obtained after optimization;
[0047] The protocol parameter tuning module adaptively adjusts the internal parameters of the selected protocol. Specifically, it adopts a lightweight multi-parameter joint adjustment strategy generation mechanism based on Dirichlet distribution. The decision nodes in the network generate parameter adjustment weight vectors according to the network status, and the adjustment direction is determined by rules. By judging topology changes and service load, it determines whether to increase or decrease the key parameters of the protocol.
[0048] This module adopts a lightweight multi-parameter joint adjustment strategy generation mechanism based on Dirichlet distribution. It outputs a multi-parameter joint adjustment weight vector through reinforcement learning and a forward inference, and determines the adjustment direction of each parameter by combining the rules, thereby realizing the multi-parameter joint adjustment of the current routing protocol.
[0049] The cluster head node's protocol parameter tuning module, based on a reinforcement learning algorithm, generates parameter adjustment weight vectors according to network conditions (network topology dynamics, service load factors, sensing cycle time, and current network performance). These m weight vectors... The output is generated based on the Dirichlet distribution, and the normalization constraint is satisfied to keep the total parameter tuning amplitude under control.
[0050]
[0051] The adjustment direction is determined by rules. By assessing topology changes and service load, it is determined whether to increase or decrease key protocol parameters (such as Hello interval, route timeout, and transmission rate). For example, when topology changes drastically, the Hello interval and active route timeout are reduced to improve the sensitivity of link failure detection; when the load is high, the RREQ transmission rate is reduced to lower the network load.
[0052]
[0053] in, This indicates increasing the k-th parameter. This indicates decreasing the k-th parameter. This indicates that the parameters remain unchanged.
[0054] The total budget for the overall dispatch is Then the first The update magnitude of each parameter for
[0055]
[0056] And update the parameter size according to the direction:
[0057]
[0058] in, It is the magnitude of the k-th parameter at time t. It is the size of the k-th parameter after the update.
[0059] 2. The cooperative adaptive routing method applied to a communication network according to claim 1, characterized in that the multiple protocols are two protocols, namely AODV protocol and OLSR protocol, corresponding to, when and When the topology is determined to be relatively stable but the load is high, the active OLSR protocol is selected as the target protocol; when and When the topology changes drastically and the load is low, On-Demand AODV is selected as the target protocol.
[0060] The beneficial effects of this invention are as follows:
[0061] 1) By introducing a multi-protocol collaborative selection mechanism based on multi-dimensional characteristics such as network scale, topology dynamics and node load, the joint adaptive optimization of routing protocol type and key parameters within the protocol is realized, which significantly improves the adaptability under different QoS requirements and highly dynamic topology scenarios.
[0062] 2) State awareness is achieved by using a k-hop local topology subgraph centered on the source-destination, which effectively reduces the control overhead caused by the collection of information from the entire network while ensuring decision accuracy, and improves the timeliness and scalability of routing decisions.
[0063] 3) Genetic algorithms are used to perform global adaptive optimization of topology thresholds, load thresholds, and sensing cycles, replacing traditional fixed threshold configurations and threshold learning algorithms based on reinforcement learning. This enhances the sensitivity and robustness of protocol switching decisions to changes in the network environment, and reduces computational overhead and algorithm convergence time.
[0064] 4) A lightweight strategy generation mechanism based on Dirichlet distribution is adopted. A multi-parameter joint adjustment weight vector is output through a single forward inference, and the adjustment direction of each parameter is determined by rules. This enables joint adjustment of multiple key parameters of the routing protocol, avoiding parameter-by-parameter enumeration and evaluation in the high-dimensional action space, thereby significantly reducing inference and decision-making overhead. At the same time, the normalization constraint of the weight vector is used to limit the total parameter adjustment amplitude and suppress the instability caused by large fluctuations of multiple parameters at the same time. Attached Figure Description
[0065] Figure 1 This is a framework for a cooperative adaptive routing method based on a communication network in this embodiment of the invention;
[0066] Figure 2 This is the method and steps of the adaptive routing protocol switching module implemented based on dynamic threshold and elastic sensing period in the embodiments of the present invention;
[0067] Figure 3 These are the steps and methods of the threshold adaptive optimization module in this embodiment of the invention;
[0068] Figure 4 This is a method for implementing a protocol parameter tuning module based on lightweight reinforcement learning in an embodiment of the present invention. Detailed Implementation
[0069] To make the objectives, technical solutions, and advantages of this application clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0070] See Figure 1 and Figure 2 , Figure 1 This is a step diagram of a cooperative adaptive routing method based on a communication network according to an embodiment of the invention. Figure 2 This is a framework for a cooperative adaptive routing method based on a communication network, as described in an embodiment of the invention.
