A knowledge-driven method and system for task scheduling in a communication network

Abstract

The application discloses a kind of knowledge-driven communication network task scheduling method and system, specifically is: according to network topology parameter, historical task data, construct weighted directed graph, four-tuple knowledge graph and dynamic updating mechanism, obtain enhanced weight matrix and Markov state transition model;According to original task request, in turn carry out decomposition subtask, establish resource demand matrix, hierarchical constraint modeling, feature vector generation, obtain double-channel task feature vector;According to double-channel task feature vector, network state, design composite cost function, construct multi-objective optimization problem, obtain meta-operator distribution by double time scale solution, and carry out joint optimization;According to meta-operator distribution, top-down traversal network layer carries out dynamic resource allocation, obtains task scheduling path and resource allocation scheme.The application can dynamically adjust network resource configuration, optimize task scheduling, so as to meet the semantics and resource demand of different tasks.

Description

Technical Field

[0001] This invention relates to the field of network information system technology, and in particular to a knowledge-driven communication network task scheduling method and system. Background Technology

[0002] With the rapid development of Intelligent Information Networks (IINs), network service paradigms are gradually shifting from traditional bit-level reliable transmission to task-driven and semantically aware intelligent service models. The core of this evolution lies in the fact that the goal of communication systems is no longer limited to simply "transmitting data," but extends to higher-level task adaptation and semantic understanding. Specifically, networks no longer focus solely on the basic transmission task of "what to transmit," but need to answer more challenging and complex task requirements such as "why transmit," "for whom to transmit," and "how to optimally achieve the task objectives."

[0003] Against this backdrop, many typical application scenarios are beginning to place higher demands on communication networks. For example, scenarios such as real-time monitoring, intelligent reasoning, collaborative sensing, and autonomous collaboration of unmanned systems require networks to not only process data but also understand its semantics and dynamically adjust their resource allocation and scheduling strategies according to task requirements. In these applications, networks need to possess the ability to handle task awareness, semantic understanding, and task optimization under resource constraints. Therefore, network scheduling strategies must have embodied semantic awareness, dynamic adaptation, and processing capabilities under multi-dimensional resource constraints to efficiently support complex task scheduling and execution.

[0004] However, existing mainstream scheduling methods, such as shortest path algorithms based on physical topology, minimum queuing delay strategies, and load balancing, generally suffer from the inability to explicitly model the semantic features of the task. These methods rely primarily on static network topology and link states during the decision-making process, lacking a deep understanding of the task's semantics. Therefore, these traditional methods struggle to perform effective path selection and resource allocation in environments with complex semantic tasks, especially in scenarios with high semantic complexity and limited resources. This can easily lead to path selection mismatches and inefficient resource allocation, ultimately affecting the overall service quality of the system. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge-driven communication network task scheduling method and system that dynamically adjusts network resource configuration and optimizes task scheduling to meet the semantic and resource requirements of different tasks.

[0006] The technical solution for achieving the objective of this invention is: a knowledge-driven communication network task scheduling method, comprising the following steps:

[0007] Step 1: Intelligent communication graph modeling: Based on network topology parameters and historical task data, construct a weighted directed graph, a quadruple knowledge graph, and a dynamic update mechanism to obtain an enhanced weight matrix and a Markov state transition model, providing a semantic network state representation for scheduling.

[0008] Step 2, Task Modeling: Based on the original task request, decompose subtasks, establish resource requirement matrix, hierarchical constraint modeling, and feature vector generation in sequence to obtain dual-channel task feature vectors. Use the enhanced weight matrix and Markov state transition model obtained in Step 1 for task knowledge embedding.

[0009] Step 3: Scheduling strategy modeling: Based on the dual-channel task feature vector and network state, design a composite cost function, construct a multi-objective optimization problem, obtain the distribution of meta-operators by solving it on two time scales, and perform joint optimization;

[0010] Step 4: Scheduling and Execution: Based on the distribution of meta-operators, the network layers are traversed from top to bottom to perform dynamic resource allocation, thereby obtaining the task scheduling path and resource configuration scheme.

[0011] A knowledge-driven communication network task scheduling system is provided. This system implements the aforementioned knowledge-driven communication network task scheduling method. The system includes an intelligent communication graph modeling module, a task modeling module, a scheduling strategy modeling module, and a scheduling execution module, wherein:

[0012] The intelligent communication graph modeling module constructs a weighted directed graph, a quadruple knowledge graph, and a dynamic update mechanism based on network topology parameters and historical task data, thereby obtaining an enhanced weight matrix and a Markov state transition model, providing a semantic network state representation for scheduling.

[0013] The task modeling module, based on the original task request, sequentially decomposes subtasks, establishes a resource requirement matrix, performs hierarchical constraint modeling, and generates feature vectors to obtain dual-channel task feature vectors. The resulting enhanced weight matrix and Markov state transition model are then used for task knowledge embedding.

[0014] The scheduling strategy modeling module designs a composite cost function based on the dual-channel task feature vector and network state, constructs a multi-objective optimization problem, obtains the distribution of meta-operators through dual time scales, and performs joint optimization.

[0015] The scheduling and execution module dynamically allocates resources by traversing the network layers from top to bottom according to the distribution of meta-operators, thereby obtaining the task scheduling path and resource configuration scheme.

[0016] A mobile terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the knowledge-driven communication network task scheduling method described above.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the knowledge-driven communication network task scheduling method.

[0018] Compared with the prior art, the significant advantages of this invention are:

[0019] (1) The semantic labels of tasks are explicitly introduced into the scheduling decision process. Through knowledge-driven methods, the network is guided from top to bottom to generate semantically consistent adaptive responses in multiple dimensions such as topology selection, path allocation and resource regulation.

[0020] (2) By organically combining the semantic information of the task, network topology, resource status and knowledge graph information, the network resource configuration can be dynamically adjusted and the task scheduling optimized, thereby meeting the semantic and resource requirements of different tasks.

[0021] (3) A task-aware knowledge-driven scheduling method is adopted, which constructs an intelligent communication graph to uniformly represent network topology, resource status and knowledge information, thereby realizing the deep integration of semantic intent and network scheduling;

[0022] (4) The dynamic knowledge evolution mechanism successfully solves the problem of adaptability to changes in semantic environment, providing a new paradigm for ultra-reliable low-latency communication in networks. Attached Figure Description

[0023] Figure 1 This is a flowchart of a knowledge-driven communication network task scheduling method proposed in this invention.

[0024] Figure 2 This is the topology diagram of the semantic matching sensitivity verification experiment.

[0025] Figure 3a It is semantic matching sensitivity verification and knowledge Figure 1 Consistency analysis diagram.

[0026] Figure 3b This is a graph showing the semantic matching sensitivity verification and average task completion delay analysis.

[0027] Figure 3c This is a graph showing the success rate analysis of semantic matching sensitivity verification and task scheduling.

[0028] Figure 4 It is the topology diagram used in the resource competition experiment.

[0029] Figure 5 This is a chart comparing the success rates of resource competition.

[0030] Figure 6 This is a comparison chart of average queuing wait times.

[0031] Figure 7 This is a comparison chart of queuing delay distribution characteristics.

[0032] Figure 8a This is a graph showing the trend of path change rate with knowledge update frequency under different semantic complexities.

[0033] Figure 8b This is a comparison chart of the variance of scheduling success rates under two knowledge graph models.

[0034] Figure 9 This is a graph analyzing the convergence speed of a knowledge graph.

[0035] Figure 10 This is a KG-Score thermodynamic analysis chart.

[0036] Figure 11 It is a load-blocking rate relationship curve.

[0037] Figure 12 This is a latency distribution map of the burst traffic.

[0038] Figure 13 This is a trend chart of knowledge overhead and semantic preservation.

[0039] Figure 14 It is a heatmap of performance under load levels. Detailed Implementation

[0040] This invention provides a knowledge-driven task scheduling method for communication networks. The core idea of ​​this method is to explicitly introduce the semantic tags of tasks into the scheduling decision-making process. Through a knowledge-driven approach, it guides the network from top to bottom to generate semantically consistent adaptive responses across multiple dimensions, including topology selection, path allocation, and resource regulation. Specifically, this invention constructs an intelligent communication graph, in which tasks are injected into the network carrying semantic intent, triggering the knowledge-driven scheduling module to automatically adapt policies. By organically combining the semantic information of tasks, network topology, resource status, and knowledge graph information, network resource configuration can be dynamically adjusted to optimize task scheduling, thereby meeting the semantic and resource requirements of different tasks.

[0041] Combination Figure 1 This invention discloses a knowledge-driven task scheduling method for communication networks, comprising the following steps:

[0042] Step 1: Intelligent communication graph modeling: Based on network topology parameters and historical task data, construct a weighted directed graph, a quadruple knowledge graph, and a dynamic update mechanism to obtain an enhanced weight matrix and a Markov state transition model, providing a semantic network state representation for scheduling.

[0043] Step 2, Task Modeling: Based on the original task request, decompose subtasks, establish resource requirement matrix, hierarchical constraint modeling, and feature vector generation in sequence to obtain dual-channel task feature vectors. Use the enhanced weight matrix and Markov state transition model obtained in Step 1 for task knowledge embedding.

