Computer system service optimization scheduling method for photoelectric system

Abstract

The invention discloses a computer system service optimization scheduling method for a photoelectric system, and particularly relates to the technical field of computer service optimization, which comprises the following steps of: sensing and extracting information such as task types and time constraints through task characteristics, and constructing a multi-dimensional resource model such as processing capability and idle degree by combining resource modeling; a weighting mechanism is used for evaluating task priorities and generating a scheduling queue, tasks are dynamically mapped to matched resources, a scheduling strategy is adjusted according to system loads and feedback, model parameters are optimized through service feedback, and efficient, self-adaptive and intelligent optimization of task scheduling of the photoelectric system is achieved; according to the method, a dynamic scheduling mechanism based on task characteristics and resource prediction is realized, and the response speed of the high-optimal task is improved; optimizing the resource utilization rate by adopting resource state modeling and prospective mapping; and reinforcement learning and a feedback closed loop are introduced, a sustainable optimization scheduling system is constructed, and the adaptability and stability of the system in a complex photoelectric task environment are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of computer service optimization, and more particularly to a computer system service optimization scheduling method for a photoelectric system. Background Art

[0002] With the widespread application of modern optoelectronic technology in high-performance scenarios such as military reconnaissance, intelligent perception, satellite remote sensing, and security monitoring, the computational requirements for tasks such as image processing, signal fusion, and target recognition have increased significantly. These optoelectronic systems often integrate a large number of heterogeneous task units and multiple computing resources. These systems must achieve efficient task scheduling and resource allocation while maintaining strong real-time performance, complex task loads, and high communication volumes.

[0003] However, existing scheduling systems often rely on static rules, preset priorities, or simplified scheduling models, making them difficult to cope with the dynamic changes in tasks and the uncertainty of system resource status. For example, faced with a surge in tasks, the preemption of critical tasks, and frequent resource contention, traditional scheduling methods often struggle to respond quickly and appropriately, leading to task delays, low system resource utilization, and even the abandonment of high-priority tasks and fluctuating system performance.

[0004] Furthermore, existing methods generally lack the ability to coordinate optimization of task dependency structures and resource topologies, and are unable to comprehensively consider the communication density between tasks, the scheduling hierarchy, and the structural matching of computing resources. Furthermore, most methods are unable to optimize policies based on scheduling history feedback and lack continuous learning and adaptive capabilities. This results in a single scheduling strategy and limited optimization capabilities, making it difficult to meet the requirements of intelligent optoelectronic systems for highly reliable, sustainable, and high-performance scheduling mechanisms. Therefore, the present invention proposes a computer system service optimization scheduling method for optoelectronic systems to address these issues. Summary of the Invention

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The computer system service optimization scheduling method for an optoelectronic system comprises the following steps:

[0007] Task characteristic perception step, collecting key characteristic information of the tasks to be scheduled in the optoelectronic system, including task type, time constraints, data dependencies and priority levels;

[0008] Resource modeling steps,modeling the heterogeneous computing resources in the computer system,constructing a multi-dimensional resource description model including processing capacity,,idleness, power consumption characteristics and communication delay;

[0009] Priority evaluation step, based on task characteristics and resource status, a weighted evaluation mechanism is used to calculate the scheduling priority of each task to form a priority queue;

[0010] Dynamic mapping step, based on the task priority queue and resource modeling results, determines the computing resource allocation target of each task in real time, and completes the binding of tasks and resources through mapping algorithms;

[0011] Scheduling strategy adjustment steps: dynamically select the appropriate scheduling strategy based on the current computer system load status and task execution feedback, and adaptively adjust the task execution order and resource allocation method;

[0012] Service optimization feedback step: After the task is completed, feedback evaluation is conducted based on the computer system operation data and service quality indicators, and scheduling parameters and modeling strategies are adjusted to optimize subsequent scheduling performance.

[0013] In a preferred embodiment, in the task feature perception step, the priority parameters of the task are automatically extracted by a machine learning model, wherein the feature extraction process includes the data flow complexity of the task, the importance index of the perception data, and the processing delay sensitivity. The machine learning model is trained based on the historical scheduling dataset of the optoelectronic system and is used to automatically predict the priority scoring function of the task:

[0014] The calculation method of the priority scoring function includes the following steps:

[0015] Set the feature vector to F = [f1, f2, f3], where f1 is the amount of task input data, f2 is the maximum tolerable delay, and f3 is the data importance identifier;

[0016] Reasoning is performed through the neural network model N to obtain the priority score S: S = N(F);

[0017] The score S is normalized and inserted into the task queue for sorting.

[0018] In a preferred embodiment, in the resource modeling step, a resource status assessment model is constructed using a multi-factor vector and time series modeling is introduced to predict resource availability in the future. The calculation method includes the following:

[0019] Set the computer system resource attribute matrix R, including the processing unit computing frequency C_f, memory availability M_a, bus occupancy B_u and thermal stability T_s;

[0020] Build a time series model LSTM_T, train and fit the historical change sequence of the attribute matrix R, and predict the resource vector R' for the next k periods;

[0021] Match the predicted resource vector with the current task demand vector to form the resource adaptation index V:

[0022]

[0023] D i Represents the i-th resource requirement dimension of the current task, R′ i is the predicted value of the i-th resource, and n represents the total number of resources.

[0024] In a preferred embodiment, the priority evaluation step not only evaluates the weight based on the attributes of the task itself, but also comprehensively considers the overall load status of the computer system and the dependencies between tasks. The specific contents are as follows:

[0025] Construct a task dependency graph G = (T, E), where T is the set of tasks and E is the set of dependency edges;

[0026] Perform topological sorting on the task graph and extract the dependency depth D of each task d and parallelism D p ;

[0027] Combined with the resource load indicator L s , calculate the priority value P by the following formula:

[0028] P=α×W t +β×D d -γ×D p +δ×L s ;

[0029] α, β, γ, and δ are all preset adjustable weight parameters, W t is the original weight value of the task.

