Master-slave hierarchical OpenClaw terminal access and cooperative management system

Abstract

This invention relates to the field of distributed intelligent agent management technology, and more particularly to a master-slave hierarchical OpenClaw terminal access and collaborative management system. The system includes an acquisition module, a judgment module, an index determination module, an update module, an intensity determination module, a generation module, and an adjustment module. This invention collects the task complexity, response latency, resource utilization, and user permissions of child nodes; classifies tasks based on thresholds to determine whether the parent entity should intervene; calculates the capability index of each child node by combining the classification results, latency, resource utilization, and the parent entity's call success rate, and aggregates these into a global distribution parameter. If the threshold is exceeded, the parent entity's evolutionary update strategy is triggered; the capability intensity is output based on the evolved strategy and permission coefficient; parameters are distributed according to intensity levels, and new nodes are incubated to address high-frequency task gaps; the threshold and permission coefficient are adjusted in reverse based on response latency and completion status, forming a closed loop to achieve efficient collaborative evolution of the master-slave hierarchical intelligent agent.

Description

Technical Field

[0001] This invention relates to the field of distributed intelligent agent management technology, and in particular to a master-slave hierarchical OpenClaw terminal access and collaborative management system. Background Technology

[0002] With the widespread adoption of smart terminals and distributed service nodes, massive user requests are mixed with a large number of routine tasks and special tasks with high complexity and high privilege requirements. The central node of the traditional master-slave architecture lacks the ability to continuously learn from distributed nodes, evolve on demand, and dynamically allocate resources. It is also difficult to provide in-depth services to advanced users while ensuring the response efficiency of ordinary users. How to achieve the convergence of group experience, adaptive evolution of master nodes, hierarchical access of users, and elastic resource expansion without downtime has become the core challenge in the field of multi-agent collaborative services.

[0003] Chinese Patent Publication No. CN118555614A discloses a method for task offloading and service caching in mobile edge computing based on hierarchical reinforcement learning. The method includes: first, constructing a task offloading decision model using an artificial neural network and training the model until convergence in a mobile edge computing simulation environment; second, constructing a service caching decision model using an artificial neural network and training it until convergence in the same environment; finally, entering the execution phase of the task offloading and service caching method, the trained model is first deployed to the corresponding server. The task offloading decision model makes task offloading decisions based on the state observed by the edge server. After a period of time, a service caching decision is triggered, and the service caching decision model makes a service caching decision based on the task generation and offloading actions of mobile terminals within the edge server's coverage area during this period. Based on the temporal characteristics of the task offloading and service caching problem, a hierarchical approach is introduced to reduce the complexity of the problem and jointly optimize system latency and energy consumption over a long time dimension.

[0004] Therefore, the existing technology has the following problems: it relies on a neural network model that has been trained to convergence in a simulated environment in advance, which is prone to insufficient model generalization ability and decision bias in real dynamic edge scenarios; it relies on an offline deployment method that triggers service caching decisions at fixed periods, which is prone to ignoring the real-time burstiness of user requests and the dynamic changes in node load; it relies on a static training and independent deployment architecture of hierarchical reinforcement learning, which is prone to lacking the ability of online aggregation of group experience and continuous adaptive evolution of master nodes. Summary of the Invention

[0005] To address this, the present invention provides a master-slave hierarchical OpenClaw terminal access and collaborative management system, which overcomes the problems of low adaptability and slow response speed in the face of mixed load scenarios with sudden high concurrency requests, dynamic interweaving of user permissions and task complexity due to the reliance on offline training and static deployment in the prior art through a master-slave hierarchical architecture and a real-time terminal access collaborative mechanism.

[0006] To achieve the above objectives, the present invention provides a master-slave hierarchical OpenClaw terminal access and collaborative management system, comprising:

[0007] The acquisition module is used to acquire in real time the interaction requests, task complexity, response latency, resource utilization, user level identifiers and permission tags of each child node in the public virtual environment, and to acquire the capability call success rate of the master node.

[0008] The determination module is used to determine whether to trigger the main node intervention based on the task complexity and the preset complexity threshold, so as to obtain the task classification result.

[0009] The index determination module is used to determine the capability index of each sub-node based on the task classification results, the response latency, the resource utilization rate, and the capability call success rate.