[0044] Step 3: Scheduling strategy modeling: Based on the dual-channel task feature vector and network state, design a composite cost function, construct a multi-objective optimization problem, obtain the distribution of meta-operators by solving it on two time scales, and perform joint optimization;

[0045] Step 4: Scheduling and Execution: Based on the distribution of meta-operators, the network layers are traversed from top to bottom to perform dynamic resource allocation, thereby obtaining the task scheduling path and resource configuration scheme.

[0046] As a specific example, step 1 models the intelligent communication graph, as follows:

[0047] Step 1.1: Define the communication network as a weighted directed graph. ,in These represent functional entities in the network, including base stations, routing nodes, and computing units; each node represents a network functional component, and nodes are interconnected through directed edges. Description, representing the physical topology; directed edges are assigned weights, weight matrix Characterize the state features of the link, including bandwidth Delay ,load Parameters in multiple dimensions;

[0048] Step 1.2: Define the Intelligent Information Network (IIN) as a quadruple. The first three elements are the same as the basic communication network model defined in step 1.1, and the knowledge graph... Include:

[0049] Entity Collection This includes network entities, task types, and resource attributes. Network entities include nodes and links. This represents multidimensional data that embodies network functions and task resources;

[0050] Relation set Semantic relationships between different entities, including relationships of support, dependency, and conflict. Provide semantic constraints for task scheduling;

[0051] Attribute Collection : Describes the characteristics of the entity, including the node's computing power and the link's QoS level. It provides detailed attribute information for the operation of the network.

[0052] Step 1.3, the update rules for the knowledge graph are as follows: ,in express A knowledge graph of time. express A knowledge graph of time. express The weight matrix at time step, express A multi-domain observation semantic fact set at any given time; packaging the task logs, network telemetry, and semantically extracted triples or quadruples captured from the system in the current round, and using them as an increment function. The input evidence is used to make three types of changes: entity expansion, relation confidence correction, and attribute update; incremental function Include:

[0053] Entity expansion: When a new task type is detected At that time, the new task entity It will be added to the knowledge graph;

[0054] Relationship Correction: Update the confidence level of a relationship based on the scheduling success rate; Relationship confidence level After each successful scheduling, the changes are corrected according to the update formula: ,in This is an indicator function for whether the scheduling was successful. For the updated learning rate; express Confidence level of the time-relationship express Confidence level of the time-relationship;

[0055] Attribute Update: Adjust node computing power attributes through online learning. ;

[0056] Represents a node At any moment The effective computing power coefficient, Indicates at time Based on the effective computing power coefficient after one online update of the observed load, Represents a node Instantaneous or windowed utilization rate This represents the congestion threshold or design hyperparameter, which indicates that the system considers significant congestion to have occurred starting from the utilization rate, and that tasks assigned to nodes need to be significantly suppressed.

[0057] Step 1.4: Construct the cross-domain knowledge mapping function as follows Establish entity alignment relationships between the physical network and the knowledge graph to form an enhanced weight matrix. ; where knowledge embedding function A graph neural network is used to achieve semantic embedding of task and network features. The specific knowledge graph embedding function is as follows:

[0058]

[0059] Represents a set of nodes. Represents the set of edges in a network. This represents a set of edges or triples in a knowledge graph labeled with relation types. Represents the weight matrix. This indicates bit-level modulation or uniform scaling with broadcast. This represents a vectorized embedding of the knowledge graph. This indicates element-wise nonlinear activation or normalization. This represents the learnable weight matrix of relation type r. Indicates the type of relation in a knowledge graph. This represents the complete set of relation types in a knowledge graph. This represents an aggregation operator for a set of edges that satisfy a given relation. A vector representing an edge or triple in a knowledge graph. This represents an edge or triple instance in a knowledge graph. Representing an edge The relation type is , Representing nodes on a knowledge graph Attention scoring / edge representation, This represents a small-leakage linear rectified function with a negative slope. This represents the learnable attention vector in graph attention. , Represents node features Apply shared linear transformation The following indicates, Indicates splicing;

[0060] Step 1.5: The dynamic state of the IIN network is modeled using a Markov process:

[0061]

[0062] Among them, the core of transfer Reflects the propagation pattern of traffic congestion; This indicates the current network weight state. This represents the network weight matrix at the next time step. This represents the network weight matrix at the current time. Represents the set of network edges. Indicates link exist Load status, Indicates link exist Load status;

[0063] Step 1.6: Expand the state transition kernel as follows:

[0064]

[0065] This represents the conditional joint transition probability when a knowledge graph context is introduced. Representing a knowledge graph, This indicates that the knowledge graph is used in the link Context subgraph or features extracted from the neighborhood;

[0066] The adjustment strategy for the knowledge perception transfer core is as follows:

[0067]

[0068] Indicates link Mean shift of load changes under conflict-free or unobstructed conditions. Indicates link Mean drift in conflict / congestion scenarios This represents the variance of the transfer kernel. Indicates a Gaussian distribution. Indicates the node indices at both ends of the link;

[0069] Step 1.7: Implement conflict detection through knowledge graph relationship queries:

[0070]

[0071] Indicates a link or node pair There is a conflict. It means if and only if, Indicates a certain relation type, Represents the universal set of relations. In a knowledge graph, and They fall under conflict labels of relation r respectively;

[0072] Step 1.8: Introduce an M / M / 1 queue model in the link load dimension to simulate link load changes:

[0073]

[0074] in, Indicates time Time Link queue length, Indicates time Time Link queue length, Indicates link The maximum load threshold, and These refer to the arrival rate and service rate of the link, respectively.

[0075] The dynamic adjustment of service rate is performed using the following formula:

[0076]

[0077] in, This represents the effective service rate after semantic adaptation. This represents the baseline service rate of the link. Indicates link Knowledge context embedding, This represents the embedding vector of the knowledge graph for the current task. Describing the vector norm, Indicates the maximum distance allowed in the embedding space;

[0078] Step 1.9, when The network state evolution process then satisfies the Markov property; Represents the mathematical expectation. Indicates link Arrival rate, Indicates the state in the knowledge graph If the link service rate and system stability meet the following conditions... If the queue length maintains a finite expected value, then when the scheduled tasks tend to be uniformly distributed, the knowledge graph update satisfies... .

[0079] As a specific example, step 2 establishes a task model, as follows:

[0080] Step 2.1: Establish a task decomposition structure model and define the semantic task flow as a seven-tuple. Subtask set Each subtask With input parameters computational complexity Input parameter quantity ; Denotes the set of positive real numbers. Represents the set of natural numbers; This represents the set of task resource requirements. Represents a set of constraints. Represents the set of semantic attributes on the task side. A knowledge graph representing a task. This represents the set of objective functions for the task;

[0081] Dependency graph , This indicates an execution order constraint that satisfies: ,in To reduce cross-task data transmission latency, Indicates the start time of the subtask. Indicates the completion time of the subtask;

[0082] Step 2.2: Establish a multi-dimensional resource demand model and define the resource matrix. To describe the requirements, among which and These represent the dimensions of hard needs and soft needs, respectively. Represents the task requirement matrix. This indicates the number of subtasks, specifically:

[0083] Hard requirements Computing resources Storage requirements Network bandwidth ; This refers to the CPU clock speed, and the storage requirement refers to the larger of the task's input and output parameters. This indicates the number of CPU cycles required for the subtask. Indicates computational complexity. Indicates CPU clock speed or equivalent processing speed. Indicates memory or storage requirements. Indicates bandwidth requirements. Indicates the amount of output data. This indicates the maximum end-to-end latency that the task can accept.

[0084] Soft needs Parallelism requirements Affinity constraints Indicates task and Deployment preferences;

[0085] Step 2.3: Establish a hierarchical constraint requirement model and construct a three-level constraint set. , Represents a set of hard constraints. Represents a set of soft constraints. Represents a set of semantic or knowledge consistency and contextual adaptability constraints; defines hard constraints as including capacity constraints. For nodes Delay constraints Reliability constraints ,in, Indicates the first Sub-tasks Indicates mapping to network nodes The set of subtasks Subtasks CPU requirements Represents a node CPU capacity limit Representing a path Directed links on, Indicates link Transmission delay, Indicates the maximum permissible end-to-end delay. Indicates link failure rate, The lower bound representing the probability of path success; soft constraints: cost constraints Energy consumption constraints ,in, Indicates link unit cost Indicates the budget ceiling. Represents the set of all subtasks. Subtasks Energy consumption expenditure, Indicates the upper limit of energy consumption; additional semantic constraints: knowledge consistency KG-Match ( , ) Context adaptability Wherein, KG-Match represents a consistency measure between task knowledge and path knowledge. Represents task knowledge, This represents a subgraph induced in the knowledge graph by path nodes. This represents the knowledge consistency threshold. Represents the knowledge-driven contextual feasible domain. Indicates the context state. This represents a set of candidate paths or a single path. Represents the empty set;

[0086] Step 2.4: Establish the knowledge-driven task representation model of IIN and construct a dual-channel task feature vector. , which includes display features , express Calculate the density of a 3D real vector space. Data stream mode Deadline margin ,in Indicates the calculated density. Indicates computational complexity. Indicates the amount of input data. This represents the ratio of output to input. Indicates the amount of output data. Indicates a time margin before the deadline. Indicates the task deadline. Indicates the current time. Represents the normalized baseline time constant; knowledge graph embedding Using the TransE algorithm to Embedding: ,in An embedded representation of the task entity. This represents the embedding features of the knowledge graph. Represents the edges of the knowledge graph involved in the task. This represents one of the edges;