[0071] Unmanned aerial vehicle (UAV) networks are a specific implementation of communication networks. In a preferred embodiment of the present invention, the physical nodes involved are communication nodes with the same data forwarding capability, and these nodes constitute a cluster. The node module comprises:
[0072] (1) Protocol switching module, which uses AODV protocol and OLSR protocol as candidate routing protocol set, is deployed in the cluster head node of the communication network as the decision node in the network.
[0073] (2) Threshold Adaptive Optimization Module, responsible for implementing adaptive threshold optimization based on differential evolution algorithm. Deployed in cluster head nodes of communication network.
[0074] (3) Protocol parameter tuning module: Based on multi-dimensional QoS requirements, it adopts a lightweight policy generation mechanism based on Dirichlet distribution and combines it with PPO reinforcement learning algorithm to jointly adjust multiple parameters. Deployed at cluster head nodes in the communication network.
[0075] (4) GNSS module, responsible for outputting the current node's position and speed information, deployed on all nodes of the communication network.
[0076] (5) Information interaction module: Based on the constructed information, each node interacts with control and perception information. Deployed on all nodes of the communication network.
[0077] (6) The underlying communication module handles the underlying communication protocols and data transmission, ensuring reliable information delivery. It is deployed on all nodes of the communication network.
[0078] Cluster head nodes will perceive multi-dimensional QoS indicators in real time and perform statistics within a preset time window, such as the perception period. These indicators include, but are not limited to, PDR, average end-to-end latency (E2ED), control overhead (OH), and average throughput (TH).
[0079] Furthermore, the protocol switching module includes:
[0080] The cluster head node, as the main body for handover decision-making and execution, adaptively selects and switches between AODV and OLSR protocols in each sensing period based on the communication sensing information of network topology changes and service load, combined with the dynamic threshold and elastic sensing period output by the threshold adaptive optimization module; the other nodes in the cluster execute the corresponding protocol handover strategy after receiving the handover control message broadcast by the cluster head.
[0081] Assuming that business types can be classified according to latency sensitivity and reliability sensitivity; the business targeted by this embodiment of the invention is a latency-tolerant and reliability-sensitive business, that is, the end-to-end latency requirements are relatively relaxed, but the requirements for packet delivery success rate are high.
[0082] S11. Construct the containing domain ;
[0083] In each perception cycle Before starting, when the source node Need to send to the host node When initiating or maintaining a service flow transmission, the currently running routing protocol first establishes a connection in the network from... to The working path. After the route is established, the source node Based on the local routing table and the path record information provided by the routing protocol, determine the node sequence on the working path s–d at time t. , As the destination node, a service path registration message is constructed carrying the source node identifier, destination node identifier, service flow identifier FlowID, and the node sequence, and reported to the cluster head node through the intra-cluster control channel.
[0084] After receiving the service path registration message, the cluster head node establishes a mapping relationship between the service flow and the working path locally. ID Map to working path This is used for subsequent domain construction and protocol switching decisions; when a working path is rerouted, the source node or a preset node in the path re-reports the business path registration message, and the cluster head node updates the mapping relationship accordingly.
[0085] Cluster head node Centered on each node, in the current network topology graph Perform a k-hop neighborhood search to obtain the corresponding set of k-hop reachable nodes. k is a positive integer, preferably k is 1 to 3. The set of nodes in the current topology. Let `x` be the set of edges in the current topology. Then, take the union of the neighborhood sets of all nodes to obtain the set of nodes in the source-destination relation subgraph. .
[0086]
[0087] At the same time, it will be by The set of edges formed by the physical links between nodes is denoted as . Thus, a subgraph containing the domain is obtained. .
