A multi-task real-time scheduling optimization method based on dependency-aware information age

Abstract

The application provides a multi-task real-time scheduling optimization method based on a dependency-aware information age, and provides a new optimization scheme in view of the deficiency of a traditional information age model in processing multi-task dependency and multi-source heterogeneity. By defining a dependency-aware information age model, the comprehensive influence of the freshness of task data and the delay of dependent tasks is quantified comprehensively, and the scheduling performance of the system in a complex industrial scene is improved significantly. The method combines the dependency relationship of the task chain, dynamically optimizes the scheduling strategy, processes the key path task preferentially, and adjusts the task priority and the resource allocation strategy in real time. In the framework of a discrete time linear time-invariant system, the global performance of the scheduling scheme is evaluated by combining a linear quadratic Gaussian control method. In addition, by using a multi-agent deep reinforcement learning framework, the task sampling frequency, the resource allocation and the transmission strategy are optimized, and the robustness and the global performance of the system are further improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial internet task scheduling and real-time optimization, and in particular, to a multi-task real-time scheduling optimization method based on dependency-aware information age. BACKGROUND

[0002] With the rapid development of industrial internet technology, industrial production gradually transforms from informatization to intelligentization. In this process, the demand for task scheduling and information processing becomes more complex, not only needing to adapt to various task types, but also needing to handle the dependency relationship between tasks and the integration demand of multi-source heterogeneous data. In modern industrial scenarios, the real-time performance of tasks and the collaborative scheduling between multiple tasks are of great significance to improve production efficiency and system stability.

[0003] Many current industrial scenarios have problems of uneven resource allocation, low overall resource utilization, and insufficient real-time performance in task scheduling. For example, different tasks have significant differences in demand for network bandwidth and computing resources, and the occupation of resources is in a competitive and shared relationship. The influence of multi-task dependency relationship and heterogeneous data sources on the overall performance of the system has not been fully quantified and optimized. Under the condition of complex multi-task dependency relationship and resource limitation, designing an efficient real-time scheduling optimization method is an important technical problem.

[0004] Traditional task scheduling methods usually evaluate the timeliness in units of single tasks, only focusing on the data update frequency or single completion time constraint of tasks, and are difficult to fully reflect the interaction between task dependency relationship and global performance. Information age is used to measure the timeliness of tasks, but in the multi-task dependency and multi-source heterogeneous scene, it is difficult to comprehensively consider the delay propagation effect in the task chain and the heterogeneous characteristics in multi-source data fusion, and the system performance optimization faces a bottleneck. SUMMARY

[0005] In view of the performance optimization problem in current industrial internet task scheduling caused by task dependency, multi-source heterogeneity and real-time demand, the present application proposes a multi-task real-time scheduling optimization method based on dependency-aware information age. By combining the update state of task data and the dependency relationship of task chain, a task timeliness evaluation index is constructed to realize the optimization of multi-task collaborative scheduling, and an efficient solution for complex scenarios in industrial internet is provided. By defining a dependency-aware information age model, combining the dependency relationship and data timeliness between multiple tasks, the task scheduling strategy and resource allocation scheme are optimized to improve the overall performance and real-time performance of the task chain. This method is suitable for complex scenarios such as industrial manufacturing, Internet of Things and edge computing that require high real-time performance and multi-task collaborative scheduling.

[0006] The technical scheme adopted by the present application to achieve the above-mentioned purpose is: a multi-task real-time scheduling optimization method based on dependent perception information age, comprising the following steps:

[0007] 1) dependent task scenario modeling: by analyzing the dependency relationship between tasks in the task chain, a dependent task scenario is established, and the attributes of each task are defined;

[0008] 2) dependent perception information age calculation: based on the attributes of the task, the dependent perception information age of the task is calculated to quantify the relationship between the data state of the current task and the completion time of its dependent task;

[0009] 3) scheduling strategy optimization: based on the dependent perception information age value of the task, the scheduling strategy of the multi-task is optimized, and the task priority is dynamically adjusted;

[0010] 4) control performance evaluation: using the discrete time linear time-invariant system framework, combined with the evaluation of the global performance of the scheduling strategy, the control cost equivalent function is obtained;

[0011] 5) multi-agent deep reinforcement learning optimization: according to the dependent perception information age of the task, the control cost equivalent function, and through the multi-agent deep reinforcement learning framework, the scheduling strategy of multiple task nodes is optimized.