[0087] Step 2.5: Use a Markov-modulated Poisson process to describe task arrival. The context state transition satisfies ;

[0088] Indicates time Task arrival rate Indicates the context state. Represents the context state space, Indicates time In state The probability, Representing state Poisson arrival rate under the following conditions Indicates the context state transition probability. Indicates the context Markov transfer kernel, This indicates the transition from the current state to the next state. Indicates time The context state, Indicates time The context state;

[0089] Step 2.6: Construct a task priority queue model and a priority vector. Service rate function Considering the semantic similarity of tasks (KG Sim), the service rate of tasks is adjusted through semantic matching embedded in knowledge graphs; the blocking probability is calculated using an M / M / 1 / K queue model. :

[0090]

[0091] in The system load rate; Indicates the basic service rate. This represents the gain coefficient of semantic similarity on service rate. Represents task knowledge, Represents the current system knowledge. Indicates the probability of blocking. Indicates the system load rate. Indicates arrival rate, Indicates service rate;

[0092] Step 2.7, the feasibility of task decomposition is: if and only if there exists a path. This ensures that the path satisfies all constraints. The task can only be executed when the time is right; resource competition conditions are determined by the preemptive relationship. Definition: If satisfying ; at the same time if , This represents the set of knowledge required for the task. If the system's current global knowledge graph is represented, then there are feasible scheduling paths, ensuring the semantic completeness of the task and guaranteeing that the task can be executed according to the expected semantic goals.

[0093] As a specific example, the scheduling strategy modeling in step 3 is as follows:

[0094] Step 3.1, Knowledge-driven scheduling strategy;

[0095] (3.1.1) The knowledge-driven scheduling strategy is formally defined as a parameterized function. Its function is to determine the network state based on the given network state. and task characteristics Output a distribution of meta-operators The formal definition is as follows:

[0096]

[0097] (3.1.2) The output is a distribution of meta-operators. Specifically, it is represented by each meta-operator. Scheduling probability:

[0098]

[0099] Denotes the set of meta-operators. Indicates the selection of the meta-operator The probability, The parameter is The value function, Indicates the first Individual operator, Indicates the first Individual operator, This represents the network state representation for knowledge enhancement. Representation of knowledge-enhanced task representations;

[0100] (3.1.3) Value Function Implemented by a neural network:

[0101]

[0102] in, For network topology embedding, For task feature vectors, For knowledge graph attention weights, this involves embedding this information into the neural network to capture the interrelationships between network state, task requirements, and semantic information. The parameter is neural networks, This represents a vector-level concatenation operation.

[0103] Step 3.2: Design the multidimensional cost function and construct the composite cost function. :

[0104]

[0105] Among them, the first item Second item These represent the cost of path latency and bandwidth, respectively. This represents the knowledge matching distance between the path and the task. This term serves as a "knowledge matching penalty" to ensure that the execution path of the task can meet its semantic requirements. It is the queue waiting time, which reflects the time cost of a task waiting for resources in the network; The weights representing latency costs The weight representing bandwidth cost Indicates link latency. Indicates link bandwidth. This represents the weight coefficient of the knowledge matching penalty term. Representing path knowledge, Represents task knowledge, Indicates the queue waiting cost weight. Indicates along the path Average waiting time;

[0106] By calculating the task path and the character itself The degree of matching between paths and tasks is used to measure the semantic consistency between them, and is defined as follows:

[0107] This indicates the degree of inconsistency between task knowledge and path knowledge.

[0108] Queue waiting time Calculated based on the M / G / 1 queue model, it represents the average time a task waits for resources in the network, specifically:

[0109] in For arrival rate, For the system load rate, It is the variance of service time. Service time distribution parameters:

[0110]

[0111] Indicates the baseline service rate. An embedding vector representing path knowledge. An embedding vector representing task knowledge. Represents the normalization constant for the maximum distance in the embedding space;

[0112] Step 3.3: Construct a constrained multi-objective optimization problem:

[0113] The specific optimization objectives are:

[0114]

[0115] in, This indicates the path to be optimized and the choice of meta-operator. Represents parameters for multi-objective trade-offs. This represents the expectation under the state distribution. Represents the combined cost of a given path;

[0116] Entropy term To ensure the exploratory nature of the strategy and avoid getting trapped in local optima, the entropy is calculated using the following formula:

[0117]

[0118] Indicates the first Individual operator, Represents the complete set of meta-operators. Indicates selection The probability of;

[0119] The following constraints also need to be met during the optimization process:

[0120] Resource capacity constraint: The computational resource requirements of each task shall not exceed the computing power of the node, specifically:

[0121] Indicates a subtask. This represents the complete set of subtasks. Subtasks The carrying path, Indicates link load. Indicates link Maximum capacity;

[0122] Knowledge consistency constraint: The matching degree between the task path and the task's knowledge graph needs to meet a certain threshold. ,Right now:

[0123]

[0124] Indicates task With path Knowledge matching degree Represents a single task instance. This indicates the path selected for it;

[0125] QoS constraints: Ensure that the bandwidth, latency, and other parameters of a task meet quality requirements during network transmission.

[0126] Indicates link load. Indicates the link load threshold. Indicates the link error rate. Indicates the error rate threshold. This indicates a logical AND operation.

[0127] Delay reliability constraint: The task completion delay must not exceed the predetermined delay reliability. :

[0128]

[0129] Represents a single task instance. This represents the set of tasks in the scheduling queue. This indicates the probability that the task is expected to be completed. This indicates an event whose completion time exceeds the deadline. This represents the upper limit of the expected probability;

[0130] Step 3.4: Construct a dynamic optimization framework using a dual-time-scale optimization architecture;

[0131] (3.4.1) Design a fast-scale (second-level) optimized architecture:

[0132] On a second-level timescale, based on the current network state and task requirements, the shortest path problem is solved to minimize task scheduling costs; the A* algorithm is used to accelerate the search, and a heuristic function is employed. To guide the search process, the heuristic function is defined as:

[0133]

[0134] in, Used to calculate the semantic similarity between the task path and the target; Indicates starting from the current node To the target node The subset of knowledge required for the task. Indicates from node To the target node The candidate paths provide a subset of knowledge.

[0135] (3.4.2) Design a slowly varying (minute-level) optimization architecture:

[0136] Update scheduling policy parameters on a minute-level timescale. The gradient is updated using the following formula:

[0137]

[0138] This represents the new value of the policy parameter on a minute-level scale. This represents the old value of the policy parameter on a minute-level scale. Indicates the learning rate. Indicates the parameter gradient operator, Indicating in strategy The expected loss below;

[0139] The gradient estimation is performed using the REINFORCE algorithm:

[0140]

[0141] Representation strategy, Represents the element operator, Indicates the state of the knowledge-enhancing network. This indicates the characteristics of knowledge-enhancing tasks;

[0142] (3.4.3) If and only if a path exists The optimization problem is feasible when all constraints are satisfied, meeting the feasibility criterion; within the finite MDP framework, the policy gradient update guarantees convergence to a local optimum, satisfying policy convergence; if If the queue length remains finite, system stability is satisfied; when task distribution is stable, knowledge graph updates satisfy:

[0143]

[0144] and This represents two adjacent knowledge graphs. The symbol for asymptotically upper bound is _____. This represents the cumulative iterations / sample length.

[0145] This invention also provides a knowledge-driven communication network task scheduling system, which implements the aforementioned knowledge-driven communication network task scheduling method. The system includes an intelligent communication graph modeling module, a task modeling module, a scheduling strategy modeling module, and a scheduling execution module, wherein:

[0146] The intelligent communication graph modeling module constructs a weighted directed graph, a quadruple knowledge graph, and a dynamic update mechanism based on network topology parameters and historical task data, thereby obtaining an enhanced weight matrix and a Markov state transition model, providing a semantic network state representation for scheduling.

[0147] The task modeling module, based on the original task request, sequentially decomposes subtasks, establishes a resource requirement matrix, performs hierarchical constraint modeling, and generates feature vectors to obtain dual-channel task feature vectors. The resulting enhanced weight matrix and Markov state transition model are then used for task knowledge embedding.

[0148] The scheduling strategy modeling module designs a composite cost function based on the dual-channel task feature vector and network state, constructs a multi-objective optimization problem, obtains the distribution of meta-operators through dual time scales, and performs joint optimization.

[0149] The scheduling and execution module dynamically allocates resources by traversing the network layers from top to bottom according to the distribution of meta-operators, thereby obtaining the task scheduling path and resource configuration scheme.

[0150] The present invention also provides a mobile terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the knowledge-driven communication network task scheduling method described above.

[0151] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the knowledge-driven communication network task scheduling method described above.

[0152] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0153] Example

[0154] This embodiment provides a task-aware knowledge-driven scheduling (KDTS) method. The core idea of ​​this method is to explicitly introduce the semantic tags of tasks into the scheduling decision-making process. Through a knowledge-driven approach, it guides the network from top to bottom to generate semantically consistent adaptive responses across multiple dimensions, including topology selection, path allocation, and resource regulation. Specifically, this invention constructs an intelligent communication graph, in which tasks are injected into the network carrying semantic intent, triggering the knowledge-driven scheduling module to automatically adapt policies. By organically combining the semantic information of tasks, network topology, resource status, and knowledge graph information, KDTS can dynamically adjust network resource configuration and optimize task scheduling, thereby meeting the semantic and resource requirements of different tasks.