[0088] By constructing the domain as described above, it necessarily includes all intermediate nodes on the source-destination work paths and alternative paths. Furthermore, through k-hop expansion, it introduces neighboring nodes strongly related to these paths, achieving localized modeling of areas highly relevant to the current business flow. This allows for perception and decision-making only within the local subgraph most relevant to the current business flow, reducing interference from non-critical nodes and lowering overhead.
[0089] S12. To Internal nodes and links collect communication-aware and load-aware information;
[0090] This embodiment provides a dynamic factor sensing method based on topology changes and node motion. Specifically, it includes acquiring node position and velocity information through a local GNSS system and estimating topology changes between adjacent nodes based on this information. The method includes the following steps:
[0091] Nodes in Obtain its current geographical location information through the local GNSS module. This includes the node's longitude, latitude, and altitude information. Additionally, the node... It also obtains its speed information at the current moment. .node and all neighboring nodes The location and speed information will be acquired through periodic updates, and the data will be time-aligned through a clock synchronization mechanism.
[0092] S121. Obtain the relative velocity of the nodes;
[0093] node Using locally acquired speed information, calculate the node with neighboring nodes relative velocity between The relative velocity is estimated using the following formula.
[0094]
[0095] in, and They are nodes and nodes At any moment The velocity vector.
[0096] node and neighboring nodes relative distance between It can be calculated using its location information, as shown in the following formula:
[0097] in, and For nodes and nodes At any moment Location information.
[0098] node and neighboring nodes Link quality between The formula is derived from the relative distance:
[0099]
[0100] in, This refers to the communication range of the drone.
[0101] Based on changes in relative node speed and link quality, nodes Calculate the node variation factor of the network topology It is used to quantify topological changes between nodes in a network, where For the perception cycle, For nodes The set of neighbors at time t. The topology change factor is calculated using the following formula:
[0102]
[0103] S122. Obtain load-related state variables
[0104] Read the current queue occupancy percentage from the MAC layer queue management module.
[0105] In each perception cycle Calculate the average occupancy rate during this period, and the nodes. The current queue occupancy status can be read by calling the queue query interface provided by the MAC layer:
[0106]
[0107] in express The number of packets to be sent in the time queue. This represents the maximum capacity of the queue. This refers to the number of statistical occurrences within the sensing period.
[0108] S13. Encapsulation status reporting message;
[0109] Each containing domain node After local sensing, the aforementioned state variables are encapsulated into a state reporting message. The message format may include, but is not limited to, the following fields. The message format is shown in Table 1.
[0110]
[0111] S14. Calculate the network topology dynamic factor and service load factor, and make cluster head switching decisions.
[0112] The cluster head node will receive and After summing and normalizing, the results are given as the current network topology dynamic factor R and service load factor L, as shown in the following formula: The current number of nodes. The current threshold is obtained from the second-layer threshold adaptive optimization module. and the next cycle of elasticity sensing cycle .like or Then proceed to the switching decision;
[0113]
[0114]
[0115] when and When the topology is determined to be relatively stable but with high load, the active OLSR protocol is selected as the target protocol, i.e., set... .
[0116] when and When the topology changes drastically and the load is low, On-Demand AODV is selected as the target protocol, i.e., it is set... .
[0117] In all other cases, the agreement remains unchanged.
[0118] S15. Cluster head broadcast switching control message
[0119] Cluster head nodes construct broadcast protocol switching control messages The field format can be defined as shown in Table 2.
[0120]
[0121] Non-cluster head node reception Then, ProtoTarget and Tsense are read; the cluster head adjusts the sensing cycle for the next cycle based on Tsense, increasing it in steady state and shortening it in volatile state, thus achieving low-overhead closed-loop adaptive control.
[0122] In a preferred embodiment of the present invention, the threshold adaptive optimization module employs a differential evolution algorithm to jointly optimize the topology dynamic threshold, load threshold, and sensing period. The module takes the network performance metrics collected by the cluster head node within several sensing periods as input, iteratively searches for near-optimal solutions in the parameter space, and transmits the optimization results to the protocol switching module.
[0123] Specifically, the threshold adaptive optimization module may include the following steps:
[0124] S21. Definition of Parameter Vector and Fitness Function
[0125] S211. Encode the parameter combination to be optimized into a parameter vector for the differential evolution algorithm. Preferably, the topology dynamic threshold, load threshold, and sensing period are denoted as...