[0010] The update module is used to determine the global capability distribution parameters based on each capability index and resource utilization rate, and to determine whether the evolution conditions of the master node are met based on the change range of the global capability distribution parameters, so as to update the global strategy parameters of the master node based on the determination result of meeting the evolution conditions.

[0011] The strength determination module is used to determine the capability output strength based on the call permission coefficient and the global capability distribution parameters, wherein the call permission coefficient is calculated based on the user level identifier, the permission tag and the task complexity;

[0012] The generation module is used to distribute global capability distribution parameters or global strategy parameters to the corresponding child nodes according to the capability output strength, and to generate new child nodes based on the real-time acquired task request frequency and capability index.

[0013] The adjustment module is used to adjust the preset complexity threshold or the call permission coefficient based on the response latency and task completion status after distribution.

[0014] Furthermore, the determination module includes:

[0015] The offset calculation unit is used to calculate the complexity offset based on the difference between the task complexity and the preset complexity threshold, and to determine the task classification result based on the mapping relationship between the complexity offset and the preset classification interval.

[0016] The determination unit is used to determine whether to trigger the main node to intervene based on the comparison result that the complexity offset is greater than a preset trigger threshold.

[0017] Furthermore, the index determination module includes:

[0018] The weight determination unit is used to determine the corresponding task weight coefficient based on the task classification result when determining the triggering of the master node intervention;

[0019] The delay determination unit is used to determine the delay deviation value based on the response delay and the preset delay reference value when it is determined that the master node intervention is triggered;

[0020] The resource determination unit is used to determine the resource deviation value based on the resource occupancy rate and the preset resource occupancy benchmark value when it is determined that the master node intervention is triggered;

[0021] The index determination unit is used to determine the basic capability value based on the capability call success rate, the latency deviation value, and the resource deviation value, and modulate the basic capability value in combination with the task weight coefficient to obtain the capability index of each sub-node.

[0022] Furthermore, the update module includes:

[0023] A distributed building unit, which is used to construct a capability distribution sequence based on the capability index and resource utilization of each child node;

[0024] A parameter extraction unit is used to extract global capability distribution parameters from the capability distribution sequence;

[0025] The change determination unit is used to determine the distribution change amplitude based on the change trend of the global capability distribution parameters within a preset change determination time period;

[0026] An evolution determination unit is used to determine whether the evolution conditions of the master node are met based on the comparison between the distribution change amplitude and the preset evolution threshold.

[0027] The strategy update unit is used to update the global strategy parameters of the master node when the evolution conditions are met.

[0028] Furthermore, the distributed building unit includes:

[0029] The sorting subunit is used to sort the capability indices of each child node according to their numerical values ​​to form an ordered capability sequence.

[0030] A difference calculation subunit is used to calculate a capability difference sequence based on the difference between adjacent capability indices in the ordered capability sequence.

[0031] A normalization subunit is used to normalize the capability difference sequence based on the resource occupancy rate of each child node to obtain the capability distribution sequence.

[0032] Furthermore, the intensity determination module includes:

[0033] A permission determination unit is used to determine a call permission coefficient based on the user level identifier, the permission tag, and the task complexity.

[0034] The gap determination unit is used to determine the capability gap value based on the difference between the capability index of each child node and the global capability distribution parameter;

[0035] The strength calculation unit is used to determine the capability output strength based on the call permission coefficient, the capability gap value, and the global capability distribution parameters.

[0036] Furthermore, the intensity calculation unit includes:

[0037] A mapping subunit is used to determine the gap level based on the mapping relationship between the capacity gap value and a preset gap range;

[0038] A constraint subunit is used to constrain the gap level according to the call permission coefficient to obtain the available capability level;

[0039] An adjustment subunit is used to adjust the available capability level according to the global capability distribution parameters to obtain the capability output intensity.

[0040] Furthermore, the generation module includes:

[0041] A distribution unit is used to distribute the global capability distribution parameters based on a comparison result where the capability output strength is less than a preset strength threshold, and to distribute the global strategy parameters based on a comparison result where the capability output strength is greater than or equal to the preset strength threshold.

[0042] The frequency determination unit is used to obtain the task request frequency of various tasks within a preset statistical time period in real time, and to determine the high-frequency task set based on the processing tasks corresponding to the task request frequency greater than the preset frequency threshold.