[0155] 1. Simulation platform and parameter settings

[0156] To verify the feasibility and effectiveness of the proposed knowledge graph-driven top-down scheduling strategy in intelligent communication networks, this invention constructs a lightweight task scheduling simulation platform that comprehensively models the communication network topology, knowledge graph evolution, task flow arrival, and multi-dimensional resource constraints. The scheduling logic and state evolution are primarily implemented using NetworkX, NumPy, and custom modules within the Python environment, exhibiting high reconfigurability and metric visualization capabilities.

[0157] (1) Regarding network topology and link model, the experiment uses a medium-sized network graph with heterogeneous link attributes. ,in:

[0158] Number of nodes This includes basic forwarding nodes and computing nodes;

[0159] Link by probability Randomly generated and assigned three-dimensional attributes: bandwidth Mbps, latency ms, initial load value ;

[0160] Link load dynamics are modeled using M / M / 1 queues, and service rates are adjusted according to the degree of knowledge matching.

[0161] (2) Regarding the initialization and evolution mechanism of knowledge graphs, knowledge graphs The initial build includes:

[0162] 20 entities (task type, network component, resource attribute);

[0163] 50 semantic relationships (support, dependency, conflict, etc.);

[0164] Constructing knowledge representations using TransE embeddings ;

[0165] After each round of task scheduling is completed, knowledge update rules are triggered, including confidence adjustment, attribute learning, and conflict correction.

[0166] (3) In terms of task flow generation and modeling, the tasks are modeled using MMPP (Markov Modulated Poisson Process), generating 10 to 30 tasks per round;

[0167] Each task is modeled as a 7-tuple. It can contain a maximum of 3 to 5 sub-tasks;

[0168] Subtask dependency graph Randomly generated, satisfying the directed acyclic graph constraint;

[0169] The tasks cover three typical scenarios: real-time video, object detection, and remote sensing image processing.

[0170] Each type of task has different knowledge segments. Simulate semantic-driven differences.

[0171] This invention implements four scheduling strategies, as shown in Table 1:

[0172] Table 1 Scheduling Strategies

[0173]

[0174] 2. Comparison Algorithm Design

[0175] To systematically evaluate the performance advantages of the proposed scheduling strategy, this invention designs three control group scheduling strategies, each representing a different scheduling paradigm. These are:

[0176] (1) Shortest-Delay (Shortest delay first, scheduling based on deep reinforcement learning) [28,29] In this invention, a deep reinforcement learning (DRL) approach was chosen as a common scheduling strategy in network scheduling. The DRL method trains a model to learn the dynamic changes in the network environment and automatically optimizes scheduling decisions. Under the shortest latency first (SLTF) strategy, the scheduling decision depends on the total link latency from the source node to the target node, selecting the path with the minimum latency for task scheduling. The specific path cost function is as follows:

[0177] This method represents the strategy that prioritizes transmission efficiency in traditional communication path scheduling. It optimizes task paths by ignoring the semantic characteristics of tasks and focusing solely on latency. However, because it does not consider the semantic or knowledge attributes of tasks, it is suitable for scheduling low-complexity tasks, but performs poorly in scenarios requiring deep semantic understanding and task adaptation.

[0178] (2) QoS-Aware (QoS-aware joint scheduling, resource scheduling based on game theory): This strategy comprehensively considers link bandwidth, latency, and current load, combined with weighted comprehensive indicators, to meet the quality requirements of the task. Its goal is to find the optimal path while maximizing resource utilization while meeting task requirements. The specific path cost is...

[0179]

[0180] in The weight parameters are selected based on experience in real-world network environments, and the scheduling objective is to find the globally optimal solution for network paths under the QoS dimension. This method can effectively adapt to various network conditions such as bandwidth, latency, and load, but it still has limitations—it lacks awareness of task semantics.

[0181] Although it can dynamically adjust based on the state of network resources, it is less effective than scheduling strategies specifically designed for task semantics in scheduling environments with high task complexity or requiring deeper semantic understanding.

[0182] (3) Random (random selection of feasible paths, scheduling based on genetic algorithm): In this strategy, a path is randomly selected from the set of all paths that meet resource constraints during task scheduling. This method simulates the performance of a system without a scheduling optimization mechanism, serving as a lower bound control group. The randomly selected path does not consider network topology, link state, or semantic information of the task, and therefore cannot effectively optimize resource utilization and task scheduling efficiency.

[0183] In contrast, Genetic Algorithms (GA) are commonly used to optimize path selection. GA simulates the process of species evolution, generating new candidate paths through operations such as crossover and mutation, and evaluating the merits of paths using a fitness function to optimize path selection and meet the resource constraints of the task. Although genetic algorithms can perform global optimization and improve scheduling performance, the random strategy is essentially a random selection and lacks path optimization capabilities. Therefore, its scheduling performance is not as good as optimization strategies based on network state or task semantic awareness.

[0184] (4) Proposed-KDTs (This method: Knowledge-driven semantic scheduling) This strategy takes the semantic features and context of the task as the core driving factors, calculates the semantic similarity between the path and the task based on knowledge graph embedding, and selects the path with the minimum composite cost function under multi-dimensional constraints:

[0185]

[0186] KG-Dist measures the degree of knowledge matching of entities along a path. This strategy represents the expectation of waiting in the path queue. It leverages the advantages of semantic-driven approaches, adaptive resource scheduling, and integration with knowledge graph reasoning.

[0187] 3. Definition of Evaluation Indicators

[0188] To comprehensively evaluate the performance of the proposed scheduling strategy in terms of semantic adaptability, resource scheduling efficiency, and system stability, this invention selects the following performance indicators across multiple dimensions for comparative analysis:

[0189] The Task Scheduling Success Rate is defined as the ratio of the number of tasks that are successfully assigned scheduling paths and have completed resource matching within a given time window to the total number of tasks.

[0190]

[0191] Average Task Completion Delay refers to the average time it takes for a task to complete from its arrival at the scheduler until the last subtask is deployed. It includes queuing time, path transmission delay, and inter-task dependency delay.

[0192]

[0193] The Knowledge Consistency Score (KG Score) measures the knowledge graph of the task objective. With the knowledge graph carried by the selected path entity Degree of matching between:

[0194]

[0195] A higher score indicates that the scheduling path is more semantically aligned with task requirements, making it a core indicator of the effectiveness of semantic scheduling.

[0196] Average Link Load represents the percentage of average load (relative to the maximum load threshold) on each link after task scheduling.

[0197]

[0198] Average Queue Waiting Time is the average waiting time of a task in the link, calculated under the M / M / 1 or M / G / 1 queue model.

[0199]

[0200] in This reflects the scheduling strategy's ability to respond to network congestion sensitivity.

[0201] 4. Experimental Results and Analysis

[0202] Based on the aforementioned typical scenarios, this section analyzes the performance of the proposed KDTS and three comparative strategies (Shortest-Delay, QoS-Aware, and Random) across multiple dimensions, verifying its advantages in semantic awareness, scheduling efficiency, and system stability.

[0203] 4.1 Semantic Matching Sensitivity Verification

[0204] This invention designs three performance metrics for tasks with different semantic complexity (low, medium, high): knowledge Figure 1 Consistency, scheduling latency, and success rate were evaluated, and comparative experiments were conducted with three typical baseline methods. Topology as follows: Figure 2 As shown, the types of tasks in the experiments included, but were not limited to, real-time video transmission, object detection, and remote sensing image processing. Each experiment group randomly generated 10–30 tasks and repeated them in 10 trials. The average value of each metric was calculated to eliminate random errors and improve the reliability of the experimental results. Finally, using these experimental data, the performance differences between the proposed KDTS strategy and the baseline method on various metrics were compared and analyzed.

[0205] Figure 3a This demonstrates the knowledge of various scheduling strategies under different semantic complexities. Figure 1KG Consistency Score Comparison. The KDTS strategy performs best across all semantic complexities, especially in low-semantic tasks (0.94), medium-semantic tasks (0.89), and high-semantic tasks (0.81), where its consistency is still significantly better than other strategies. This indicates that KDTS can effectively understand the semantic information of the task and maintain high consistency in complex tasks. In contrast, the KG consistency of the Shortest-Delay strategy gradually decreases with increasing semantic complexity, reaching 0.52 in high-semantic tasks, significantly lower than KDTS, reflecting its reliance on network topology and resource status and lack of semantic understanding. The QoS-Aware strategy performs similarly to Shortest-Delay, with a consistency of 0.55 in high-semantic tasks, failing to effectively consider the semantic structure of the task, leading to a decline in scheduling performance. The Random strategy performs worst across all semantic complexities, especially in high-semantic tasks where its consistency is only 0.41, indicating that it selects paths completely randomly, failing to consider task semantics, resulting in a severe drop in consistency.