[0012] In step 1), the attributes include the data acquisition time of the task, the dependent task set and the dependent task completion time.

[0013] In step 1), the dependent task scenario modeling includes at least one of the following modeling methods:

[0014] 1.1) series dependent modeling:

[0015] Based on the task chain, the series dependent relationship of the task is determined, and a series task dependency graph is generated; wherein the start time of the task satisfies: t start，i is the start time of task i, t end，j is the end time of dependent task j, D i represents the dependent task set of task i; the dependent task represents the prerequisite task that must be completed before the execution of a certain task;

[0016] 1.2) parallel dependent modeling:

[0017] Based on the task chain, the parallel dependent relationship of the task is determined, and a parallel task dependency graph is generated, and the multi-path dependent relationship of the task is labeled; the dependent task completion time is: t dep，i is the dependent completion time of task i;

[0018] 1.3) mixed dependent modeling:

[0019] determine the series and parallel dependency relationships of the tasks based on the task chain, and generate a hybrid dependency task chain graph; the hybrid dependency task chain graph comprises a series task dependency graph and a parallel task dependency graph.

[0020] In step 2), the dependency-aware information age is calculated, specifically as follows:

[0021] The dependency-aware information age of the task is calculated, and the value is the difference between the current time and the latest data collection time of the task and the latest completion time of the dependent task:

[0022]

[0023] wherein, is the dependency-aware information age, k is the current time, τ i (k) is the latest data collection time of the task i, D i is the dependent task set of the task i, C j (k) is the completion time of the dependent task j∈D i .

[0024] In step 3), the scheduling strategy optimization comprises the following steps:

[0025] Based on the dependency relationship of the task in the dependency task scenario modeling, the task priority is dynamically adjusted by using a topological sorting algorithm to identify the key tasks in the critical path that affect the global performance as the task nodes for priority scheduling.

[0026] wherein, the task priority is dynamically adjusted as follows:

[0027]

[0028] wherein, P i is the task priority, R i is the resource requirement of the task i, is the delay cost of the task i, and f represents a priority sorting function in the topological sorting algorithm, which is used to assign priority to the task according to the task attributes including the dependency-aware information age, the resource requirement, and the delay cost.

[0029] In step 4), the control performance evaluation comprises the following steps:

[0030] (1) For a system composed of terminal devices and edge servers, a discrete-time linear time-invariant system model is constructed to describe the state evolution of the task; the evolution relationship of the task state is as follows:

[0031] x i，k+1 = A i x i，k + B iu i，k +ω i，k

[0032] in It is a system task state vector used to represent the state of the terminal device. It is a control input vector generated by the edge server, used to represent the control quantity of the terminal device, and is a constant matrix. It is a system matrix. This is the input matrix for the edge server, i d i represents the state space dimension of the task. n This indicates the dimension of the control input vector space; It is zero-mean Gaussian random external interference noise;

[0033] (2) The linear quadratic Gaussian function is used to evaluate the equivalent function of the control cost of each task system, and the calculation is as follows:

[0034]

[0035] Among them W i，N and W i It is the system state weight matrix, U i It controls the input weight matrix. It is used to measure and control consumption. It is used to measure the state error of an actual system based on available measurements; What is used to express is the final error. This represents the expected value calculation; k represents the k-th time step, and N represents the total number of time steps;

[0036] (3) Based on the evolutionary relationship of task states and the control cost equivalence function, construct a control cost equivalence function based on age dependent on perceived information:

[0037]

[0038] Among them, e i,k To estimate the error in the control system, It is the system task state estimation vector. Age that relies on perceived information;

[0039]

[0040] Among them, J C，i To control the cost equivalence function, ∑ represents the expected value operation. i,k For e i，k covariance, S i Let Tr(·) be a solution to the Riccati equation, and let Tr(·) denote the trace operation of the matrix.