[0043] A gap determination unit is used to determine the capability gap value based on the difference between the capability index corresponding to the high-frequency task set and the global capability distribution parameter;

[0044] A generation unit is used to generate new child nodes based on the comparison result between the capability gap value and the preset gap threshold, wherein the new child nodes are initialized based on the global policy parameters.

[0045] Furthermore, the adjustment module includes:

[0046] The preprocessing unit is used to sample the response delay and task completion status after capacity distribution within a preset observation period, and to standardize the sampling results to obtain a standardized delay sequence and a standardized completion sequence, respectively.

[0047] The observation and determination unit is used to determine the time delay fluctuation characteristic value based on the standardized time delay sequence, and to determine the completion stability characteristic value based on the standardized completion sequence.

[0048] The indicator calculation unit is used to determine the short-cycle performance indicator value based on the time delay fluctuation characteristic value and the completion stability characteristic value, and to determine the comprehensive performance indicator value according to the changing trend of all cycle performance indicator values ​​within the preset adjustment period.

[0049] An adjustment unit is used to adjust the preset complexity threshold or the call permission coefficient based on the comprehensive performance index value.

[0050] Furthermore, the adjustment unit includes:

[0051] The interval determination subunit is used to divide the comprehensive performance index value into a performance insufficiency interval, a performance stability interval, or a performance excess interval based on the comparison results between the comprehensive performance index value and the preset first index threshold and the preset second index threshold.

[0052] A threshold adjustment subunit is used to reduce the preset complexity threshold when the comprehensive performance index value is in the performance insufficiency range, and to increase the preset complexity threshold when the comprehensive performance index value is in the performance excess range.

[0053] The license adjustment subunit is configured to increase the call license coefficient when the overall performance index value is in the underperformance range, and to decrease the call license coefficient when the overall performance index value is in the overperformance range.

[0054] Compared with existing technologies, the beneficial effects of this invention are as follows: The acquisition module collects real-time data on the task complexity, response latency, resource utilization, and user level identifiers and permission tags of each child node; the judgment module classifies tasks based on a preset complexity threshold to determine whether the master node should intervene; the index determination module combines the task classification results with response latency, resource utilization, and the success rate of parent node capability invocation to calculate a capability index characterizing the real-time service level of each child node; and the update module aggregates the capability indices and resource utilization of all child nodes to form a global capability distribution parameter. When the change in this distribution parameter exceeds a threshold, the parent node is determined to meet the evolution conditions, thus updating its global strategy parameters. The more inefficient or differentiated the overall capability distribution of the child node group becomes, the more the parent node needs to evolve to adjust its strategy. The evolved strategy parameters will further affect the capability output strength calculated by the strength determination module. This strength is also constrained by the invocation permission coefficient. The capability coefficient is jointly determined by user level identifier, permission tag, and task complexity, ensuring that ordinary users only obtain basic capabilities while advanced users receive more comprehensive capability feedback. The generation module selects to distribute global capability distribution parameters or global strategy parameters to corresponding child nodes based on the strength of capability output. It also automatically incubates new child nodes for high-frequency task gaps where the task request frequency is high and the capability index is insufficient. The adjustment module reversely adjusts the preset complexity threshold and call permission coefficient based on the actual observed response latency and task completion status after capability distribution. This allows the system to automatically lower the parent intervention threshold and relax the permission coefficient to call the parent capability to alleviate the bottleneck when facing performance insufficiency, and raise the intervention threshold and tighten the permission coefficient when performance is excessive to avoid resource waste. This forms a complete closed-loop logic from data collection capability quantification group distribution evolution trigger intensity output performance feedback parameter adaptation, ultimately achieving efficient collaboration and adaptive evolution of the master-slave hierarchical intelligent agent system. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the master-slave hierarchical OpenClaw terminal access and collaborative management system in this embodiment;

[0056] Figure 2 This is a logic diagram for determining whether the triggering master node needs to intervene in the determination unit of this embodiment;

[0057] Figure 3 This is a logic diagram for the distribution unit in this embodiment to determine the distribution process.

[0058] Figure 4 This is a schematic diagram of the adjustment module in this embodiment. Detailed Implementation

[0059] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0060] Please see Figure 1 As shown, this is a schematic diagram of the master-slave hierarchical OpenClaw terminal access and collaborative management system of this embodiment. This embodiment provides a master-slave hierarchical OpenClaw terminal access and collaborative management system, including:

[0061] The acquisition module is used to acquire in real time the interaction requests, task complexity, response latency, resource utilization, user level identifiers and permission tags of each child node in the public virtual environment, and to acquire the capability call success rate of the master node.