[0030] In a preferred embodiment, a graph matching optimization algorithm is used in the dynamic mapping step to perform a structured mapping between the task graph and the resource topology graph, ensuring that tasks with high communication intensity are preferentially mapped to processing unit nodes with higher interconnection bandwidth. The specific process includes:

[0031] Construct the task directed graph T_G and computing resource graph R_G, respectively defining the communication edge weights between nodes;

[0032] Use the weighted graph embedding matching algorithm GMatchNet to extract structural similarity indicators between tasks and resources;

[0033] Execute the matching strategy that maximizes the embedding similarity to obtain the best mapping pair M * ;

[0034]

[0035] m represents the number of tasks to be mapped, represents the kth task node, represents the kth resource node, and sim() represents the structural similarity function.

[0036] In a preferred embodiment, a reinforcement learning scheduling agent is introduced into the scheduling strategy adjustment step, which adaptively adjusts the scheduling strategy based on the feedback data of the computer system to optimize the long-term scheduling benefits, using a deep Q learning algorithm, specifically including:

[0037] The current state of the system S is defined as the task queue state, resource utilization, and scheduling history;

[0038] Define scheduling action A as a set of different scheduling strategies;

[0039] Set the reward function R, combining task completion rate, response time and energy consumption;

[0040] Learn the optimal strategy through the deep Q network DQN(S,A):

[0041] π * (S) represents the optimal policy function.

[0042] In a preferred embodiment, during the service optimization feedback step, the computer system dynamically adjusts scheduling parameters based on a rolling time window. This window uses a sliding average to predict future scheduling bottlenecks and adjusts parameters in advance. The feedback includes: the actual execution time of tasks, the frequency of resource contention, the task abandonment rate, and the QoS achievement rate.

[0043] After each round of scheduling, the computer system compares the predicted value with the actual feedback, calculates the deviation Δ, or TD error, and updates the parameters for the next round as follows:

[0044] θ new =θ old -η×Δ; η is a positive number less than 1, indicating the feedback learning rate, θ old Represents the parameter vector of the current Q network, θ new The updated parameter vector of the Q network is represented by . The computer system supports a scheduling policy version rollback mechanism. When an abnormal state occurs, it reverts to the historical stable policy state and restarts the scheduling process.

[0045] In a preferred embodiment, the task feature perception step further performs logical compression on the task flow by analyzing redundant task chains and merging aggregatable subtask nodes, thereby shortening the scheduling path. Specifically, the following contents are included:

[0046] Identify a set of tasks T_c in the task graph that have the same resource requirements and no external dependencies in the continuous execution path;

[0047] Merge the tasks in the set to form a super task node T_super;

[0048] Update the task graph dependency structure and re-evaluate and map priorities to reduce redundant scheduling decisions and improve scheduler response efficiency.

[0049] Technical effects and advantages of the present invention:

[0050] This invention introduces a task-characteristics perception mechanism, combining data flow complexity, processing delay sensitivity, and the importance index of the perceived data to automatically extract task features and generate scheduling priority scores using a neural network model. This effectively avoids the subjectivity and limitations of traditional scheduling methods that rely on manual priority setting. The system makes refined judgments based on information at the task semantic level and can dynamically adjust the scheduling order within the task flow, ensuring that critical tasks are executed first, improving overall service response speed and scheduling accuracy. It is particularly suitable for real-time scenarios in optoelectronic systems with complex task types and variable priority distributions.

[0051] This method incorporates a multi-factor resource vector and time series prediction model into the resource modeling process. Using an LSTM neural network, it models and predicts multidimensional resource attributes such as processing frequency, memory usage, communication load, and thermal stability, thereby constructing a priori models of resource availability within future scheduling cycles. The system makes resource mapping decisions based on the fit between predicted resource states and task resource requirements, making task allocation more forward-looking and structured, significantly improving resource utilization efficiency, reducing scheduling conflicts and congestion, and ensuring the stability and sustainability of optoelectronic systems in high-density task environments.

[0052] This invention establishes a complete closed loop of scheduling learning and feedback optimization by introducing a deep Q-learning scheduling agent and service optimization feedback mechanism. The system continuously explores the relationship between scheduling strategies and system operating results through reinforcement learning. After each round of scheduling, it calculates the strategy error based on feedback indicators such as task completion rate, response time, and energy consumption, and updates the model parameters in real time, thereby achieving self-adjustment and continuous evolution of the strategy. Combined with the task flow logic compression and strategy version rollback mechanism, the system is not only highly adaptable and robust, but also continuously improves decision-making quality during operation, effectively responding to complex challenges such as dynamic environments, multi-task heterogeneity, and resource conflicts in optoelectronic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0054] Figure 1 This is a schematic diagram of an overview of the computer system service optimization scheduling method for a photovoltaic system in the present invention.

[0055] Figure 2 This is a detailed schematic diagram of the computer system service optimization scheduling method for optoelectronic systems in the present invention. DETAILED DESCRIPTION

[0056] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0057] Reference Figure 1-2 The following examples were obtained:

[0058] Example 1:

[0059] Modern optoelectronic systems, such as infrared detection, lidar, and image recognition, generally rely on high-performance computing platforms for real-time data processing and task decision-making. These systems often face challenges such as complex heterogeneous computing resources, a diverse number of tasks, strong data dependencies during execution, and stringent real-time requirements. Traditional static scheduling methods and scheduling algorithms based on fixed priority rules or simplified models are no longer able to meet the requirements for efficient, dynamic, and intelligent service scheduling, especially when task loads fluctuate significantly, resource utilization is uneven, or communication is limited. Furthermore, optoelectronic systems often incorporate compute- and data-intensive tasks such as image processing, signal fusion, target recognition, and environmental modeling. These tasks place high demands on computing power, response time, and resource allocation for high concurrency and adaptability. Existing scheduling methods suffer from the following shortcomings: a lack of in-depth modeling of task characteristics, preventing the dynamic assessment of scheduling priorities based on task complexity; overly coarse resource modeling, ignoring resource volatility and predictability; difficulty in adaptively optimizing strategies based on actual system feedback; a lack of graph structure awareness, preventing intelligent mapping and scheduling of complex task flows; and a lack of learning capabilities, preventing continuous improvement based on historical behavior.