[0206] Figure 3b The average task completion delay of each scheduling strategy is compared under different semantic complexities. The KDTS strategy performs excellently in terms of delay control, especially for low (3.1ms) and medium (3.7ms) semantic tasks, where the scheduling delay is significantly lower than other strategies. Even for high semantic tasks, KDTS has a delay of 4.3ms, which, although slightly higher, is still significantly better than other strategies. This indicates that KDTS can accurately select the scheduling path by constraining the matching degree between the task semantic vector and the network link knowledge vector, thereby maintaining low latency.

[0207] In comparison, the Shortest-Delay strategy exhibits higher latency, especially under complex tasks: 4.2ms for low-semantic tasks, 5.8ms for medium-semantic tasks, and 7.2ms for high-semantic tasks. Although the Shortest-Delay strategy selects the shortest latency path, it does not consider task semantics and dynamic resource allocation, resulting in a gradual increase in latency as semantic complexity increases. The QoS-Aware strategy has similar latency to Shortest-Delay: 5.3ms for low-semantic tasks, 6.1ms for medium-semantic tasks, and 6.9ms for high-semantic tasks.

[0208] While QoS-Aware strategies consider factors such as bandwidth and latency, they fail to effectively account for the semantic requirements of tasks, leading to increased latency as task complexity grows. The Random strategy consistently exhibits the highest latency, reaching 7.8ms, particularly under high-semantic tasks. Due to its random path selection, the Random strategy cannot effectively control latency, especially under high-semantic tasks, further exacerbating scheduling delays.

[0209] Figure 3c This paper compares the task scheduling success rates of various scheduling strategies under different semantic complexities. The KDTS strategy performs excellently across all semantic complexities, achieving a success rate of 0.98 for low-semantic tasks, 0.96 for medium-semantic tasks, and 0.94 for high-semantic tasks. This demonstrates that KDTS can accurately schedule tasks and maintain a high success rate, especially showing stable performance under complex tasks. In contrast, the Shortest-Delay strategy has a success rate of 0.92 for low-semantic tasks, but the success rate gradually decreases with increasing semantic complexity, especially dropping to 0.62 for high-semantic tasks. This reflects that the strategy fails to effectively consider task semantics, leading to performance degradation. The QoS-Aware strategy is similar to Shortest-Delay, with a success rate of 0.87 for low-semantic tasks and 0.68 for high-semantic tasks. This proves that although network resources are considered, the lack of accurate understanding of task semantics prevents effective scheduling optimization. The Random strategy has the lowest success rate, at only 0.33 for high-semantic tasks, indicating that it selects paths completely randomly without considering task requirements and network resources, resulting in a significant decrease in success rate.

[0210] The KDTS strategy maintains a high success rate across all semantic complexities: 0.98 for low-semantic tasks, 0.96 for medium-semantic tasks, and 0.94 for high-semantic tasks. This indicates that its success rate is relatively stable across various task types, and remains high even in high-semantic tasks, effectively completing task scheduling. The success rates of the Shortest-Delay and QoS-Aware strategies gradually decrease with increasing semantic complexity. Particularly in high-semantic tasks, the success rates of Shortest-Delay (0.58) and QoS-Aware (0.62) are significantly lower than KDTS, indicating poor scheduling performance for these two strategies in high-complexity tasks. The Random strategy has a lower success rate, starting at 0.42 for low-semantic tasks and gradually decreasing to 0.33 for high-semantic tasks, demonstrating its inadequacy in scheduling, especially in complex tasks where the success rate drops significantly.

[0211] 4.2 Resource Competitiveness Verification

[0212] In this experiment, to evaluate the performance of different scheduling strategies under resource contention conditions, the link bandwidth settings in the network topology were first determined. This simulates varying degrees of bandwidth limitation. The experimental topology is as follows: Figure 4 As shown, the remaining resources of the computing nodes are not evenly distributed, and some nodes are set as "hot zones", that is, their remaining computing resources are low, to simulate the situation of computing bottleneck.

[0213] Task generation employs a Markov-modulated Poisson process, generating 10 to 30 tasks per round. Each task has different resource requirements, specifically bandwidth and computational power requirements. Experiments compare the task scheduling success rate and average queuing time of four scheduling strategies under resource contention environments. By simulating scenarios with varying bandwidth limitations and scarce computational resources, the experiments verify the performance of each strategy under high resource contention, particularly the resource allocation and scheduling efficiency of the KDTS strategy in task scheduling.

[0214] like Figure 5 As shown, different scheduling strategies exhibit significant differences in task success rates under resource contention environments. KDTS significantly outperforms other strategies in both success rate (mean 0.92, 95% confidence interval [0.90, 0.94]) and stability (highly concentrated data points). Its confidence intervals do not overlap with those of other strategies, indicating that the difference is statistically significant (p < 0.001).

[0215] In comparison, the rule-based Shortest-Delay and QoS-Aware strategies performed moderately well and exhibited partially overlapping confidence intervals (means 0.75 and 0.80, respectively), while the Random strategy had the lowest success rate (mean 0.55) and the greatest volatility (widest confidence interval). Notably, KDTS's high success rate and low dispersion validate its robustness in resource-constrained scenarios—dynamically matching task requirements with network states through a knowledge graph effectively avoids hotspot nodes and bandwidth bottlenecks. This result aligns with theoretical assumptions, demonstrating the necessity of knowledge-driven mechanisms in scheduling complex semantic tasks: compared to traditional strategies, KDTS not only improves the success rate by approximately 30% but also reduces task blocking risk through semantic-resource joint optimization. Furthermore, the suboptimal performance of the Shortest-Delay strategy (success rate 0.75) reveals the limitations of solely relying on path delay—in situations of high resource contention, scheduling lacking semantic awareness can easily lead to conflicts and resource waste. Experimental results further demonstrate that the dynamic updating capabilities of knowledge graphs (such as relation confidence adjustment and entity expansion) play a key role in maintaining a high success rate (>90%), providing an interpretable and verifiable decision framework for task scheduling in intelligent communication networks.

[0216] Figure 6The average queuing wait times and their 95% confidence intervals for four scheduling strategies are presented. KDTS demonstrates a significant advantage with an average latency of 3.4 seconds (95% CI: [3.2, 3.6]) and the narrowest error bar, indicating strong policy robustness. In contrast, the rule-based Shortest-Delay and QoS-Aware strategies have latencies of 4.2 seconds (95% CI: [4.0, 4.4]) and 5.0 seconds (95% CI: [4.8, 5.2]), respectively, while the Random strategy has the highest latency (7.8 seconds, 95% CI: [7.5, 8.1]). Statistical tests show that the differences between KDTS and other strategies all pass the p < 0.001 significance test, validating its efficiency advantage in resource-contention environments. This result is consistent with the theoretical model: KDTS dynamically matches task requirements with network states through a knowledge graph, reducing path conflicts and link congestion, while traditional strategies suffer from increased latency due to ignoring resource contention or overly conservative constraints.

[0217] Figure 7 The study further revealed the queuing delay distribution characteristics of each strategy. KDTS delay data was highly concentrated (median 3.4 seconds, interquartile range 0.6 seconds), with no significant outliers, demonstrating its stability. The delay distributions of the Shortest-Delay and QoS-Aware strategies were more dispersed (interquartile ranges of 1.2 seconds and 1.4 seconds, respectively), and some high-latency outliers existed, indicating that resource contention could lead to bursty delays. The Random strategy exhibited the most extreme delay distribution (median 7.8 seconds, interquartile range 2.3 seconds) and contained multiple outliers (up to 9.5 seconds), reflecting severe resource waste due to the lack of optimization logic. Box plots visually validated KDTS's significant advantages in reducing tail latency (95th percentile 3.8 seconds vs. Random's 9.1 seconds) and reducing volatility (standard deviation 0.8 seconds vs. Random's 1.5 seconds), providing a solution for task scheduling in intelligent communication networks that combines efficiency and stability.

[0218] Combination Figure 6 , Figure 7 As a result, KDTS outperformed other strategies in both average latency (3.4 seconds) and distribution stability (standard deviation 0.8 seconds), validating the effectiveness of the knowledge-driven mechanism. Traditional strategies (such as Shortest-Delay) suffer from congestion on hotspot links due to a lack of semantic awareness, while the high latency of the QoS-Aware strategy reflects its overly conservative resource reservation mechanism. The extreme performance of the Random strategy highlights the necessity of optimized scheduling.

[0219] Experimental data shows that KDTS achieves precise task-resource matching through dynamic updates of the knowledge graph (such as relation confidence adjustment and entity expansion), thereby maintaining low latency and high success rate in complex resource competition scenarios.

[0220] 4.3 Impact of Knowledge Graph Evolution

[0221] To systematically evaluate the adaptability of KDTS to the dynamic evolution of knowledge graphs, this experimental design focuses on simulating the incremental update process of knowledge graphs and its impact on scheduling performance. In the simulation environment, new task types, entities, and relations are periodically introduced (e.g., a new semantic task is added every 10 rounds). By comparing the performance of KDTS under two modes—fixed knowledge graph (static initial state) and dynamic knowledge graph (real-time update mechanism)—the path selection change rate (defined as the ratio of the difference between dynamic and static paths) and scheduling stability (e.g., the variance of task success rate) are analyzed. Simultaneously, the convergence speed of the knowledge graph is observed (measured when the L2 norm change of the knowledge embedding vector falls below a threshold). The impact of the required number of rounds (as a metric) on core performance indicators, including task completion latency and path knowledge consistency score (KG-Score), is used to verify the robustness of KDTS in scenarios with dynamically changing semantic environments. The experiment controls the knowledge update frequency (e.g., learning rate) to... =0.05~0.2) and the semantic complexity gradient of task flow (low / medium / high), comprehensively quantifying the impact mechanism of knowledge evolution on system efficiency.