[0062] The determination module, which is connected to the acquisition module, is used to determine whether to trigger the main node intervention based on the task complexity and the preset complexity threshold, so as to obtain the task classification result.

[0063] An index determination module, which is connected to the acquisition module and the judgment module respectively, is used to determine the capability index of each sub-node based on the task classification result, the response latency, the resource utilization rate and the capability call success rate.

[0064] The update module is connected to the acquisition module and the index determination module respectively. It is used to determine the global capability distribution parameters based on each capability index and resource utilization rate, and to determine whether the evolution conditions of the master node are met based on the change range of the global capability distribution parameters. The global strategy parameters of the master node are updated based on the determination result that the evolution conditions are met.

[0065] The strength determination module, which is connected to the update module, is used to determine the capability output strength based on the call permission coefficient and the global capability distribution parameters, wherein the call permission coefficient is calculated based on the user level identifier, the permission tag and the task complexity;

[0066] The generation module, which is connected to the strength determination module, is used to distribute global capability distribution parameters or global strategy parameters to the corresponding child nodes according to the capability output strength, and to generate new child nodes according to the real-time acquired task request frequency and capability index.

[0067] The adjustment module, which is connected to the generation module, the strength determination module, and the judgment module, is used to adjust the preset complexity threshold or the call permission coefficient based on the response latency and task completion status after distribution.

[0068] In this embodiment, multiple ordinary users access the public virtual environment and complete identity registration through applications on their respective mobile terminals. Multiple sub-nodes have been deployed in the public virtual environment, and each sub-node establishes bidirectional communication with its corresponding ordinary user. An advanced user runs a client virtual environment on their local computing device and completes identity verification through user level identifiers and permission tags. The advanced user's terminal has not yet established a connection with any sub-nodes and is in a state of waiting for the master node to intervene. The master node is unique and deployed on the publisher's local computer. At this time, it is silently collecting data on the interaction requests, task complexity, response latency, and resource utilization of each sub-node in the public virtual environment. The entire system is in an initial steady state awaiting task triggering.

[0069] The acquisition module continuously monitors the interaction request messages written by each child node to the public data table through a real-time subscription interface of the SpacetimeDB database deployed in the public virtual environment. Each message carries a unique node identifier, the user terminal's level identifier and permission tag, and the estimated task complexity corresponding to the request. Simultaneously, the acquisition module collects the current response latency and resource utilization of each child node every 500 milliseconds using lightweight probes embedded within each child node. Response latency is defined as the time interval from receiving a user request to returning the first byte of the response, and resource utilization is the weighted sum of the node's CPU utilization and memory utilization. For the success rate of capability calls to the master node, the acquisition module calculates it in real-time by statistically analyzing the proportion of successful responses among the most recent 100 capability call results recorded in the public virtual environment. All the parameters collected in real-time are converted into timestamp key-value pairs in a unified format for subsequent module use.

[0070] The preset complexity threshold is a dimensionless complexity scoring threshold. It is set based on the following: during the initial operation of the public virtual environment, the acquisition module collects statistical relationships between the response latency and task complexity of all child nodes processing various tasks over a continuous 24-hour period. With constraints that the response latency of each child node does not exceed 2000 milliseconds and the capability call success rate is not less than 95%, the maximum complexity value that the set of child nodes can independently handle is calculated. 90% of this maximum complexity value is taken as the baseline threshold. Typically, this threshold is set between 0.6 and 0.9. In this embodiment, it is set to 0.75, which allows approximately 85% of regular user requests to be directly responded to by the child nodes without waking up the main node, while ensuring that high-difficulty tasks with a complexity exceeding 0.75 promptly trigger the main node's intervention to avoid child node response timeouts or processing failures.