[0060] The present invention proposes a computer system service optimization scheduling method for an optoelectronic system, which aims to:

[0061] Build a scheduling system that combines task feature perception and resource dynamic modeling; design a scheduling strategy with learning and predictive capabilities to improve the intelligence of scheduling decisions and system adaptability; introduce a multi-dimensional priority evaluation mechanism and a graph structure matching algorithm to achieve a high-matching mapping between tasks and resources; utilize a deep reinforcement learning mechanism (DQN) and a service quality feedback tuning mechanism to continuously optimize the scheduling strategy; perform logic compression and dependency reorganization at the task flow level to further improve the compactness and efficiency of the scheduling path; increase the abnormal rollback capability to enhance the stability and robustness of the system under sudden loads or abnormal behavior.

[0062] The computer system service optimization scheduling method for an optoelectronic system comprises the following steps:

[0063] Task characteristic perception step, collecting key characteristic information of the tasks to be scheduled in the optoelectronic system, including task type, time constraints, data dependencies and priority levels;

[0064] Resource modeling steps,modeling the heterogeneous computing resources in the computer system,constructing a multi-dimensional resource description model including processing capacity,,idleness, power consumption characteristics and communication delay;

[0065] Priority evaluation step, based on task characteristics and resource status, a weighted evaluation mechanism is used to calculate the scheduling priority of each task to form a priority queue;

[0066] Dynamic mapping step, based on the task priority queue and resource modeling results, determines the computing resource allocation target of each task in real time, and completes the binding of tasks and resources through mapping algorithms;

[0067] Scheduling strategy adjustment steps: dynamically select the appropriate scheduling strategy based on the current computer system load status and task execution feedback, and adaptively adjust the task execution order and resource allocation method;

[0068] Service optimization feedback step: After the task is completed, feedback evaluation is conducted based on the computer system operation data and service quality indicators, and scheduling parameters and modeling strategies are adjusted to optimize subsequent scheduling performance.

[0069] The present invention first constructs a unified cognitive mechanism for external tasks and internal resources in the scheduling system through two basic steps: task characteristic perception and resource modeling. Task characteristic perception not only realizes the dynamic identification of task type, execution urgency, data dependency and business priority, but also provides high-quality input for subsequent decision-making through structured feature extraction; while resource modeling comprehensively depicts the multi-dimensional performance characteristics of the computing platform, including computing power, power consumption, communication delay and resource utilization status, thereby providing an accurate resource capability portrait for the subsequent scheduling process. This stage constructs the semantic bridge of "task-resource", laying the foundation for the system's intelligent scheduling from static pairing to dynamic matching.

[0070] After completing perception and modeling, the present invention achieves intelligent scheduling strategy generation and precise task mapping through two steps: priority evaluation and dynamic mapping. Priority evaluation not only combines the characteristic weights of the tasks themselves, but also integrates task dependencies and system load status to ensure that scheduling decisions are global and context-aware. On this basis, the dynamic mapping mechanism matches high-priority tasks to the most suitable resource nodes. Through a structural matching algorithm, it fully considers system topology and communication requirements to achieve optimal "task-to-resource" coupling. This part constitutes the core decision-making chain of the scheduling system, enabling the system to have real-time response, multi-task concurrency, and structure-aware mapping capabilities.

[0071] Finally, through the two steps of scheduling strategy adjustment and service optimization feedback, the present invention realizes a complete adaptive scheduling closed-loop control. Scheduling strategy adjustment dynamically switches or fine-tunes strategy parameters in real time according to the current state of the system and execution feedback, so as to cope with changes in task flow, resource bottlenecks or system load fluctuations; service optimization feedback quantitatively analyzes the task execution effect, and adjusts the scheduling parameters and model structure based on feedback indicators, so that the system can continuously learn and improve from historical operation data. This part gives the system the intelligent ability of "continuous optimization", which enables the scheduling behavior to evolve from "passive execution" to "active learning", and ultimately achieves the goal of maximizing task processing efficiency and optimizing resource utilization.

[0072] During the task feature perception step, the task priority parameters are automatically extracted by a machine learning model. The feature extraction process includes the task's data flow complexity, the importance index of the perceived data, and the sensitivity to processing delay. The machine learning model is trained based on the historical optoelectronic system scheduling dataset and is used to automatically predict the task priority scoring function:

[0073] The calculation method of the priority scoring function includes the following steps:

[0074] Set the feature vector to F = [f1, f2, f3], where f1 is the amount of task input data, f2 is the maximum tolerable delay, and f3 is the data importance identifier;

[0075] Reasoning is performed through the neural network model N to obtain the priority score S: S = N(F);

[0076] The score S is normalized and inserted into the task queue for sorting.

[0077] f1 represents the data flow complexity of the task, which is measured by quantifying the amount of input data that the task needs to process, reflecting the computational load scale of the task; f2 represents the processing delay sensitivity of the task, specifically the maximum delay time that can be tolerated from the scheduling to the completion of the task, which is used to reflect the task's requirements for execution timeliness; f3 represents the importance index of the perception data processed by the task. This index is obtained through the expert assignment method, combining the semantic value of the data itself and its influence on the subsequent decision-making process, and is used to identify the priority level of the task in the global task.

[0078] Extract features of the scheduled task and construct a feature vector F, which is used to fully describe the scheduling-related attributes of the task. F contains three elements:

[0079] f1: Task input data size, which indicates the size of the original data that needs to be processed before the task is executed. It is an important factor affecting the task's computing resource consumption and execution time.

[0080] f2: Maximum tolerable delay, which represents the maximum time window allowed from the start of scheduling to the completion of a task, reflecting the task's sensitivity to response speed;

[0081] f3: Data importance indicator, which indicates the value level of the data processed by the task or its criticality to the overall function completion of the system. It is an important semantic indicator for determining task priority.

[0082] After constructing the feature vector F, a pre-trained neural network model N is used for inference. Model N is trained offline based on a historical optoelectronic system scheduling dataset. Its architecture can be a multilayer perceptron network or a convolutional neural network. The training objective is to map task features to their priorities. The inference process can be expressed as: S = N(F), where S is the task priority score output by model N, representing a comprehensive measure of the task's importance, urgency, and execution priority within the current context.

[0083] After obtaining the score S, it is normalized to eliminate the impact of differences in scoring scales across tasks and ensure a fair and comparable ranking process. This normalization method applies a linear transformation based on the historical mean and variance of task scores, mapping the score S to a uniform interval (e.g., between 0 and 1). After normalization, the computer system inserts the task and its corresponding score into the task queue.