[0222] Figure 8a This demonstrates the path change rate under different semantic complexities as a function of knowledge update frequency. The changing trend of the path in the Dynamic Knowledge Graph (Dynamic-KG). In low-semantic scenarios, the path change rate is... When =0.05, it is 0.21, and as... The path change rate increased to 0.2, and decreased to 0.16 (p < 0.05), indicating that high-frequency updates reduced path oscillations. In contrast, the path change rate of static knowledge graphs (Static-KG) was fixed at 0.19–0.23, failing to adapt to semantic changes introduced by new task types. In high-semantic scenarios, the path change rate of Dynamic-KG further decreased to 0.15 (p < 0.05). =0.2), verifying its adaptive capability in complex environments. This result is consistent with the theoretical model: dynamic knowledge graphs achieve precise task-resource matching through incremental updates (such as entity expansion and relation confidence correction), while static graphs suffer from rigid path selection due to the lack of evolutionary mechanisms.

[0223] Figure 8bThe variance of scheduling success rate was compared between the two knowledge graph models. The variance of static knowledge graphs increased significantly with semantic complexity (from 0.10 to 0.20), especially reaching 0.20 in high-semantic scenarios, reflecting its insufficient adaptability. Conversely, the variance of dynamic knowledge graphs remained consistently below 0.15. The variance was only 0.08 at 0.2, and remained stable even in high-semantic scenarios (p < 0.01). This result indicates that dynamic knowledge graphs mitigate scheduling fluctuations caused by semantic uncertainty through real-time updates, while static graphs, due to their fixed knowledge representation, cannot cope with new task types, leading to a 3-fold increase in success rate variance (0.20 vs. 0.07). This finding validates the crucial role of knowledge evolution mechanisms in system robustness.

[0224] Figure 9 The convergence speed of dynamic knowledge graphs was quantified (measured by L2 norm change being less than 100%). The required number of rounds is the metric. (With update frequency...) As the value increased from 0.05 to 0.2, the number of convergence rounds decreased from 31 to 26 (Δ=5 rounds, p < 0.05), showing a clear negative correlation. When the value is 0.1, the convergence speed reaches its peak (32 rounds), and then decreases. The frequency continues to increase and then gradually decreases. This result indicates that a moderate update frequency (η=0.1–0.2) can accelerate knowledge graph convergence, but an excessively high frequency... This may lead to overfitting. Experimental data further validated the theoretical assumptions: the rapid convergence of the knowledge graph (<30 rounds) can effectively shorten the system response time and improve scheduling efficiency in dynamic environments.

[0225] Figure 10 This demonstrates how the KGScore (Path Knowledge Consistency Score) of a dynamic knowledge graph changes with update frequency in scenarios with high semantic complexity. When the KGS score is 0.2, the KGS score reaches 0.95, which is significantly higher than the KGS score of 0.2. =0.05 (0.86) and =0.1 (0.82) (p < 0.001). This result indicates that high-frequency updates can capture semantic changes more promptly, keeping the knowledge graph highly consistent with task requirements. For example, in When the value is 0.2, the knowledge graph completes 80% of the entity relationship updates within 10 rounds, while =0.05 requires 20 rounds to achieve the same effect. This finding verifies the direct impact of update frequency on the expressive power of knowledge graphs and provides empirical evidence for optimizing parameter settings.

[0226] Combination Figure 9 , Figure 10The results show that dynamic knowledge graphs significantly outperform static models in path stability (40% reduction in change rate), scheduling robustness (65% reduction in variance), and knowledge consistency (11% improvement in KGScore). Experimental data further demonstrate that update frequency... =0.2 achieves the optimal trade-off between convergence speed (26 rounds) and performance metrics, providing a configurable optimization strategy for knowledge-driven scheduling in intelligent communication networks.

[0227] 4.4 Large-scale task load stress test

[0228] To verify the scalability and stability of KDTS under large-scale, high-load scenarios, this experiment constructed a rigorous task stress test environment: a hybrid task stream was continuously injected through a Markov-modulated Poisson process, with 70% of the time remaining in a stationary state. =25 tasks / round), 30% of the time period introduces sudden traffic ( =75 tasks / round for 5 rounds), accumulating T=500 simulation rounds to simulate the high-frequency impact of a real network; simultaneously constructing a resource bottleneck environment—randomly selecting 30% of nodes and compressing their computing power to 50% of their initial value (triggering a dynamic resource adjustment mechanism), and limiting the bandwidth of 40% of links to 5-15Mbps (lower than the bandwidth requirements of regular tasks), thereby systematically monitoring three core indicators in a heterogeneous task flow with semantic complexity gradients (low / medium / high = 4:3:3): average path load rate, task blocking rate, and knowledge graph maintenance overhead (the proportion of CPU time spent by the knowledge embedding function Emb(𝒦)). Through comparison with the load-blocking rate curves of baseline strategies such as Shortest-Delay and QoS-Aware, and through stability analysis of knowledge overhead, the dual-timescale optimization framework is comprehensively verified. Maintaining scheduling efficiency (latency ≤ 5.2ms) and avoiding link overload (load variance) under extreme load conditions >0.85 Robustness capability of <0.05). Experimental reuse. Figure 4 Topology and setting link overload thresholds =0.85. At the burst flow trigger point (t∈{50,150,250,350,450}), we focused on observing the semantic preservation capability (KGScore≥0.75 for high semantic tasks) and real-time performance (knowledge maintenance overhead<15%) of KDTS, providing empirical support for high-load scheduling of 5G-A networks.

[0229] Figure 11 The curves reveal the differences in stability among various scheduling strategies under load pressure. When the average path load rate... When the value increases from 0.6 to 0.95, the traditional strategy exhibits exponential degradation: Shortest-Delay... After the value exceeds 0.65, the blocking rate increases sharply, and QoS-Aware... >0.8 and then broke through the 0.15 warning line; on the other hand, KDTS... It maintains a low blocking state with Pblock < 0.08 until the overload threshold of < 0.85, and even then... At a value of 0.9, the blocking rate (0.15) is still significantly lower than that of QoS-Aware (0.22). This is because the knowledge-driven value function embeds a bandwidth utilization efficiency factor into path selection. By combining the M / G / 1 queue optimization model, the system can still dynamically avoid hotspot links (such as...) even when ρ>0.85. Figure 4 The path around node n7 in the topology controls the overall network blocking rate, improving overload tolerance by 35.7% compared to the optimal baseline strategy.

[0230] Figure 12 The various strategies are clearly presented in response to sudden traffic surges. Latency characteristics under the impact of =75). The median latency of KDTS (4.1ms) is reduced by 40.6% and 42.3% compared with QoS-Aware (6.9ms) and Shortest-Delay (7.1ms), respectively, and the interquartile range (IQR=0.8ms) is only 1 / 3 of that of the traditional method, proving that the fast-scale A* search effectively reduces jitter; more importantly, all latency samples of KDTS are below the 5.2ms target line (while 27.5% of samples of the traditional method exceed the limit), which is attributed to the slow-scale policy gradient optimization continuously adjusting the knowledge matching weights on a minute-by-minute basis.

[0231] Experiments have shown that KDTS triggers the knowledge-aware transfer kernel during sudden events. The probability of (⋅) increases by 3.2 times, through dynamic avoidance of conflict relationships ( This ensures deterministic latency for high semantic tasks.

[0232] like Figure 13 As shown, the time sequence diagram reveals the dynamic balance between knowledge maintenance and semantic accuracy during 500 rounds of stress testing. KGOverhead initially increased to 12.8 ms due to incremental updates ΔK, but after optimization using LeakyReLU on the embedding function Emb(K), it stabilized at 11.2 ± 0.6 ms (accounting for 13.1% of the total overhead). Crucially, the five burst traffic events ( Under the impact of 50, 150, 250, 350, 450, KGScore only briefly dropped by 0.05. =At its lowest value of 0.80 (at 75), it recovers to 0.85+ within two rounds, demonstrating the robustness of the empirical knowledge consistency condition. Data shows that the satisfaction rate of the semantic constraint Csem is... Even at >0.9, the reliability rate still reaches 92.3%, providing a guarantee for time-sensitive scenarios such as the Industrial Internet of Things.

[0233] like Figure 14 As shown, this heatmap quantitatively reveals the multidimensional performance of KDTS under different load intensities. The vertical axis is divided into four load levels (low / medium / high / severe), and the horizontal axis covers four core indicators: blocking rate, latency, KGScore, and knowledge overhead.

[0234] Experimental data clearly shows that when the load increases from low (0.6-0.7) to severe (>0.9), KDTS still maintains excellent service quality—the blocking rate only gradually increases from 0.05 to 0.18, and the average latency remains within the critical value of 5.0ms (4.0ms→5.0ms), fully verifying the resilience of the dual-timescale optimization framework. Particularly noteworthy is that even under high load, the KGScore still reaches 0.75 (only a 18.5% decrease from the initial value), and the knowledge overhead remains stable at 14ms (accounting for 15.2% of the total latency). This data empirically demonstrates that the dynamic adjustment mechanism can maintain semantic matching accuracy under extreme loads, providing crucial support for ultra-reliable low-latency communication in the network.