[0126]
[0127] in, For network topology dynamic factor threshold, For the business load factor threshold, This refers to the time interval between two consecutive sensing operations. All three parameters are limited to a preset physically feasible range:
[0128]
[0129] S212. Construct a fitness function based on the above normalization index to measure a given parameter vector. The performance of the next routing switching strategy is evaluated. Preferably, a fitness function of the following form can be used:
[0130]
[0131] in, The weights are non-negative and satisfy the following conditions: This reflects the emphasis placed on delivery rate, throughput, latency, and overhead under different business scenarios. It demonstrates the adaptability of the method to different scenarios. A higher fitness value indicates better overall network performance under the corresponding parameter vector. In this embodiment of the invention, the business type is reliability-sensitive, so preferably... maximum.
[0132] S22. Initial Population Generation
[0133] The threshold adaptive optimization module randomly generates parameters within a preset parameter space. The candidate solutions serve as the initial population, with each individual corresponding to a different set of solutions. The value is denoted as:
[0134]
[0135] Each component can be initialized as follows:
[0136]
[0137] in It represents a uniform random number in the interval [0,1].
[0138] Within the initial sensing periods, the threshold adaptive optimization module executes protocol switching logic according to the threshold and sensing period configuration for each individual. The statistically obtained network performance indicators are then substituted into the fitness function to calculate the initial fitness value for all individuals. .
[0139] S23. Differential Mutation Operation
[0140] In the In the iteration, for each target individual in the population The cluster head node randomly selects three distinct individuals from the population. ,and Construct the mutation vector:
[0141]
[0142] in, This is the differential variation scaling factor, used to control the variable asynchronous length.
[0143] For the mutated vector If any component exceeds the allowable range of the corresponding parameter, it is truncated or projected back into the feasible range to ensure that all candidate thresholds and sensing cycles meet the preset physical constraints.
[0144] S24. Crossover Operation
[0145] To enhance population diversity and avoid premature convergence, for each target individual Its corresponding mutation vector Perform crossover operations to generate experimental individuals. Preferably, a binary crossover strategy can be used:
[0146] For the parameter vector of the first Each component ( Generate random numbers ;like or Then let Otherwise, let .
[0147] in, For crossover probability, The index is randomly selected from {1,2,3} to ensure that each experimental individual has at least one component from the mutation vector.
[0148] S25. Selection and Iterative Update
[0149] The threshold adaptive optimization module in subsequent experimental individuals The corresponding parameter vector is used to configure the protocol switching strategy, and network performance is statistically analyzed and its fitness value is calculated. Then, the optimal selection step is performed:
[0150]
[0151] The above selection mechanism ensures that the average fitness of the new generation population does not decrease. S23–S25 are repeated for several generations until the preset maximum number of iterations is reached. Alternatively, the iteration can be stopped when the improvement in the optimal fitness over several consecutive generations is less than a threshold.
[0152] After the iteration is complete, select the individual with the highest fitness value from the final population. As the current optimization result, its parameter components are:
[0153]
[0154] S26. Threshold and Sensing Cycle Deployment and Online Re-optimization
[0155] The threshold adaptive optimization module will optimize the obtained and As the current working threshold for network topology dynamics factors and service load factors, As the next sensing cycle, the elastic sensing cycle is passed to the protocol switching module for subsequent sensing and routing protocol selection.
[0156] During network operation, when a significant change in the service scenario is detected (e.g., change in service type, large adjustment of node size, long-term average PDR or latency deviating from the target range), the threshold adaptive optimization module can trigger the reinitialization of the population and re-execute S21-S25, thereby achieving [the desired effect]. The online re-optimization enables the protocol switching module to continuously track changes in the network environment and business requirements.
[0157] Regarding the protocol parameter tuning module, assuming that the current node is running the AODV protocol, the following uses AODV as an example to illustrate the multi-parameter adaptive scheme based on different QoS requirements and current communication network environment awareness, using reinforcement learning algorithms.