[0041] In step 5), the multi-agent deep reinforcement learning optimization comprises the following steps:

[0042] 5.1) State space definition: define the state space in multi-agent deep reinforcement learning

[0043]

[0044] where s i (k) represents the state of task i at the kth time step, Φ i (k) is the data size, C i，m，k (k) is the computing resource requirement, B i,m,k (k) is the bandwidth, h i，m，k (k) is the channel power gain, is the dependent task set of the task, is the subsequent task set of the task, is the dependency-aware information age;

[0045] 5.2) Action space design: construct the action space of multi-agent deep reinforcement learning to present the strategy of all tasks:

[0046]

[0047] where a i (k) represents the execution action, is the computing resource type matching decision, is the task division ratio, p i,m,k (k) is the transmission power of the sensor in the terminal device, f i，m，k (k) is the edge server computing resource allocation strategy, τ i，m，k (k) is the task sampling frequency;

[0048] 5.3) The reward function is:

[0049] where J C，i is the control cost, J N，i is the network energy consumption cost, ρ i is the dependency-aware information age reward item weight factor, Δ th is the threshold of the dependency-aware information age, α∈[0, 1] and β∈[0, 1] are corresponding weight factors, and α+β=1;

[0050] 5.4) Multi-agent collaborative optimization: according to the obtained edge server computing resource allocation strategy f i，m，k (k) and the task sampling frequency τ i，m，k(k), to achieve the scheduling strategy optimization of multiple task nodes.

[0051] A multi-task real-time scheduling optimization system based on dependency-aware information age, comprising:

[0052] A dependency task scenario construction module is configured to establish a dependency task scenario by analyzing the dependency relationship between tasks in a task chain, and define the attributes of each task.

[0053] A dependency-aware information age calculation module is configured to calculate the dependency-aware information age of a task based on the attributes of the task, to quantify the relationship between the data state of the current task and the completion time of its dependent task.

[0054] A scheduling strategy optimization module is configured to optimize the scheduling strategy of multiple tasks based on the dependency-aware information age value of the task, and dynamically adjust the task priority.

[0055] A control performance evaluation module is configured to obtain a control cost equivalent function by utilizing a discrete-time linear time-invariant system framework and combining the evaluation of the global performance of the scheduling strategy.

[0056] A multi-agent deep reinforcement learning optimization module is configured to optimize the scheduling strategy of multiple task nodes through a multi-agent deep reinforcement learning framework based on the dependency-aware information age of the task and the control cost equivalent function.

[0057] A computer-readable storage medium, the storage medium storing a computer program, when the computer program is executed by a processor, realizing the multi-task real-time scheduling optimization method based on dependency-aware information age.

[0058] The present application has the following advantages and benefits:

[0059] 1. The present application introduces a dependency-aware information age model to comprehensively quantify the data freshness and timeliness of the dependency relationship of the task, and optimizes the task priority and scheduling strategy. Compared with traditional methods, it can effectively reduce the influence of delay propagation in the task chain on the global performance, and significantly improve the real-time performance of complex task scenarios and the global optimization ability of the system.

[0060] 2. In a multi-source heterogeneous scenario, the present application dynamically adjusts the data source sampling frequency, transmission delay and resource allocation strategy to reduce the uncertainty of data synchronization and improve the efficiency and accuracy of data fusion tasks. At the same time, combined with the multi-agent deep reinforcement learning framework, the resource allocation and task scheduling scheme is optimized to improve the resource utilization efficiency.

[0061] 3. The present application is based on a discrete-time linear time-invariant system model and an error feedback mechanism, ensuring the stability and fault tolerance of task scheduling in a dynamically changing environment, suitable for complex scenarios such as industrial manufacturing, Internet of Things and edge computing, and has good scalability and application value. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 The method principle diagram of the present application is shown in the figure.

[0063] Figure 2 The overall system architecture considered by the present application is shown in the figure. DETAILED DESCRIPTION

[0064] The present application will be further described in detail below in conjunction with the drawings and examples

[0065] The present application provides a multi-task real-time scheduling optimization method based on dependency-aware information age, which provides a new optimization scheme for the shortcomings of traditional information age model in handling multi-task dependency and multi-source heterogeneity. By defining a dependency-aware information age model, the freshness of the task's own data and the comprehensive influence of the delay of dependent tasks are quantified, which significantly improves the scheduling performance of the system in complex industrial scenarios. This method combines the dependency relationship of the task chain, dynamically optimizes the scheduling strategy, and prioritizes the processing of critical path tasks, while adjusting the task priority and resource allocation strategy in real time. Under the framework of discrete-time linear time-invariant system, combined with linear quadratic Gaussian control method to evaluate the global performance of the scheduling scheme. In addition, using the multi-agent deep reinforcement learning framework, the task sampling frequency, resource allocation and transmission strategy are optimized, further improving the robustness and global performance of the system. The present application is suitable for complex scenarios such as industrial internet and edge computing, and meets the needs of high real-time, multi-source heterogeneous data fusion and task chain optimization.