[0084] The task queue is sorted by the standardized score S value, with tasks with higher scores placed at the front of the queue and prioritized for scheduling. This sorting ensures that when resources are limited or the system is busy, the most important or urgent tasks receive computing resources first, thereby improving overall system service efficiency and scheduling responsiveness.

[0085] In the resource modeling step, a multi-factor vector is used to construct a resource status assessment model, and time series modeling is introduced to predict resource availability in the future. The calculation method includes the following:

[0086] Set the computer system resource attribute matrix R, including the processing unit computing frequency C_f, memory availability M_a, bus occupancy B_u and thermal stability T_s;

[0087] Build a time series model LSTM_T, train and fit the historical change sequence of the attribute matrix R, and predict the resource vector R' for the next k periods;

[0088] Match the predicted resource vector with the current task demand vector to form the resource adaptation index V.

[0089] The core function of the resource modeling step is to create a multidimensional model of the heterogeneous resource states within a computer system. Combined with time series forecasting methods, this method estimates the availability of various resources over multiple future scheduling cycles. This resource modeling allows the scheduling system to proactively identify potential bottlenecks or idle areas in system resources, enabling more proactive task allocation decisions.

[0090] The specific technical process is as follows:

[0091] First, we define the system resource attribute matrix R. This matrix uses time as the sequence dimension and records the key performance parameters of system resources at several moments in the past. The matrix contains four basic resource attributes, each of which represents an important dimension that affects task execution efficiency:

[0092] C_f: The computing frequency of the processing unit, which indicates the main frequency or processing speed of the current computing unit, in GHz;

[0093] M_a: Memory availability, indicating the free memory capacity currently available for task execution in the system, in GB;

[0094] B_u: Bus occupancy, which indicates the current usage ratio of the system I / O bus. The value range is 0 to 1. A higher value indicates more congested bus.

[0095] T_s: Thermal stability, which indicates the temperature safety status of the current system operating environment. The higher the value, the closer the system temperature is to the warning threshold. It is usually related to power consumption and concurrency. In the present invention, thermal stability is used as an important indicator to evaluate the carrying capacity of system scheduling. Its calculation takes into account two key influencing factors: the actual power consumption level of the current system and the number of concurrent tasks. Specifically, the thermal safety reference point of the system is set as a two-dimensional vector, which includes the maximum tolerable power consumption value and the maximum number of safe concurrent tasks. At the same time, the corresponding vector composed of the current system state is used as another point, and the Euclidean distance between the current state and the safety reference point is calculated to quantify the risk degree of thermal imbalance of the system. The smaller the distance, the more stable the system operation. Conversely, it indicates that the system is in a high temperature and high load state with poor thermal stability. Finally, the Euclidean distance is normalized and used as the thermal stability index value to participate in subsequent resource adaptation evaluation and scheduling strategy adjustment.

[0096] Each row in the resource attribute matrix R represents the system state at a specific historical moment, and each column corresponds to a resource attribute. After collecting a resource state sequence of length k, the system analyzes and fits it using the trained time series prediction model LSTM_T. LSTM_T is a recurrent neural network based on a long-short-term memory (LSTM) structure that learns how different resource attributes change over time and predicts the changing trends of each resource attribute over the next k time units.

[0097] Using the LSTM_T model, the system derives a predicted resource vector R', which represents the availability of resources across all dimensions of the system during the future scheduling period. R' is a vector of length n, where n represents the total number of resource attributes. In this solution, it is 4, containing the predicted values ​​of C_f, M_a, B_u, and T_s, corresponding to R'_1, R'_2, R'_3, and R'_4, respectively.

[0098] Next, the system obtains the resource requirement vector for the currently scheduled task, denoted as [D_1, D_2, D_3, D_4], where D_1 represents the task's processing power requirement; D_2 represents the task's memory requirement; D_3 represents the task's reliance on bus transmission capacity; and D_4 represents the task's sensitivity to system thermal stability. (_) represents a subscript symbol. To measure the degree of match between the predicted resource state and the task's resource requirements, the system calculates the resource adaptability index V, which is expressed as:

[0099] This formula comprehensively considers the coupling between task resources and resource forecasts. The numerator measures whether the task's resources are available within the system, while the denominator represents the system's burden. This allows V to reflect both the degree of matching and the degree of pressure. V typically ranges from 0 to 1, with higher values ​​indicating better resource matching.

[0100] Assume that a task's resource requirements are [D_1 = 3, D_2 = 4, D_3 = 0.5, D_4 = 1], with units of GHz, GB, ratio, and stability, respectively. The system predicts that the future resource availability vector is [R'_1 = 5, R'_2 = 6, R'_3 = 0.8, R'_4 = 1.5], indicating sufficient system resources. Substituting this into the formula: numerator = 40.9; denominator = 21.8; the fitness factor V = 40.9 / 21.8 ≈ 1.877. In this case, V > 1, indicating sufficient system resources and a high resource compatibility between the task and the resources, allowing it to be prioritized for execution.

[0101] The priority evaluation step not only evaluates the weight based on the attributes of the task itself, but also comprehensively considers the overall load status of the computer system and the dependencies between tasks. The specific contents are as follows:

[0102] Construct a task dependency graph G = (T, E), where T is the set of tasks and E is the set of dependency edges;

[0103] Perform topological sorting on the task graph and extract the dependency depth D_d and parallelism D_p of each task;

[0104] Combined with the resource load indicator L_s, the priority value P is calculated.

[0105] Construct a task dependency graph G = (T, E), where T represents the set of all tasks to be scheduled, with each task corresponding to a node in the graph. E represents the data or execution dependencies between tasks, with each directed edge indicating that one task must complete before another can begin. This graph is typically a directed acyclic graph, used to represent the sequential logic and concurrent structure of the task flow.

[0106] The computer system then topologically sorts the dependency graph G and extracts two structural indicators for each task node: D d : Dependency depth, which indicates the longest path length from the starting node to the current task node in the graph, reflecting the execution level of the task; D p : Parallelism, which indicates the number of tasks at the same level as the current task that can be executed simultaneously, reflecting the parallel adjustability of the current task.