[0071] The acquisition module collects real-time data on task complexity, response latency, resource utilization, and user level identifiers and permission tags for each child node. The judgment module classifies tasks based on a preset complexity threshold to determine whether the master node should intervene. The index determination module combines the task classification results with response latency, resource utilization, and the success rate of parent node capability invocation to calculate a capability index characterizing the real-time service level of each child node. The update module aggregates the capability indices and resource utilization of all child nodes to form a global capability distribution parameter. When the change in this distribution parameter exceeds a threshold, the parent node is deemed to meet the evolution conditions, thus updating its global strategy parameters. The more inefficient or differentiated the overall capability distribution of the child node group, the more the parent node needs to evolve to adjust its strategy. The evolved strategy parameters further influence the capability output strength calculated by the strength determination module. This strength is also constrained by the call permission coefficient, which is determined by the user level identifier. The system is determined by both permission tags and task complexity, ensuring that ordinary users only receive basic capabilities while advanced users receive more comprehensive capability feedback. The generation module selects and distributes global capability distribution parameters or global strategy parameters to corresponding child nodes based on the strength of capability output. It also automatically incubates new child nodes for high-frequency tasks with high request frequency and insufficient capability index. The adjustment module reversely adjusts the preset complexity threshold and call permission coefficient based on the actual observed response latency and task completion status after capability distribution. This allows the system to automatically lower the parent intervention threshold and relax the permission coefficient to call parent capabilities to alleviate bottlenecks when facing performance insufficiency, and raise the intervention threshold and tighten the permission coefficient when performance is excessive to avoid resource waste. This forms a complete closed-loop logic from data collection capability quantification group distribution evolution trigger intensity output performance feedback parameter adaptation, ultimately achieving efficient collaboration and adaptive evolution of the master-slave hierarchical intelligent agent system.

[0072] Please see Figure 2 As shown, this is the determination logic diagram of whether the triggering master node needs to intervene in the determination unit of this embodiment. In this embodiment, the determination module includes:

[0073] The offset calculation unit is used to calculate the complexity offset based on the difference between the task complexity and the preset complexity threshold, and to determine the task classification result based on the mapping relationship between the complexity offset and the preset classification interval.

[0074] The determination unit is used to determine whether to trigger the main node to intervene based on the comparison result that the complexity offset is greater than a preset trigger threshold.

[0075] In this embodiment, the determination unit is also used to determine, based on the comparison results that the complexity offset is less than or equal to the preset trigger threshold, that the trigger master node does not need to intervene, so as to maintain the independent execution of the child node.

[0076] In this embodiment, the offset calculation unit first calculates the difference between the task complexity and a preset complexity threshold to obtain the complexity offset, which ranges from negative to positive. The offset calculation unit internally presets three hierarchical intervals: the first interval corresponds to a complexity offset less than or equal to -0.1, the second interval corresponds to a complexity offset greater than -0.1 and less than 0.1, and the third interval corresponds to a complexity offset greater than or equal to 0.1. The task hierarchical results are then mapped to low-level, mid-level, and high-level tasks, respectively. The judgment unit further reads a preset trigger threshold, set to 0.05. When the complexity offset is greater than 0.05, the judgment unit outputs a judgment result that triggers the main node's intervention; when the complexity offset is less than or equal to 0.05, the judgment unit outputs a judgment result that the main node does not need to intervene, and the current child node executes the task independently. With the above settings, tasks with a complexity offset between -0.1 and 0.05 (including some intermediate tasks and all low-level tasks) will be handled by the child nodes themselves, while high-level tasks with an offset exceeding 0.05 and some intermediate-level tasks will trigger the parent node to intervene, thereby realizing the joint determination of task classification and parent node intervention conditions.

[0077] The preset grading interval is set based on the following: the acquisition module collects the average response latency and success rate of each sub-node when processing offsets of different complexity over a continuous 24-hour period. Statistical analysis shows that when the offset is less than or equal to -0.1, the average response latency of all tasks is less than 500 milliseconds and the success rate is 100%. When the offset is between -0.1 and 0.1, the response latency fluctuates between 500 and 1500 milliseconds and the success rate is not less than 98%. When the offset is greater than or equal to 0.1, the response latency exceeds 2000 milliseconds and the success rate drops below 92%. Therefore, -0.1 and 0.1 are used as the boundaries to distinguish between low-level, medium-level, and high-level tasks. Typically, the two boundary values ​​of the preset grading interval are set between -0.2 and -0.05 and between 0.05 and 0.2, respectively. In this embodiment, they are set to -0.1 and 0.1, which allows approximately 60% of low-complexity tasks to be classified as low-level, 30% of medium-complexity tasks as medium-level, and 10% of high-difficulty tasks as high-level, thus achieving a reasonable separation of task levels.