[0066] For multi-dependent relationship process tasks executed by industrial control scene terminals, task modeling, parameter calculation and numerical analysis are performed in the edge server, and the multi-agent deep reinforcement learning framework is used to optimize the task sampling frequency, resource allocation and transmission strategy. Figure 1 The method principle diagram of the present application is shown in the figure, and the specific implementation process is as follows:

[0067] Step 1: dependency task scenario modeling:

[0068] For industrial control scenarios, based on the multi-dependent relationship process tasks executed by terminals, the following any way of dependency relationship modeling is constructed.

[0069] Step 1.1: series dependency modeling:

[0070] Modeling the serial relationship in the task chain, analyzing the relationship between the direct precedent task and the subsequent task of the task. In the modeling process, the completion time of the precedent task is assigned to each task, ensuring that the start time of the subsequent task is after the completion of the precedent task. By combing the order of tasks in the task chain, a clear structure of serial task dependency graph is generated, accurately reflecting the order dependency relationship in the task chain. Among them, the start time of the task satisfies: Where t start，i is the start time of task i, t end，j is the end time of dependent task j, and D i represents the dependent task set of task i.

[0071] Step 1.2: Parallel dependency modeling:

[0072] For the parallel relationship in the task chain, the multiple precedent tasks and their mutual relationship of task dependency are analyzed. For each task, the maximum value of the completion time of all precedent tasks is calculated as the dependent completion time of the task, that is: Where t dep，i is the dependent completion time of task i. Based on the in-depth analysis of the multi-path dependency between tasks, a parallel task dependency graph is generated, marking the parallel relationship and dependency path between tasks, ensuring the logical consistency of parallel tasks in the dependency chain.

[0073] Step 1.3: Mixed dependency modeling:

[0074] When dealing with mixed dependency task chains, the structural characteristics of series and parallel are considered, and the task chain model is constructed. By analyzing each task in the mixed dependency task chain, the dependency relationship between the task and the precedent task and the parallel task is clarified. Combined with the dependency relationship and the overall structure of the task chain, a complete mixed dependency task chain model is generated.

[0075] Step 2: Dependency-aware information age calculation:

[0076] Step 2.1: Definition of dependency-aware information age:

[0077] The dependency-aware information age of the task is calculated, which is the difference between the current time and the latest data collection time of the task and the latest completion time of its dependent tasks:

[0078]

[0079] Where k is the current time, τ i (k) is the latest data collection time of task i, D i is the dependent task set of task i, and C j (k) is the completion time of dependent task j∈D i .

[0080] Step 2.2: Global Latency Propagation Effect Quantification:

[0081] Evaluate the data freshness and timeliness of dependencies for the current task using the dependency-aware information age value. Measure the real-time deviation of the task chain by comparing the actual completion time with the ideal completion time, and dynamically adjust the task priority P and resource allocation strategy based on the evaluation results to ensure that the overall real-time requirements of the task chain are met.

[0082] Step 2.3: Multi-source Heterogeneous Data Analysis:

[0083] Based on the calculation results of the dependency-aware information age, analyze the delay characteristics and update frequency of different data sources, optimize the data fusion strategy, and dynamically adjust the sampling frequency and transmission delay of each data source.

[0084] Step 3: Scheduling Strategy Optimization:

[0085] Step 3.1: Critical Path Identification:

[0086] Based on the dependency relationship of the task chain, use the topological sorting algorithm to identify the key tasks in the critical path that affect the global performance. By analyzing the contribution of tasks in the task chain to the global delay, determine the task nodes in the critical path that need to be prioritized.

[0087] Step 3.2: Dynamic Adjustment of Task Priority:

[0088] According to the dependency-aware information age value and the global dependency relationship of the task chain, update the priority of the task in real time. Prioritize the completion of critical path tasks, reduce the propagation effect of global delay, and optimize the resource allocation strategy. The calculation formula of task priority P is:

[0089]

[0090] where R i is the resource requirement of task i, is the delay cost of task i. f represents the priority sorting function in the topological sorting algorithm, which adds the dependency-aware information age to the existing priority sorting function. It assigns priority to tasks based on task attributes (dependency-aware information age, resource requirement, delay cost).