[0235] In summary, this invention adopts a task-aware knowledge-driven scheduling method, which constructs an intelligent communication graph to uniformly represent network topology, resource status, and knowledge information, thereby achieving a deep integration of semantic intent and network scheduling. Through a dynamic knowledge evolution mechanism, it successfully solves the problem of adaptability to changes in the semantic environment, providing a new paradigm for ultra-reliable low-latency communication in networks and helping to explore the application expansion of knowledge graphs in cross-domain federated learning and integrated air-space-ground-sea networks.

Claims

1. A knowledge-driven task scheduling method for communication networks, characterized in that, Includes the following steps: Step 1: Intelligent communication graph modeling: Based on network topology parameters and historical task data, construct a weighted directed graph, a quadruple knowledge graph, and a dynamic update mechanism to obtain an enhanced weight matrix and a Markov state transition model, providing a semantic network state representation for scheduling. Step 2, Task Modeling: Based on the original task request, decompose subtasks, establish resource requirement matrix, hierarchical constraint modeling, and feature vector generation in sequence to obtain dual-channel task feature vectors. Use the enhanced weight matrix and Markov state transition model obtained in Step 1 for task knowledge embedding. Step 3: Scheduling strategy modeling: Based on the dual-channel task feature vector and network state, design a composite cost function, construct a multi-objective optimization problem, obtain the distribution of meta-operators by solving through dual time scales, and perform joint optimization; Step 4: Scheduling and Execution: Based on the distribution of meta-operators, the network layers are traversed from top to bottom to perform dynamic resource allocation, thereby obtaining the task scheduling path and resource configuration scheme.

2. The knowledge-driven communication network task scheduling method according to claim 1, characterized in that, Step 1 is described in detail as follows: Step 1.1: Define the communication network as a weighted directed graph. ,in These represent functional entities in the network, including base stations, routing nodes, and computing units; each node represents a network functional component, and nodes are interconnected through directed edges. Description, representing the physical topology; directed edges are assigned weights, weight matrix Characterize the state features of the link, including bandwidth Delay ,load Parameters in multiple dimensions; Step 1.2: Define the Intelligent Information Network (IIN) as a quadruple. The first three elements are the same as the basic communication network model defined in step 1.1, and the knowledge graph... Include: Entity Collection This includes network entities, task types, and resource attributes. Network entities include nodes and links. This represents multidimensional data that embodies network functions and task resources; Relation set Semantic relationships between different entities, including relationships of support, dependency, and conflict. Provide semantic constraints for task scheduling; Attribute Collection : Describes the characteristics of the entity, including the node's computing power and the link's QoS level. Provides detailed attribute information for network operation; Step 1.3, the update rules for the knowledge graph are as follows: ,in express A knowledge graph of time. express A knowledge graph of time. express The weight matrix at time step, express A multi-domain observation semantic fact set at any given time; packaging the task logs, network telemetry, and semantically extracted triples or quadruples captured from the system in the current round, and using them as an increment function. The input evidence is used to make three types of changes: entity expansion, relation confidence correction, and attribute update. Increment function Include: Entity expansion: When a new task type is detected At that time, the new task entity It will be added to the knowledge graph; Relationship Correction: Update the confidence level of a relationship based on the scheduling success rate; Relationship confidence level After each successful scheduling, the changes are corrected according to the update formula: ,in This is an indicator function for whether the scheduling was successful. For the updated learning rate; express Confidence level of the time-relationship. express Confidence level of the time-relationship; Attribute Update: Adjust node computing power attributes through online learning. ; Represents a node At any moment The effective computational capability coefficient, Indicates at time Based on the effective computing power coefficient after one online update of the observed load, Represents a node Instantaneous or windowed utilization rate Indicates the congestion threshold or design hyperparameter; Step 1.4: Construct the cross-domain knowledge mapping function as follows Establish entity alignment relationships between the physical network and the knowledge graph to form an enhanced weight matrix. ; where knowledge embedding function A graph neural network is used to achieve semantic embedding of task and network features. The specific knowledge graph embedding function is as follows: ； Represents a set of nodes. Represents the set of edges in a network. This represents a set of edges or triples in a knowledge graph labeled with relation types. Represents the weight matrix. This indicates bit-level modulation or uniform scaling with broadcast. This represents a vectorized embedding of the knowledge graph. This indicates element-wise nonlinear activation or normalization. This represents the learnable weight matrix of relation type r. Indicates the type of relation in a knowledge graph. This represents the complete set of relation types in a knowledge graph. This represents an aggregation operator for a set of edges that satisfy a given relation. A vector representing an edge or triple in a knowledge graph. This represents an edge or triple instance in a knowledge graph. Representing an edge The relation type is , Representing nodes on a knowledge graph Attention scoring / edge representation This represents a small-leakage linear rectified function with a negative slope. This represents the learnable attention vector in graph attention. , Represents node features Apply shared linear transformation The following indicates, Indicates splicing; Step 1.5: The dynamic state of the IIN network is modeled using a Markov process: ； Among them, the core of transfer Reflects the propagation pattern of traffic congestion; This indicates the current network weight state. This represents the network weight matrix at the next time step. This represents the network weight matrix at the current time. Represents the set of network edges. Indicates link exist Load status, Indicates link exist Load status; Step 1.6: Expand the state transition kernel as follows: ； This represents the conditional joint transition probability when a knowledge graph context is introduced. Representing a knowledge graph, This indicates that the knowledge graph is used in the link Context subgraph or features extracted from the neighborhood; The adjustment strategy for the knowledge perception transfer core is as follows: ； Indicates link Mean shift of load changes under conflict-free or unobstructed conditions. Indicates link Mean drift in conflict / congestion scenarios This represents the variance of the transfer kernel. Indicates a Gaussian distribution. Indicates the node indices at both ends of the link; Step 1.7: Implement conflict detection through knowledge graph relationship queries: ； Indicates a link or node pair There is a conflict. It means if and only if, Indicates a certain relation type, Represents the universal set of relations. In a knowledge graph, and They fall under conflict labels of relation r respectively; Step 1.8: Introduce an M / M / 1 queue model in the link load dimension to simulate link load changes: ； in, Indicates time Time Link queue length, Indicates time Time Link queue length, Indicates link The maximum load threshold, and These refer to the arrival rate and service rate of the link, respectively. The dynamic adjustment of service rate is performed using the following formula: ； in, This represents the effective service rate after semantic adaptation. This represents the baseline service rate of the link. Indicates link Knowledge context embedding, This represents the embedding vector of the knowledge graph for the current task. Describing the vector norm, Indicates the maximum distance allowed in the embedding space; Step 1.9, when The network state evolution process then satisfies the Markov property; Represents the mathematical expectation. Indicates link Arrival rate, Indicates the state in the knowledge graph If the link service rate and system stability meet the following conditions... If the queue length maintains a finite expected value, then when the scheduled tasks tend to be uniformly distributed, the knowledge graph update satisfies... .