[0078] The preset trigger threshold is set based on the following: based on the classification results, the processing capacity of the child node when the offset is near the upper limit of the intermediate task range is used as the benchmark. The critical point where the proportion of child node response timeouts starts to exceed 5% when the complexity offset of the intermediate task exceeds 0.05 is selected as the trigger condition. Usually, the preset trigger threshold is set between 0.03 and 0.08. In this embodiment, it is set to 0.05. This allows the child node to handle the intermediate task entirely on its own when the offset is below 0.05, thus saving the parent system resources. When the offset exceeds 0.05, the parent system intervenes in time to avoid overloading the child node. At the same time, it ensures that all high-level tasks with an offset ≥ 0.1 will trigger the parent system to intervene, thus balancing system efficiency and service reliability.

[0079] The offset calculation unit subtracts the task complexity from a preset complexity threshold to obtain the complexity offset. Based on the level range of the offset, the task is divided into low-level, medium-level, or high-level. The judgment unit then compares the offset with the trigger threshold to determine whether to trigger the master node intervention. A negative offset with a larger absolute value indicates that the task complexity is much lower than the child node's normal processing capacity, and the child node can achieve low latency and high success rate service by executing independently. When the offset is within a certain negative range and does not exceed the trigger threshold, the child node is still in a reliable service range and can stably complete the task without the master node's intervention. A positive offset that continues to increase means that the task complexity exceeds the child node's independent processing capacity. The child node's response latency begins to increase significantly and the success rate decreases. At this time, the master node needs to intervene to maintain service quality. When the offset reaches a higher positive value, the child node is approaching the overload boundary and must be taken over by the master node to avoid processing failure. By linking the hierarchical intervals and trigger thresholds, frequent false triggers or missed triggers caused by a single threshold are avoided. The processing capability boundary of the child node is precisely aligned with the intervention time of the parent system. This ensures efficient autonomy for low-complexity tasks while ensuring timely support from upper-level capabilities for high-complexity tasks, thus achieving a refined and adaptive intervention strategy in the master-slave hierarchical system.

[0080] Specifically, the index determination module includes:

[0081] The weight determination unit is used to determine the corresponding task weight coefficient based on the task classification result when determining the triggering of the master node intervention;

[0082] The delay determination unit is used to determine the delay deviation value based on the response delay and the preset delay reference value when it is determined that the master node intervention is triggered;

[0083] The resource determination unit is used to determine the resource deviation value based on the resource occupancy rate and the preset resource occupancy benchmark value when it is determined that the master node intervention is triggered;

[0084] The index determination unit is used to determine the basic capability value based on the capability call success rate, the latency deviation value, and the resource deviation value, and modulate the basic capability value in combination with the task weight coefficient to obtain the capability index of each sub-node.

[0085] In this embodiment, after the judgment unit triggers the main node's intervention, the weight determination unit maps low-level tasks, medium-level tasks, and high-level tasks to task weight coefficients of 0.8, 1.0, and 1.2 respectively, based on the task classification results output by the offset calculation unit. High-level tasks are assigned higher weights because they are more dependent on the parent node's capabilities. The latency determination unit uses the median response latency of the child node over the past hour as a preset latency benchmark value, reads the response latency of the current sampling period, and calculates the latency deviation value, which is equal to the current response latency divided by the latency benchmark value. A value greater than 1 indicates a slower response, and a value less than 1 indicates a faster response. The resource determination unit uses the resource utilization rate of the child node over the past hour... The average value is the preset resource utilization benchmark value. The current resource utilization rate is read, and the resource deviation value is calculated as the current resource utilization rate divided by the resource utilization benchmark value. A value greater than 1 indicates increased load, and a value less than 1 indicates decreased load. The index determination unit calculates the basic capability value, which is equal to the capability call success rate divided by the product of the latency deviation value and the resource deviation value. Both the latency deviation value and the resource deviation value are not less than 0.5, and the capability call success rate is between 0 and 1. Then, the basic capability value is multiplied by the task weight coefficient to obtain the capability index of the sub-node. The higher the capability index value, the stronger the comprehensive service capability of the sub-node under the current task level and load state.