[0107] At the same time, the computer system obtains the current resource load status indicator L s This value can be integrated with multiple factors such as system CPU occupancy, memory occupancy, bus pressure, etc. For example, the value is output through fuzzy logic method. The value reflects the overall scheduling pressure of the system. The higher the value, the busier the system. Finally, the computer uses the following priority evaluation function to calculate the task priority value P: P = α × W t +β×D d -γ×D p +δ×L s ;

[0108] W t The original weight of the task is typically assigned a constant value based on task importance, the value of the processed data, or external policies. α, β, γ, and δ are adjustable weight parameters that control the influence of each factor in the priority assessment and can be set based on system characteristics and scenario requirements. The higher the original weight of a task, the higher its priority. The greater the dependency depth (the later the task is in the process, or the longer the path), the greater its urgency, and the higher its priority should be. The higher the parallelism, the greater the degree of freedom it can be scheduled, and the priority can be appropriately lowered to allow more urgent tasks to execute first. The heavier the computer system load, the more dynamic the priority of certain tasks should be to guide tasks smoothly into the execution phase.

[0109] For example: Assume that the indicator of a task is: W t =5 (task importance is high), D d =3 (deeper in the task flow), D p =4 (more parallel tasks), L s=0.7 (the system is under moderate load). If the weight parameters are set to: α = 1.0, β = 0.5, γ = 0.3, δ = 2.0; then the priority is calculated as follows: P = 1.0 × 5 + 0.5 × 3 - 0.3 × 4 + 2.0 × 0.7 = 5 + 1.5 - 1.2 + 1.4 = 6.7

[0110] Finally, the task is sorted with other tasks according to the P value. The tasks with higher values ​​are prioritized to enter the front of the scheduling queue to ensure that critical tasks, long-path tasks, or tasks that are highly sensitive to resources are executed first.

[0111] In the dynamic mapping step, a graph matching optimization algorithm is used to perform a structured mapping between the task graph and the resource topology graph, ensuring that tasks with high communication intensity are preferentially mapped to processing unit nodes with higher interconnection bandwidth. The specific process includes:

[0112] Construct the task directed graph T_G and computing resource graph R_G, respectively defining the communication edge weights between nodes;

[0113] Use the weighted graph embedding matching algorithm GMatchNet to extract structural similarity indicators between tasks and resources;

[0114] Execute the matching strategy that maximizes the embedding similarity to obtain the best mapping pair M * .

[0115] The dynamic mapping step is used to achieve structurally optimal allocation between tasks and computing resources. Due to the complex dependencies and communication requirements of tasks in optoelectronic systems, scheduling based solely on task weights or resource availability often fails to align with the underlying communication structure, easily leading to bottleneck nodes or resource congestion. To address this issue, this paper proposes a dynamic mapping method based on structured graph matching optimization. This method embeds and compares the task dependency structure with the resource topology, achieving mapping combinations with high structural similarity and optimizing overall communication efficiency.

[0116] First, the computer system constructs a directed task graph, T_G, with task nodes as the graph elements. Edges between nodes represent data or execution dependencies between tasks. Each edge is assigned a weight, representing the communication overhead or data transfer required between tasks. Tasks with high communication intensity have larger adjacent edge weights, indicating their greater communication sensitivity within the computer system.

[0117] Simultaneously, the computer system constructs a resource topology graph, R_G, where each node represents a processing unit, edges represent the network connections between nodes, and edge weights represent the communication bandwidth or latency of the corresponding link. Higher values ​​represent stronger communication capabilities. This allows the resource graph R_G to truly reflect the interconnectivity and load paths between computing resources.

[0118] The computer system then used the weighted graph embedding matching algorithm GMatchNet to perform a structural similarity analysis on T_G and R_G. GMatchNet is a graph alignment model based on deep graph neural networks. It embeds nodes in two graphs into a unified vector space and measures the structural similarity between each pair of nodes in this space. This embedding takes into account node properties, adjacency structure, and edge weights, accurately reflecting the structural coupling between "task nodes" and "resource nodes."

[0119] After completing the embedded calculation, the computer system performs and resource nodes The structural similarity between them is measured to obtain a similarity score. The higher the score, the higher the structural matching degree between the task node and the resource node, which may lead to higher communication efficiency after mapping.

[0120] The computer system finally executes the following optimal matching strategy to obtain the optimal mapping relationship M* of all task nodes:

[0121] m represents the number of tasks to be scheduled; represents the kth task node; represents the kth computing resource node; sim() is the structural similarity function output by the GMatchNet embedding model. The computer system prioritizes mapping pairs with the highest total similarity score, ensuring that high-communication-intensive tasks are mapped to combinations of nodes with strong communication capabilities, maximizing data flow efficiency, minimizing communication latency, and improving overall scheduling performance.

[0122] For example, assume there are three task nodes. Tasks A and B require a large amount of image data transmission, and Task C has no strong dependencies on other tasks, forming a task graph T_G. Simultaneously, there are three resource nodes, where the bandwidth between Node 1 and Node 2 is extremely high, while Node 3 has weaker communication capabilities, forming a resource graph R_G. Through GMatchNet embedding calculations, the computer system discovers that the structural adjacency of Tasks A and B highly matches the interconnection characteristics of Nodes 1 and 2, resulting in a high score. Task C, on the other hand, has a low dependency on communication and can be assigned to Node 3. Ultimately, the computer system achieves the optimal mapping: Task A → Node 1, Task B → Node 2, and Task C → Node 3. This structural matching not only improves task processing efficiency but also reduces communication bottlenecks between tasks. This can significantly optimize the overall computer system scheduling quality, particularly when processing large-scale image or laser point cloud data in optoelectronic systems.

[0123] The scheduling strategy adjustment step introduces a reinforcement learning scheduling agent, which adaptively adjusts the scheduling strategy based on the feedback data from the computer system to optimize the long-term scheduling benefits. It uses a deep Q-learning algorithm, specifically including:

[0124] The current state of the system S is defined as the task queue state, resource utilization, and scheduling history;

[0125] Define scheduling action A as a set of different scheduling strategies;

[0126] Set the reward function R, combining task completion rate, response time and energy consumption;

[0127] Learn the optimal strategy through the deep Q network DQN(S,A); π * (S) represents the optimal policy function.