3. The knowledge-driven communication network task scheduling method according to claim 1, characterized in that, Step 2 is described in detail below: Step 2.1: Establish a task decomposition structure model and define the semantic task flow as a seven-tuple. Subtask set Each subtask With input parameters computational complexity Input parameter quantity ; Denotes the set of positive real numbers. Represents the set of natural numbers; This represents the set of task resource requirements. Represents a set of constraints. Represents the set of semantic attributes on the task side. A knowledge graph representing a task. This represents the set of objective functions for the task; Dependency graph , This indicates an execution order constraint that satisfies: ,in To reduce cross-task data transmission latency, Indicates the start time of the subtask. Indicates the completion time of the subtask; Step 2.2: Establish a multi-dimensional resource demand model and define the resource matrix. To describe the requirements, among which and These represent the dimensions of hard needs and soft needs, respectively. Represents the task requirement matrix. This indicates the number of subtasks, specifically: Hard requirements Computing resources Storage requirements Network bandwidth ; This refers to the CPU clock speed, and the storage requirement refers to the larger of the task's input and output parameters. This indicates the number of CPU cycles required for the subtask. Indicates computational complexity. Indicates CPU clock speed or equivalent processing speed. Indicates memory or storage requirements. Indicates bandwidth requirements. Indicates the amount of output data. This indicates the maximum end-to-end latency that the task can accept. Soft needs Parallelism requirements Affinity constraints Indicates task and Deployment preferences; Step 2.3: Establish a hierarchical constraint requirement model and construct a three-level constraint set. , Represents a set of hard constraints. Represents a set of soft constraints. Represents a set of semantic or knowledge consistency and contextual adaptability constraints; defines hard constraints as including capacity constraints. For nodes Delay constraints Reliability constraints ,in, Indicates the first Sub-tasks Indicates mapping to network nodes The set of subtasks Subtasks CPU requirements Represents a node CPU capacity limit Representing a path Directed links on, Indicates link Transmission delay, Indicates the maximum permissible end-to-end delay. Indicates link failure rate, The lower bound representing the probability of path success; soft constraints: cost constraints Energy consumption constraints ,in, Indicates link unit cost Indicates the budget ceiling. Represents the set of all subtasks. Subtasks Energy consumption expenditure, Indicates the upper limit of energy consumption; additional semantic constraints: knowledge consistency KG-Match ( , ) Context adaptability Wherein, KG-Match represents a consistency measure between task knowledge and path knowledge. Represents task knowledge, This represents a subgraph induced in the knowledge graph by path nodes. This represents the knowledge consistency threshold. Represents the knowledge-driven contextual feasible domain. Indicates the context state. This represents a set of candidate paths or a single path. Represents the empty set; Step 2.4: Establish the knowledge-driven task representation model of IIN and construct a dual-channel task feature vector. , which includes display features , express Calculate the density of a 3D real vector space. Data stream mode Deadline margin ,in Indicates the calculated density. Indicates computational complexity. Indicates the amount of input data. This represents the ratio of output to input. Indicates the amount of output data. Indicates a time margin before the deadline. Indicates the task deadline. Indicates the current time. Represents the normalized baseline time constant; knowledge graph embedding Using the TransE algorithm to Embedding: ,in An embedded representation of the task entity. This represents the embedding features of the knowledge graph. Represents the edges of the knowledge graph involved in the task. This represents one of the edges; Step 2.5: Use a Markov-modulated Poisson process to describe task arrival. The context state transition satisfies ; Indicates time Task arrival rate Indicates the context state. Represents the context state space, Indicates time In state The probability, Representing state Poisson arrival rate under the following conditions Indicates the context state transition probability. Indicates the context Markov transfer kernel, This indicates the transition from the current state to the next state. Indicates time The context state, Indicates time The context state; Step 2.6: Construct a task priority queue model and a priority vector. Service rate function Considering the semantic similarity of tasks (KG Sim), the service rate of tasks is adjusted through semantic matching embedded in knowledge graphs; the blocking probability is calculated using an M / M / 1 / K queue model. : ； in The system load rate; Indicates the basic service rate. This represents the gain coefficient of semantic similarity on service rate. Represents task knowledge, Represents the current system knowledge. Indicates the probability of blocking. Indicates the system load rate. Indicates arrival rate, Indicates service rate; Step 2.7, the feasibility of task decomposition is: if and only if there exists a path. This ensures that the path satisfies all constraints. The task can only be executed when the time is right; resource competition conditions are determined by the preemptive relationship. Definition: If satisfying ; at the same time if , This represents the set of knowledge required for the task. If the system's current global knowledge graph is represented, then there are feasible scheduling paths, ensuring the semantic completeness of the task and guaranteeing that the task can be executed according to the expected semantic goals.

4. The knowledge-driven communication network task scheduling method according to claim 1, characterized in that, Step 3 is described in detail below: Step 3.1, Knowledge-driven scheduling strategy; (3.1.1) The knowledge-driven scheduling strategy is formally defined as a parameterized function. Its function is to determine the network state based on the given network state. and task characteristics Output a distribution of meta-operators The formal definition is as follows: ； (3.1.2) The output is a distribution of meta-operators. Specifically, it is represented by each meta-operator. Scheduling probability: ； Denotes the set of meta-operators. Indicates the selection of the meta-operator The probability, The parameter is The value function, Indicates the first Individual operator, Indicates the first Individual operator, This represents the network state representation for knowledge enhancement. Task representations that represent knowledge enhancement; (3.1.3) Value Function Implemented by a neural network: ； in, For network topology embedding, For task feature vectors, For knowledge graph attention weights, this involves embedding this information into the neural network to capture the interrelationships between network state, task requirements, and semantic information. The parameter is neural networks, This represents a vector-level concatenation operation; Step 3.2: Design the multidimensional cost function and construct the composite cost function. : ； Among them, the first item Second item These represent the cost of path latency and bandwidth, respectively. This represents the knowledge matching distance between the path and the task. This item serves as a "knowledge matching penalty" to ensure that the execution path of the task can meet its semantic requirements. It is the queue waiting time, which reflects the time cost of a task waiting for resources in the network; The weights representing latency costs The weight representing bandwidth cost Indicates link latency. Indicates link bandwidth. This represents the weight coefficient of the knowledge matching penalty term. Representing path knowledge, Represents task knowledge, Indicates the queue waiting cost weight. Indicates along the path Average waiting time; By calculating the task path and the character itself The degree of matching between paths and tasks is used to measure the semantic consistency between the paths and tasks, and is defined as follows: ； This indicates the degree of inconsistency between task knowledge and path knowledge. Queue waiting time Calculated based on the M / G / 1 queue model, it represents the average time a task waits for resources in the network, specifically: ； in For arrival rate, For the system load rate, It is the variance of service time. Service time distribution parameters: ； Indicates the baseline service rate. An embedding vector representing path knowledge. An embedding vector representing task knowledge. Represents the normalization constant for the maximum distance in the embedding space; Step 3.3: Construct a constrained multi-objective optimization problem: The specific optimization objectives are: ； in, This indicates the path to be optimized and the choice of meta-operator. Represents parameters for multi-objective trade-offs. This represents the expectation under the state distribution. Represents the combined cost of a given path; Entropy term To ensure the exploratory nature of the strategy and avoid getting trapped in local optima, the entropy is calculated using the following formula: ； Indicates the first Individual operator, Represents the complete set of meta-operators. Indicates selection The probability of; The following constraints also need to be met during the optimization process: Resource capacity constraint: The computational resource requirements of each task shall not exceed the computational capacity of the node, specifically: ； Indicates a subtask. This represents the complete set of subtasks. Subtasks The carrying path, Indicates link load. Indicates link Maximum capacity; Knowledge consistency constraint: The matching degree between the task path and the task's knowledge graph needs to meet a certain threshold. ,Right now: ； Indicates task With path Knowledge matching degree, Represents a single task instance. This indicates the path selected for it; QoS constraints: Ensure that the bandwidth, latency, and other parameters of a task meet quality requirements during network transmission. ； Indicates link load. Indicates the link load threshold. Indicates the link error rate. Indicates the error rate threshold. This indicates a logical AND operation. Delay reliability constraint: The task completion delay must not exceed the predetermined delay reliability. : ； Represents a single task instance. This represents the set of tasks in the scheduling queue. This indicates the probability that the task is expected to be completed. This indicates an event whose completion time exceeds the deadline. This represents the upper limit of the expected probability; Step 3.4: Construct a dynamic optimization framework using a dual-time-scale optimization architecture; (3.4.1) Design a rapidly changing, second-level optimized architecture: On a second-level timescale, based on the current network state and task requirements, the shortest path problem is solved to minimize task scheduling costs; the A* algorithm is used to accelerate the search, and a heuristic function is employed. To guide the search process, the heuristic function is defined as: ； in, Used to calculate the semantic similarity between the task path and the target; Indicates starting from the current node To the target node The subset of knowledge required for the task. Indicates from node To the target node The subset of knowledge that the candidate paths can provide; (3.4.2) Design a slowly varying, minute-level optimized architecture: Update scheduling policy parameters on a minute-level timescale. The gradient is updated using the following formula: ； This represents the new value of the policy parameter on a minute-level scale. This represents the old value of the policy parameter on a minute-level scale. Indicates the learning rate. Indicates the parameter gradient operator, Indicating in strategy The expected loss below; The gradient estimation is performed using the REINFORCE algorithm: ； Representation strategy, Represents the element operator, Indicates the state of the knowledge-enhancing network. This indicates the characteristics of knowledge-enhancing tasks; (3.4.3) If and only if a path exists The optimization problem is feasible when all constraints are satisfied, meeting the feasibility criterion; within the finite MDP framework, the policy gradient update guarantees convergence to a local optimum, satisfying policy convergence; if If the queue length remains finite, system stability is satisfied; when task distribution is stable, knowledge graph updates satisfy: ； and This represents two adjacent knowledge graphs. The symbol for asymptotically approaching the upper bound is _____. This represents the cumulative iterations / sample length.

5. A knowledge-driven communication network task scheduling system, characterized in that, This system is used to implement the knowledge-driven communication network task scheduling method according to any one of claims 1 to 4, the system comprising an intelligent communication graph modeling module, a task modeling module, a scheduling strategy modeling module, and a scheduling execution module, wherein: The intelligent communication graph modeling module constructs a weighted directed graph, a quadruple knowledge graph, and a dynamic update mechanism based on network topology parameters and historical task data, thereby obtaining an enhanced weight matrix and a Markov state transition model, providing a semantic network state representation for scheduling. The task modeling module, based on the original task request, sequentially decomposes subtasks, establishes a resource requirement matrix, performs hierarchical constraint modeling, and generates feature vectors to obtain dual-channel task feature vectors. The resulting enhanced weight matrix and Markov state transition model are then used for task knowledge embedding. The scheduling strategy modeling module designs a composite cost function based on the dual-channel task feature vector and network state, constructs a multi-objective optimization problem, obtains the distribution of meta-operators through dual-time scale solution, and performs joint optimization. The scheduling and execution module dynamically allocates resources by traversing the network layers from top to bottom according to the distribution of meta-operators, thereby obtaining the task scheduling path and resource configuration scheme.

6. A mobile terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the knowledge-driven communication network task scheduling method as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the knowledge-driven communication network task scheduling method as described in any one of claims 1 to 4.

Images