[0158] In a preferred embodiment of the present invention, a protocol parameter tuning module is deployed at the cluster head node. When the protocol switching module determines that the currently selected routing protocol is AODV, it performs joint adaptive adjustment of multiple key parameters within the AODV protocol. This module employs a lightweight strategy generation mechanism based on Dirichlet distribution for weight generation and rule-based direction determination. This allows for the output of a multi-parameter joint adjustment weight vector after a single forward inference, and parameter updates are completed without parameter-by-parameter enumeration and evaluation in the high-dimensional action space. This reduces inference and decision-making overhead and suppresses instability caused by simultaneous large fluctuations in multiple parameters.
[0159] S31. Constructing the state space
[0160] A state vector is constructed by using network performance metrics collected from cluster head nodes and network topology dynamic factors and service load factors from the protocol switching module. The state vector includes: network topology dynamic factors. Business load factor Group delivery rate End-to-end delay Routing control overhead and perception cycle ;
[0161]
[0162] Where at time t, R L is a network topology dynamic factor containing a domain. For the service load factor of the included domain, PDR E2ED is the packet delivery rate during the sensing cycle. OH is the average end-to-end delay. For routing control overhead;
[0163] Preferably, to avoid bias in the strategy output caused by different units, the protocol parameter tuning module normalizes the state vector to obtain... The normalization adopts linear extremum normalization, so that the state variables of each dimension are mapped to the interval [0,1].
[0164] S32. Construct the action space, i.e., the set of adjustable parameters and their range of values for AODV.
[0165] The protocol parameter tuning module builds a set of adjustable parameters for AODV. , is represented as:
[0166]
[0167] in, Preferably, the parameters include, but are not limited to:
[0168] Hello interval ;
[0169] Active route timeout ;
[0170] RREQ Sending Rate Limit ;
[0171] RERR (Redirect Rate Limit) ;
[0172] Extended ring search TTL increment ;
[0173] Furthermore, to ensure the deployability and stability of the protocol, a range of allowed values is set for each parameter:
[0174]
[0175] And out-of-bounds values are truncated when parameters are updated.
[0176] S33. Generating a multi-parameter joint adjustment weight vector based on Dirichlet distribution.
[0177] The protocol parameter tuning module will normalize the state. Input Policy Network Output the concentration parameter vector of the Dirichlet distribution. It satisfies that all its components are positive:
[0178]
[0179] To reduce online sampling jitter and inference overhead, the expectation of the Dirichlet distribution is preferably used as the deterministic weight vector:
[0180]
[0181] in,
[0182]
[0183] The This represents the proportion of the joint adjustment range of each adjustable parameter under the current network state. Since the weight vector naturally satisfies the normalization constraint, the total parameter adjustment range in this cycle is controllable, thereby suppressing large fluctuations of multiple parameters simultaneously.
[0184] S34. Determine the adjustment direction of each parameter based on rules.
[0185] In this embodiment, the adjustment direction of each parameter is determined by rules, and the Dirichlet weights are only used to allocate the adjustment magnitude. Specifically, the cluster head node calculates the network topology dynamic factors based on the protocol switching module. Business load factor And performance feedback, to determine the direction vector:
[0186]
[0187] in, Indicates increasing the parameter , Indicates reducing the parameter , This indicates that the parameters remain unchanged.
[0188] Preferably, the following rules (examples) can be used to determine the direction:
[0189] when When a drastic topology change is detected, to improve the sensitivity of link failure detection and repair, the following settings are configured:
[0190]
[0191] That is, reduce and reduce ;
[0192] when and When the topology is determined to be relatively stable but the load is high, in order to suppress control overhead and congestion caused by flooding, the following settings are configured:
[0193]
[0194] That is, reduce and And it can be set as follows:
[0195]
[0196] To reduce the number of search rounds for the extended cycle;
[0197] when and When the network is in a relatively steady state, the parameters remain unchanged.
[0198] The above rules can be extended or modified according to the service type and QoS priority without affecting the basic principles of this invention.
[0199] S35. Update AODV multiple parameters based on weights and budget and apply range constraints.
[0200] Define the total parameter tuning budget for this tuning cycle as follows: Then the first The update magnitude of each parameter is:
[0201]
[0202] And update the parameters:
[0203]
[0204] in This is a truncation function used to ensure that the updated parameters fall within the allowed range.