[0128] The scheduling strategy adjustment step is the core component of achieving intelligent and adaptive scheduling capabilities. To adapt to changing task characteristics and resource status fluctuations in complex dynamic environments, this step introduces a scheduling agent based on reinforcement learning. This agent uses historical scheduling feedback to continuously optimize the strategy, thereby maximizing long-term system performance.

[0129] A scheduling agent is built using the Deep Q-Learning algorithm. This agent uses a reward function to evaluate the long-term benefits of different scheduling strategies and continuously adjusts its strategy selection behavior. This process does not rely on manual empirical rules and can be self-trained and continuously improved through interactive data during system operation.

[0130] First, the current scheduling state is defined as a state vector S, which includes at least the following three dimensions:

[0131] Task queue status: indicates the number of tasks to be executed in the current task queue, priority distribution, resource demand structure, etc.

[0132] Resource utilization: reflects the usage of each processing unit in the current system, such as the occupancy rate of processors, memory, communication buses and other resources;

[0133] Scheduling history: Contains the scheduling result data from the previous round or multiple cycles, such as the strategy number used, strategy effect evaluation, etc.

[0134] Then, we define the scheduling action A, which is the set of currently available scheduling strategies. Each action represents a predefined scheduling strategy, such as shortest task first, least load first, or priority weighted strategy. The system will select the best one among these strategies to execute.

[0135] To guide the agent to make effective decisions, the system designs a reward function R, which is used to evaluate the scheduling results after adopting a certain strategy in the current state. The reward value can be composed of the following weighted indicators:

[0136] Task completion rate (the percentage of tasks successfully executed);

[0137] Average response time (the average time from task enqueuing to task completion);

[0138] Energy consumption (the total energy consumption per unit time after the system executes the scheduling strategy);

[0139] The above three indicators can be combined and weighted by pre-trained machine learning algorithms such as convolutional neural network models and regression models to form a scalar score R, which is used to guide the scheduling agent to update the strategy value estimation.

[0140] Then, a deep Q network (DQN(S,A)) is constructed. This network takes state S and action A as input and outputs a Q value, which is the expected long-term cumulative reward for taking action A in the current state S. The goal of the policy function is to maximize this Q value. The system selects the optimal policy according to the following formula:

[0141] π * (S) represents the optimal strategy selection under the current state S, and Q(S,A) is the long-term return value of the strategy predicted by the deep network.

[0142] For example, assume the current state S contains 10 tasks in the task queue, 60% of which are high-priority tasks; system CPU utilization is 85%, and memory usage is 70%; the strategy used in the last two rounds is "shortest task first," which has a completion rate of 90% and a low average response time, but high energy consumption. The system scheduling action set A contains three strategies: Action 1: Shortest task first; Action 2: Priority-weighted scheduling; and Action 3: Minimum resource consumption scheduling.

[0143] In the current state, the system uses the DQN(S,A) model to evaluate the long-term Q-values ​​of each of the three strategies. Assume that the Q-value for action 1 is 3.2, the Q-value for action 2 is 4.1, and the Q-value for action 3 is 2.9. The system then selects action 2 based on the maximum Q-value principle. The corresponding strategy is "priority weighted scheduling," which will be used in this round of scheduling. The system also records the actual reward R and uses it to update model parameters in the next round of training.

[0144] In the service optimization feedback step, the computer system dynamically adjusts the scheduling parameters based on a rolling time window. This window uses a sliding average to predict future scheduling bottlenecks and adjusts the parameters in advance. The feedback includes: the actual execution time of the task, the frequency of resource contention, the task abandonment rate, and the QoS achievement rate.

[0145] After each round of scheduling, the computer system compares the predicted value with the actual feedback, calculates the deviation Δ, or TD error, and updates the parameters for the next round.

[0146] A rolling time window mechanism is used for scheduling performance monitoring and trend modeling. A rolling window is a fixed-length sequence of historical data that is continuously updated with the scheduling cycle and always retains the statistical values ​​of the scheduling results of the most recent rounds. The data in this window is used to perform a sliding average forecast, which takes a weighted average of key scheduling indicators across multiple periods within the window to obtain the expected value of system performance over a period of time, thereby predicting potential bottlenecks. Predictive indicators can include: system response time trend; high-priority task accumulation trend; resource utilization fluctuation trend. Through the sliding average model, the system can identify potential scheduling risks such as continued increases in task delays and increased frequency of task abandonment, providing early warnings and preparing parameter adjustments.

[0147] After each round of task scheduling, the system collects the following feedback information for evaluation: Task execution time: The actual time it takes for a task to be scheduled and completed; Resource contention frequency: The number of times multiple tasks conflict or wait for the same resource; Task abandonment rate: The proportion of tasks that were not executed in the current scheduling round due to resource conflicts, timeouts, or low priority; and Quality of Service (QoS) achievement rate: The proportion of completed tasks in the current round that met response time and success rate requirements. This feedback data reflects the actual performance of the scheduling strategy and is compared with the predicted value obtained by the sliding average to form a performance deviation Δ, also known as the time difference error (TD error).

[0148] The system's deep Q-network model, built on a reinforcement learning mechanism, updates its internal parameters after each round of scheduling to improve future strategic decision-making capabilities. Parameter updates use the following standard gradient descent form:

[0149] θ new =θ old -η×Δ;θ old is the parameter vector in the current deep Q network, including the weights and biases in the model; Δ is the TD error between the predicted value and the actual feedback; η is the learning rate, which ranges from (0, 1) and is used to control the step size of each update; θ new is the updated parameter vector, which will take effect in the next round of scheduling decisions.

[0150] For example, suppose the policy model currently used by the system predicts an average task response time of 2 seconds over the past five rounds, while the actual measured average response time is 2.6 seconds, indicating that the system processing latency is higher than expected. At the same time, feedback indicates that the task abandonment rate is 8%, exceeding the upper tolerance limit of 5%, and resource contention is high. The system calculates the error value Δ for this round and updates the policy model parameter vector based on a preset learning rate (e.g., 0.1).

[0151] The computer system supports a scheduling policy version rollback mechanism. When an abnormal state occurs, it reverts to a historically stable policy state and restarts the scheduling process. For example, during actual operation, if the system detects a significant decrease in task completion rate over two consecutive scheduling cycles, a task abandonment rate exceeding a set threshold, and an abnormal increase in resource contention frequency, it may determine that the current scheduling policy is at risk of failure. At this time, the system automatically triggers the rollback mechanism, retrieves the last verified stable policy parameter set from the historical record, restores that policy, and reinitializes the scheduling process. This prevents the system from continuing to operate under an inefficient or even crashing policy, ensuring the continuity of task execution and the stability of the system service quality.