[0091] Step 3.3: Multi-task Cooperative Optimization:

[0092] Consider the series and parallel relationship in the task chain, balance resource allocation, and improve the overall completion efficiency and system performance of the task chain by dynamically optimizing the scheduling order.

[0093] Step 4: Control Performance Evaluation:

[0094] Step 4.1: Control system modeling:

[0095] For the overall architecture of the system as shown in Figure 2 There are multiple sensor networks and multiple edge servers in the overall system. There are multiple sensor networks, multiple edge servers in the physical space. Multiple sensor networks will generate tasks with series and parallel dependent relationship, and these tasks will be distributed to different edge servers through scheduling optimization strategy to ensure the efficient operation of the overall system. Therefore, a discrete time linear time-invariant system model is constructed to describe the state evolution of the task. Combined with the dependency relationship and completion time constraint of the task, the change of the task state in the scheduling process is calculated. Each sensor network includes multiple terminal devices, and each edge server covers one or more sensor networks. For a system composed of multiple terminal devices (robots) and multiple edge servers, the evolution relationship of the task state is as follows:

[0096] x i，k+1 =A i x i,k +B i u i，k +ω i，k

[0097] wherein, is the system task state vector, is the control input vector generated by the edge control decision unit, the constant matrix is the system matrix, is the input matrix of the control unit. The random process is a zero-mean Gaussian random exogenous disturbance noise with covariance Ξ. For the actual robot control scene, x i，k is the position and speed of the robot, etc., u i，k is the steering angle and acceleration of the robot, etc. For the temperature control scene, x i，k is the temperature information of the system, and u i，k is the fan speed of the heat dissipation device and other information.

[0098] Step 4.2: Global performance index calculation:

[0099] Based on the global performance index, the scheduling scheme is evaluated, and the control cost of the system is calculated. A linear quadratic Gaussian function is used to evaluate the control cost of each task system, which is calculated as follows:

[0100]

[0101] wherein, W i，N and W i are the system state weight matrices, and U i is the control input weight matrix, is used to measure control consumption, is used to measure the state error of the actual system based on the available measurement value. is used to express the terminal error. represents the expected value operation, and N represents the total time step.

[0102] Step 4.3: Error feedback and optimal control analysis:

[0103] Error feedback and optimal control analysis are performed on the execution effect of the scheduling strategy, and the task priority and resource allocation strategy are adjusted in real time. Combined with error accumulation, propagation analysis and optimal control theory, the stability and timeliness of global scheduling are optimized to ensure the robustness and high efficiency of the system in a dynamic changing environment.

[0104] The system state evolution equation of the control system is:

[0105]

[0106] The estimated state equation of the control system is:

[0107]

[0108] The estimated error of the control system is:

[0109]

[0110] According to the definition of covariance: And combined with the optimal control law The control cost equivalent function can be obtained:

[0111] Where, ∑ i，k is the covariance of e i，k , L i is the gain matrix, S i is the solution of the Riccati equation, and Tr(·) represents the trace operation of the matrix, and r represents the rth time step.

[0112] The specific implementation process of the multi-agent algorithm in the application is as follows.

[0113] Step 5: Multi-agent deep reinforcement learning optimization:

[0114] Step 5.1: State space definition:

[0115] Define the state space in multi-agent deep reinforcement learning The state space describes the running state of the task and the edge computing and communication resources, and the state s i(k) is characterized by data size, computing resource requirement, bandwidth, channel power gain, dependency of tasks (predecessor task and successor task), dependency-aware information age, i.e.,

[0116]

[0117] where Φ i (k) is data size, C i,m,k (k) is computing resource requirement, B i,m,k (k) is bandwidth, h i，m，k (k) is channel power gain, is a set of predecessor tasks of the task, is a set of successor tasks of the task, is dependency-aware information age. Dependency task is for the current task, successor task represents the next task or multiple tasks for the current task.

[0118] Step 5.2: Action space design:

[0119] Action space design for multi-agent deep reinforcement learning Action space presents the strategy of all agents, at each decision-making period k, i-agent performs action a i (k) according to the whole state s i (k). Action space a i (k) is given by the action of computing resource type matching decision, task partition ratio, sensor system transmit power, edge server computing resource allocation strategy, task sampling frequency as follows:

[0120]

[0121] where, is computing resource type matching decision, is task partition ratio, p i,m,k (k) is sensor system transmit power, f i，m，k (k) is edge server computing resource allocation strategy, τ i，m，k (k) is task sampling frequency.