[0205] Preferably, to further suppress parameter jitter caused by short-term fluctuations, exponential smoothing is performed on the updated parameters:
[0206]
[0207] in This is a smoothing factor.
[0208] S36. Sending Protocol Parameters
[0209] The updated AODV parameter set P(t+ This is encapsulated in a control message and passed to other nodes within the cluster via the cluster head or routing node. This control message contains parameter update information and a period identifier, ensuring that each node adjusts its protocol according to the new parameter set in the next period.
[0210] Cluster head nodes construct broadcast protocol parameter update messages The field format can be defined as shown in Table 3.
[0211]
[0212] The above describes the adaptive parameter conditions of the present invention using AODV as an example. For other protocols, the same method can be used to adjust the adaptive parameters.
[0213] Compared to discretizing multiple protocol parameters and outputting them as a set of joint actions, this invention uses a method based on Dirichlet distribution. The weight vector is used as a joint adjustment strategy representation. The output dimension is linearly related to the number of adjustable parameters. The total parameter tuning budget is controlled by the normalization constraint of the weight vector. This ensures the ability of multi-parameter coordinated adjustment while reducing the inference and decision-making overhead in the high-dimensional action space and suppressing parameter oscillations.
Claims
1. A cooperative adaptive routing method applied to a communication network, defining that the nodes in the communication network have the same data forwarding capacity and all the nodes form a cluster, characterized in that, include: A protocol switching module, a threshold adaptive optimization module, and a parameter tuning module are deployed on the cluster head node. The protocol switching module adaptively switches between multiple protocols in each sensing period based on the communication sensing information of network topology changes and service load, and in combination with dynamic thresholds and elastic sensing periods. The switching control message is broadcast to the other nodes, and the other nodes execute the corresponding protocol switching strategy according to the received switching control message. Cluster head nodes can be configured to collect multi-dimensional QoS metrics in real time to quantify network performance, including packet delivery rate (PDR), average end-to-end latency (E2ED), control overhead (OH), and average throughput (TH), and perform statistics within a set evaluation window, such as the sensing period. The method for obtaining communication-aware information based on network topology changes and service load is as follows: Construct containing domain Before the start of each sensing cycle, the source node is established in the network by the currently running routing protocol. To the destination node Working path, source node Based on the local routing table and the path record information provided by the routing protocol, determine the node sequence on the working path s–d at time t. , For the first in the path Each node identifier , The path endpoint sequence number is used to construct a service path registration message carrying the source node identifier, destination node identifier, service flow identifier, and the node sequence. This message is reported to the cluster head node via the intra-cluster control channel. The cluster head node establishes a mapping relationship between service flows and working paths locally. Map to working path , Indicates the current service flow identifier; cluster head node uses Centered on each node, in the current network topology graph Perform a k-hop neighborhood search to obtain the corresponding set of k-hop reachable nodes. where k is a positive integer. The set of nodes in the current topology. Given the set of edges in the current topology; take the union of the neighborhood sets of all nodes to obtain the set of nodes in the source-destination relational subgraph. : ， Will be The set of edges formed by the physical links between nodes is denoted as . Thus, a subgraph containing the domain is obtained. ; right Internal nodes and links collect communication-aware and load-aware information: Nodes in Obtain your own geographical location information Current speed information At the same time, the nodes are synchronized via a clock synchronization mechanism. and all neighboring nodes The location and velocity information are time-aligned; then, the obtained location and velocity information are used to calculate the node values respectively. with neighboring nodes relative velocity between ,node and neighboring nodes relative distance between ,node and neighboring nodes Link quality between Based on changes in relative node speed and link quality, nodes Calculate the node variation factor of the network topology , used to quantify topological changes between nodes in a network; In each perception cycle ,node The current queue occupancy status can be read by calling the queue query interface provided by the MAC layer. ; The result and Encapsulated as a status reporting message, it is sent to the cluster head node. The cluster head node will then receive the message. and After summing and normalizing, we obtain the current network topology dynamic factor R and service load factor L: ， ， in, The current number of nodes, based on the threshold. Determine whether to include it in the handover decision. For network topology dynamic factor threshold, The business load factor threshold is specifically when... and When, the active protocol is selected as the target protocol, when