[0152] In the task feature perception step, the task flow is further logically compressed by analyzing redundant task chains and merging aggregatable subtask nodes, thereby shortening the scheduling path. The specific contents include:

[0153] Identify a set of tasks T_c in the task graph that have the same resource requirements and no external dependencies in the continuous execution path;

[0154] Merge the tasks in the set to form a super task node T_super;

[0155] Update the task graph dependency structure and re-evaluate and map priorities to reduce redundant scheduling decisions and improve scheduler response efficiency.

[0156] This logical compression process is based on a structural analysis of the task dependency graph. It identifies sets of tasks in the task graph that have identical resource requirements and no external dependencies, denoted as T_c. A task set T_c satisfies the following two conditions: all task nodes are located on the same directed path in the task graph, with no other task branches intervening; the resource types and resource specifications of all tasks are identical, avoiding the additional scheduling costs associated with switching resource types; and there are no cross-set dependencies between tasks within the set, meaning that their execution does not depend on tasks outside the set, nor is it dependent on by external tasks.

[0157] After the above conditions are met, the system merges all tasks in set T_c into a supertask node, T_super. This node represents the logical aggregation of the original set of tasks in the task graph. The resource requirements of T_super are equal to the resource allocation of any task in the original set. The execution time can be the sum of the execution times of the original tasks, or a conservative maximum estimate.

[0158] After the merge operation is complete, the system replaces node T_c in the original task graph with node T_super and reconnects all dependency edges connected to T_c in the graph to maintain logical consistency in the dependency structure. All dependency edges that originally pointed to a task in T_c are redirected to T_super, and all dependency edges from a task in T_c pointing to the outside are also updated.

[0159] Finally, the system re-evaluates priorities and remaps resources based on the updated task graph. With fewer tasks and shorter dependency paths, the scheduler's complexity in topological sorting, weight calculation, and task allocation is significantly reduced, improving scheduling decision speed and system response efficiency.

[0160] For example, consider an optoelectronic image analysis system with three consecutive tasks: image preprocessing, image noise reduction, and image format conversion. These tasks form a linear path in the task graph, share identical resource requirements, require access to the same type of image processing accelerator module, and interact with no other tasks in the task dependency graph. The system identifies these three tasks as forming T_c, meeting the merge criteria, and aggregates them into T_super, denoted as the "image input module unified processing node." This merge replaces the original three nodes and two dependency edges with a single node and its forward and backward edges. This simplifies the task graph, and the scheduler no longer schedules these three tasks separately, reducing system overhead and improving scheduling efficiency.

[0161] Implementation method 1: Real-time scheduling optimization based on image target recognition:

[0162] In a certain optoelectronic imaging system, there is a large amount of image data collected by multiple cameras or infrared sensors that needs to be processed simultaneously. Its tasks include multiple steps such as image preprocessing, feature extraction, target recognition and classification, and these tasks have clear execution dependencies and delay constraints. In this embodiment, the system first analyzes the data volume, processing delay tolerance and image content importance indicators of each task through the task characteristic perception module, and automatically generates a priority score by the pre-trained neural network model. Subsequently, the scheduling system uses the LSTM model to predict the resource availability of each processor, calculates the compatibility of tasks and resources, and executes the graph matching algorithm to map key tasks to the computing node with the best communication performance. During the task execution process, the system uses the deep Q learning method to continuously evaluate the effects of different strategies, dynamically optimizes the scheduling strategy according to the task completion rate and response time, and ultimately achieves rapid recognition and response to the target image, ensuring that the system still has high efficiency and stability under load fluctuation conditions.

[0163] Implementation Method 2: Heterogeneous Computing Task Scheduling Based on Multi-Mode Sensor Fusion:

[0164] In a certain multi-mode optoelectronic sensor fusion platform, infrared, laser, and visible light sensors simultaneously collect multi-channel data, forming a large number of parallel computing tasks. These tasks have different computing requirements and priorities, and some tasks have strong dependencies. This implementation automatically analyzes the execution order and parallelism of tasks by constructing a task dependency graph, and combines the current resource status with a multi-factor weighted priority calculation model to comprehensively consider task complexity, dependency depth, and system load to generate a task scheduling priority list. The scheduling engine compares the task graph with the resource topology graph through a graph embedding model to obtain the optimal mapping relationship, ensuring that bandwidth-sensitive tasks are preferentially allocated to high-throughput nodes. During system operation, the scheduler uses a rolling time window feedback mechanism to perform real-time backtracking of execution results, fine-tune resource allocation parameters and scheduling strategies, and improve resource utilization and QoS stability. When sudden abnormal loads or resource bottlenecks occur, the system can roll back to the historical stable strategy to ensure uninterrupted service.

[0165] Implementation method three: Adaptive scheduling optimization mechanism based on intelligent unmanned systems:

[0166] In intelligent unmanned reconnaissance platforms, optoelectronic systems must perform complex tasks such as environmental perception, threat analysis, and path planning during flight. These tasks not only require strong real-time performance but also face a frequently changing resource scheduling environment. In this implementation, the scheduling system perceives task characteristics and resource status in real time, automatically extracts key task features, and assesses their priorities. A logical compression mechanism is also introduced to identify mergable subtasks within the task flow, reconstructing them into supertasks to reduce scheduling complexity. The resource model utilizes a time series prediction model to dynamically predict the load of each computing unit in future cycles, enabling predictive scheduling of future resource status. The scheduling engine leverages reinforcement learning methods to automatically optimize strategy selection based on system feedback and dynamically updates model parameters based on metrics such as task completion rate, response latency, and power consumption. The system as a whole possesses the capabilities of intelligent task reconstruction, scheduling path compression, and continuous strategy self-learning, enabling efficient and stable task scheduling services under the conditions of resource constraints and rapidly changing scenarios on unmanned platforms.

[0167] These three implementation methods respectively emphasize the practical applications and innovations of the present invention in typical optoelectronic system scenarios such as image processing, sensor fusion and unmanned platforms, and all demonstrate a complete process closed loop and the ability to solve technical problems.