[0122] Step 5.3: Reward function design:

[0123] Reward function is: where J C，i is control cost, is J N，i network energy consumption cost, ρ i is dependency-aware information age reward term weight factor, Δ thwhere is the threshold of the age of perception information, and are the corresponding weight factors, and. The multi-task scheduling strategy is optimized based on the reward function, and the scheduling strategy is dynamically adjusted by evaluating the task completion rate, system energy consumption and the change of the age of perception information, to balance the timeliness of the task and the resource utilization.

[0124] Step 5.4: Multi-agent collaborative optimization:

[0125] The computing resource allocation strategy f of the edge server is obtained through the above steps i，m，k (k), terminal task sampling frequency τ i，m，k (k) and the like, the scheduling strategies of multiple task nodes are collaboratively optimized under a multi-agent deep reinforcement learning framework. Through information interaction among the agents, global resource allocation optimization of the task nodes is achieved, and the overall completion efficiency and system performance of the task chain are improved.

[0126] Finally, it should be finally noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.

Claims

1. A multi-task real-time scheduling optimization method based on age-dependent information, characterized in that, Includes the following steps: 1) Dependency task scenario modeling: By analyzing the dependency relationships between tasks in the task chain, a dependency task scenario is established, and the attributes of each task are defined; 2) Dependency-aware information age calculation: Based on the attributes of the task, calculate the age of the task's dependency-aware information to quantify the relationship between the data state of the current task and the completion time of its dependent tasks. 3) Scheduling strategy optimization: Based on the age value of task dependency awareness information, optimize the scheduling strategy for multiple tasks and dynamically adjust task priorities; 4) Control performance evaluation: Using the discrete-time linear time-invariant system framework and combining it with the global performance evaluation of the scheduling strategy, the equivalent function of control cost is obtained; 5) Multi-agent deep reinforcement learning optimization: Based on the task dependency perception information age and control cost equivalence function, the scheduling strategy of multiple task nodes is optimized through a multi-agent deep reinforcement learning framework. In step 2), the age calculation based on perceived information is as follows: The age of the task's dependent information is calculated as the difference between the current time and the minimum of the task's latest data acquisition time and the latest completion time of its dependent tasks. ； in, Age that relies on perceived information For the current time, For the task The latest data collection time, For the task The set of dependent tasks, For dependent tasks Completion time; Step 3), the scheduling strategy optimization, includes the following steps: Based on the task dependencies in the task dependency scenario modeling, the topology sorting algorithm is used to dynamically adjust the task priority in order to identify the critical tasks that affect the global performance in the critical path and schedule them as priority task nodes. The dynamic adjustment of task priorities is as follows: ； in, As a task priority, For the task resource requirements, For the task The delay cost, This represents the priority sorting function in the topology sorting algorithm, used to assign priorities to tasks based on task attributes including dependency awareness age, resource requirements, and latency costs.

2. The multi-task real-time scheduling optimization method based on age-aware information according to claim 1, characterized in that, In step 1), the attributes include the data acquisition time of the task, the set of dependent tasks, and the completion time of the dependent tasks.

3. The multi-task real-time scheduling optimization method based on age-aware information according to claim 1, characterized in that, In step 1), the dependent task scenario modeling includes at least one of the following modeling methods: 1.1) Series Dependency Modeling: Based on the task chain, the sequential dependencies of tasks are determined, and a sequential task dependency graph is generated; wherein the start time of the tasks satisfies: , For the task The start time, For dependent tasks End time, Indicates task A set of dependent tasks; the dependent task represents a preceding task that must be completed before a certain task can be executed; 1.2) Parallel dependency modeling: Based on the task chain, determine the parallel dependencies of tasks, generate a parallel task dependency graph, and mark the multi-path dependencies of tasks. The completion time of the dependent task is: ; For the task Dependency completion time; 1.3) Hybrid dependency modeling: Based on the task chain, the serial and parallel dependencies of tasks are determined, and a hybrid dependency task chain graph is generated; the hybrid dependency task chain graph includes a serial task dependency graph and a parallel task dependency graph.