and In this case, select the on-demand protocol as the target protocol; otherwise, leave it unchanged. The method for broadcasting the handover control message to the other nodes is as follows: the protocol handover module of the cluster head node constructs and broadcasts the protocol handover control message. , This includes the target protocol identifier and the perception cycle. Non-cluster head node receives Then, execute the protocol corresponding to the target protocol identifier, according to... Adjust the next sensing cycle; The threshold adaptive optimization module is used to optimize the threshold. , Perception cycle Optimization is performed, specifically based on the differential evolution algorithm to achieve adaptive threshold optimization, including: The threshold to be optimized , Perception cycle The parameter vector encoded as a differential evolution algorithm is denoted as: ， The range of values is limited to: ， in, , , , These are all set thresholds; Construct a fitness function to measure the fitness of a given parameter vector. Performance advantages and disadvantages of routing switching strategies: ， in, The weights are non-negative and satisfy the following conditions: ; The threshold adaptive optimization module randomly generates parameters within a preset parameter space. The candidate solutions serve as the initial population, with each individual corresponding to a different set of solutions. The value is denoted as: ， Each component is initialized as follows: ， ， ， in, Represents a uniformly random number in the interval [0,1]. The threshold adaptive optimization module substitutes the network performance metrics obtained by the cluster head nodes based on the elastic sensing period time window into the fitness function to calculate the initial fitness value of all individuals. ; Perform differential mutation operation: at the first In the iteration, for each target individual in the population The threshold adaptive optimization module randomly selects three distinct individuals from the population. ,and Construct the mutation vector: ， in, This is the differential variation scaling factor, used to control the variable asynchronous length; Perform cross operations for each target individual Its corresponding mutation vector Perform crossover operations to generate experimental individuals. ; for the parameter vector of the first One portion, Generate random numbers ;like or Then let Otherwise, let ;in, For crossover probability, The index is randomly selected from {1,2,3} to ensure that each experimental individual has at least one component from the mutation vector; Execution Selection and Iterative Update: The Threshold Adaptive Optimization Module in Subsequent Experimental Individuals The corresponding parameter vector is used to configure the protocol switching strategy, and network performance is statistically analyzed and its fitness value is calculated. Then, the optimal selection step is performed: ， Repeat the differential mutation, crossover, selection, and iterative update processes until the preset maximum number of iterations is reached. The iteration stops when the optimal fitness improvement is less than a threshold for several consecutive generations; after the iteration ends, the individual with the highest fitness value is selected from the final population. As the current optimization result, its parameter components are: ， in, , , These are the parameters obtained after optimization; The protocol parameter tuning module adaptively adjusts the internal parameters of the selected protocol. Specifically, it adopts a lightweight multi-parameter joint adjustment strategy generation mechanism based on Dirichlet distribution. Cluster head nodes in the network generate parameter adjustment weight vectors according to the network status. The adjustment direction is determined by rules. By judging topology changes and service load, it is determined whether to increase or decrease the key parameters of the protocol. The cluster head node's protocol parameter tuning module is based on a reinforcement learning algorithm. It generates parameter adjustment weight vectors according to the network state (network topology dynamic factors, service load factors, sensing cycle, and current network performance), defining m weight vectors. The output is generated based on the Dirichlet distribution, and the normalization constraint is satisfied to keep the total parameter tuning amplitude t under control. ， The adjustment direction is determined by rules. By judging topology changes and service load, it is determined whether to increase or decrease key protocol parameters, and a direction vector is defined: ， in, This indicates increasing the k-th parameter. This indicates decreasing the k-th parameter. This indicates that the parameters remain unchanged; Define the total tuning parameter budget as Then the first The update magnitude of each parameter for: ， And update the parameter size according to the direction: ， in, It is the magnitude of the k-th parameter at time t. It is the size of the kth parameter after the update. The updated parameter set is then distributed to the remaining nodes in the cluster or the local protocol stack so that it takes effect in the next cycle.
2. The cooperative adaptive routing method for communication networks according to claim 1, characterized in that, The various protocols mentioned are two protocols: AODV protocol and OLSR protocol. Correspondingly, when... and When the topology is determined to be relatively stable but the load is high, the active OLSR protocol is selected as the target protocol; when and When the topology changes drastically and the load is low, On-Demand AODV is selected as the target protocol.