[0168] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0169] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0170] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0171] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0172] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A computer system service optimization scheduling method for an optoelectronic system, characterized in that: The following steps are involved: Task characteristic perception step, collecting key characteristic information of the tasks to be scheduled in the optoelectronic system, including task type, time constraints, data dependencies and priority levels; Resource modeling steps,modeling the heterogeneous computing resources in the computer system,constructing a multi-dimensional resource description model including processing capacity,,idleness, power consumption characteristics and communication delay; Priority evaluation step, based on task characteristics and resource status, a weighted evaluation mechanism is used to calculate the scheduling priority of each task to form a priority queue; Dynamic mapping step, based on the task priority queue and resource modeling results, determines the computing resource allocation target of each task in real time, and completes the binding of tasks and resources through mapping algorithms; Scheduling strategy adjustment steps: dynamically select the appropriate scheduling strategy based on the current computer system load status and task execution feedback, and adaptively adjust the task execution order and resource allocation method; Service optimization feedback step: After the task is completed, feedback evaluation is conducted based on the computer system operation data and service quality indicators, and scheduling parameters and modeling strategies are adjusted to optimize subsequent scheduling performance.

2. The computer system service optimization scheduling method for optoelectronic systems according to claim 1, characterized in that: During the task feature perception step, the task priority parameters are automatically extracted by a machine learning model. The feature extraction process includes the task's data flow complexity, the importance index of the perceived data, and the sensitivity to processing delay. The machine learning model is trained based on the historical optoelectronic system scheduling dataset and is used to automatically predict the task priority scoring function: The calculation method of the priority scoring function includes the following steps: Set the feature vector to F = [f1, f2, f3], where f1 is the amount of task input data, f2 is the maximum tolerable delay, and f3 is the data importance identifier; Reasoning is performed through the neural network model N to obtain the priority score S: S = N(F); The score S is normalized and inserted into the task queue for sorting.

3. The computer system service optimization scheduling method for optoelectronic systems according to claim 2, characterized in that: In the resource modeling step, a multi-factor vector is used to construct a resource status assessment model, and time series modeling is introduced to predict resource availability in the future. The calculation method includes the following: Set the computer system resource attribute matrix R, including the processing unit computing frequency C_f, memory availability M_a, bus occupancy B_u and thermal stability T_s; Build a time series model LSTM_T, train and fit the historical change sequence of the attribute matrix R, and predict the resource vector R' for the next k periods; Match the predicted resource vector with the current task demand vector to form the resource adaptation index V: D i Represents the i-th resource requirement dimension of the current task, R′ i is the predicted value of the i-th resource, and n represents the total number of resources.

4. The computer system service optimization scheduling method for optoelectronic systems according to claim 3, characterized in that: The priority evaluation step not only evaluates the weight based on the attributes of the task itself, but also comprehensively considers the overall load status of the computer system and the dependencies between tasks. The specific contents are as follows: Construct a task dependency graph G = (T, E), where T is the set of tasks and E is the set of dependency edges; Perform topological sorting on the task graph and extract the dependency depth D of each task d and parallelism D p ; Combined with the resource load indicator L s , calculate the priority value P by the following formula: P=α×W t +β×D d -γ×D p +δ×L s ; α, β, γ, and δ are all preset adjustable weight parameters, W t is the original weight value of the task.

5. The computer system service optimization scheduling method for optoelectronic systems according to claim 4, characterized in that: In the dynamic mapping step, a graph matching optimization algorithm is used to perform a structured mapping between the task graph and the resource topology graph, ensuring that tasks with high communication intensity are preferentially mapped to processing unit nodes with higher interconnection bandwidth. The specific process includes: Construct the task directed graph T_G and computing resource graph R_G, respectively defining the communication edge weights between nodes; Use the weighted graph embedding matching algorithm GMatchNet to extract structural similarity indicators between tasks and resources; Execute the matching strategy that maximizes the embedding similarity to obtain the best mapping pair M * ; m represents the number of tasks to be mapped, represents the kth task node, represents the kth resource node, and sim() represents the structural similarity function.

6. The computer system service optimization scheduling method for optoelectronic systems according to claim 5, characterized in that: The scheduling strategy adjustment step introduces a reinforcement learning scheduling agent, which adaptively adjusts the scheduling strategy based on the feedback data from the computer system to optimize the long-term scheduling benefits. It uses a deep Q-learning algorithm, specifically including: The current state of the system S is defined as the task queue state, resource utilization, and scheduling history; Define scheduling action A as a set of different scheduling strategies; Set the reward function R, combining task completion rate, response time and energy consumption; Learn the optimal strategy through the deep Q network DQN(S,A): π * (S) represents the optimal policy function.

7. The computer system service optimization scheduling method for optoelectronic systems according to claim 6, characterized in that: In the service optimization feedback step, the computer system dynamically adjusts the scheduling parameters based on a rolling time window. This window uses a sliding average to predict future scheduling bottlenecks and adjusts the parameters in advance. The feedback includes: the actual execution time of the task, the frequency of resource contention, the task abandonment rate, and the QoS achievement rate. After each round of scheduling, the computer system compares the predicted value with the actual feedback, calculates the deviation Δ, or TD error, and updates the parameters for the next round as follows: θ new -θ old -η×Δ; η is a positive number less than 1, indicating the feedback learning rate, θ old Represents the parameter vector of the current Q network, θ new represents the parameter vector of the updated Q network.

8. The computer system service optimization scheduling method for optoelectronic systems according to claim 7, characterized in that: The computer system supports a scheduling policy version rollback mechanism. When an abnormal state occurs, it will go back to the historical stable policy state and restart the scheduling process.

9. The computer system service optimization scheduling method for optoelectronic systems according to claim 8, characterized in that: In the task feature perception step, the task flow is further logically compressed by analyzing redundant task chains and merging aggregatable subtask nodes, thereby shortening the scheduling path. The specific contents include: Identify a set of tasks T_c in the task graph that have the same resource requirements and no external dependencies in the continuous execution path; Merge the tasks in the set to form a super task node T_super; Update the task graph dependency structure and re-evaluate and map priorities to reduce redundant scheduling decisions and improve scheduler response efficiency.

Images