4. The multi-task real-time scheduling optimization method based on age-aware information according to claim 1, characterized in that, Step 4), the control performance evaluation, includes the following steps: (1) For the system consisting of terminal devices and edge servers, a discrete-time linear time-invariant system model is constructed to describe the state evolution of the task; the evolution relationship of the task state is as follows: ； in It is a system task state vector used to represent the state of the terminal device. It is a control input vector generated by the edge server, used to represent the control quantity of the terminal device, and is a constant matrix. It is a system matrix. It is the input matrix of the edge server. This represents the state space dimension of the task. This indicates the dimension of the control input vector space; It is zero-mean Gaussian random external interference noise; (2) The linear quadratic Gaussian function is used to evaluate the equivalent function of control cost for each task system, and the calculation is as follows: ； in and It is the system state weight matrix. It controls the input weight matrix. It is used to measure and control consumption. It is used to measure the state error of an actual system based on available measurements; What is used to express is the final error. This represents the expected value operation; Indicates the first Each time step Indicates the total time step; (3) Based on the evolutionary relationship of task states and the control cost equivalence function, construct a control cost equivalence function based on age dependent on perceived information: ； in, To estimate the error in the control system, It is the system task state estimation vector. Age that relies on perceived information; ； in, To control the cost equivalence function, This represents the expected value operation. for covariance, , This is a solution to the Riccati equation. Represents the trace operation of a matrix.

5. The multi-task real-time scheduling optimization method based on age-aware information according to claim 1, characterized in that, Step 5), the multi-agent deep reinforcement learning optimization, includes the following steps: 5.1) Definition of State Space: Define the state space in multi-agent deep reinforcement learning. : ； in, Indicates the first Time step, task state, For data size, To meet computing resource requirements, For bandwidth, For channel power gain, For the set of dependent tasks of the task, This is a set of subsequent tasks for the given task. Age that relies on perceived information; 5.2) Action Space Design: Constructing the Action Space for Multi-Agent Deep Reinforcement Learning , to present a strategy for all tasks: ； in, Indicates the execution of an action. To calculate resource type matching decisions, Divide the tasks into proportions, This refers to the transmit power of the sensor in the terminal device. For edge server computing resource allocation strategies, The sampling frequency for the task; 5.3) The reward function is: ; in, To control costs, For network energy consumption costs, For age-related reward items that rely on perceived information, weighting factors A threshold for age based on perceived information. and For the corresponding weighting factor, and ; 5.4) Multi-agent cooperative optimization: Based on the obtained edge server computing resource allocation strategy Task sampling frequency This enables the optimization of scheduling strategies for multiple task nodes.

6. A multi-task real-time scheduling and optimization system based on age-dependent information, characterized in that, include: The dependency task scenario building module is used to establish dependency task scenarios by analyzing the dependencies between tasks in the task chain and to define the attributes of each task. The dependency-aware information age calculation module is used to calculate the dependency-aware information age of a task based on its attributes, in order to quantify the relationship between the current data state of the task and the completion time of its dependent tasks. The age calculation module that relies on perceived information is configured to execute: The age of the task's dependent information is calculated as the difference between the current time and the minimum of the task's latest data acquisition time and the latest completion time of its dependent tasks. ； in, Age that relies on perceived information For the current time, For the task The latest data collection time, For the task The set of dependent tasks, For dependent tasks Completion time; The scheduling strategy optimization module is used to optimize the scheduling strategy of multiple tasks based on the age value of task dependency awareness information and dynamically adjust task priorities. The scheduling strategy optimization module is configured to execute: Based on the task dependencies in the task dependency scenario modeling, the topology sorting algorithm is used to dynamically adjust the task priority in order to identify the critical tasks that affect the global performance in the critical path and schedule them as priority task nodes. The dynamic adjustment of task priorities is as follows: ； in, As a task priority, For the task resource requirements, For the task The delay cost, This represents the priority sorting function in the topology sorting algorithm, which is used to assign priorities to tasks based on task attributes including dependency-aware information age, resource requirements, and latency costs. The control performance evaluation module is used to obtain the control cost equivalent function by utilizing the discrete-time linear time-invariant system framework and combining the evaluation of the global performance of the scheduling strategy. The multi-agent deep reinforcement learning optimization module is used to optimize the scheduling strategy of multiple task nodes based on the task's dependency perception information age and control cost equivalence function through a multi-agent deep reinforcement learning framework.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements a multi-task real-time scheduling optimization method based on age-dependent information as described in any one of claims 1-5.

Images