[0086] Through collaborative calculations by the weight determination unit, latency determination unit, resource determination unit, and index determination unit, the task classification results are mapped to weight coefficients. The latency deviation relative to the baseline, the resource occupancy deviation relative to the baseline, and the capability invocation success rate of the master node are all incorporated into the calculation of the basic capability value. A higher capability invocation success rate results in a larger basic capability value, while a larger latency deviation (slower response) or a larger resource deviation (heavier load) decreases the basic capability value. The inherent logic of this ratio is that the quality of the parent node's capability support directly improves the service level of the child node, but response latency and resource congestion weaken service quality; the two are mutually restrictive. Based on this, a task weight coefficient is introduced, allowing higher-level tasks to receive higher weights due to their stronger reliance on the parent node, thus amplifying the sensitivity to response speed and load conditions in the capability index. The calculated capability index value accurately characterizes the comprehensive service capability of a child node when handling the current classification task, considering the parent node's support quality, its own response efficiency, and load pressure. This provides a reliable basis for the quantification of child node capabilities and subsequent scheduling and allocation in the master-slave hierarchical system.

[0087] Specifically, the update module includes:

[0088] A distributed building unit, which is used to construct a capability distribution sequence based on the capability index and resource utilization of each child node;

[0089] A parameter extraction unit is used to extract global capability distribution parameters from the capability distribution sequence;

[0090] The change determination unit is used to determine the distribution change amplitude based on the change trend of the global capability distribution parameters within a preset change determination time period;

[0091] An evolution determination unit is used to determine whether the evolution conditions of the master node are met based on the comparison between the distribution change amplitude and the preset evolution threshold.

[0092] The strategy update unit is used to update the global strategy parameters of the master node when the evolution conditions are met.

[0093] Specifically, the distributed building unit includes:

[0094] The sorting subunit is used to sort the capability indices of each child node according to their numerical values ​​to form an ordered capability sequence.

[0095] A difference calculation subunit is used to calculate a capability difference sequence based on the difference between adjacent capability indices in the ordered capability sequence.

[0096] A normalization subunit is used to normalize the capability difference sequence based on the resource occupancy rate of each child node to obtain the capability distribution sequence.

[0097] In this embodiment, the distribution construction unit first obtains the capability indices of all child nodes from the index determination module. The sorting subunit sorts these capability indices from smallest to largest to obtain an ordered capability sequence. The difference calculation subunit traverses each pair of adjacent elements in the ordered capability sequence, calculates the difference between the previous capability index and the next capability index, and obtains a capability difference sequence, which reflects the density of the capability index distribution in the sorting space. The normalization subunit reads the current resource utilization rate of each child node, divides each difference in the capability difference sequence by the resource utilization rate of the next child node corresponding to that difference. The higher the resource utilization rate, the more the node is currently under high load, and the weaker the representativeness of its capability difference to the overall group capability distribution. Therefore, normalization suppresses the abnormal pull of high-load nodes on the capability distribution sequence, so that the contribution of nodes with higher resource utilization rates in the distribution sequence is compressed, thereby obtaining a load-weighted capability distribution sequence. The parameter extraction unit extracts a scalar value from the capability distribution sequence as a global capability distribution parameter. In this embodiment, the median of the capability distribution sequence is used, which represents the center position of the capability distribution of the current child node group after load weighting. The change determination unit continuously records the global capability distribution parameter extracted at each sampling time within a sliding window of a preset change determination duration. It calculates the absolute value of the change in parameter value at the end of the window relative to the parameter value at the beginning of the window, and then divides it by the parameter value at the beginning of the window to obtain the relative change amplitude, which is used as the distribution change amplitude. The evolution determination unit compares the distribution change amplitude with a preset evolution threshold. If the distribution change amplitude is greater than or equal to the preset evolution threshold, it is determined that the evolution condition of the master node is met, indicating that the overall capability level of the child node group has significantly migrated, and the parent needs to adapt to the new distribution state; otherwise, the evolution condition is not met. When the evolutionary conditions are met, the policy update unit reads the current global policy parameters of the master node from the public virtual environment, including the weights of the deep reinforcement learning policy network or the thresholds of the decision tree nodes. It then uses a proximal policy optimization algorithm to perform one or more gradient updates on the global policy parameters based on the child node interaction samples collected recently. The updated parameters are then stored back in the public virtual environment for subsequent capability output. The system can automatically perceive the migration of the overall capability level of the child node group and drive the parent agent to evolve at appropriate times, thereby maintaining its leadership over the entire cluster.

[0098] The preset change determination duration depends on the balance between the system's sensitivity to changes in capability distribution and sampling stability. The rationale is that too short a duration can cause drastic jumps in the variance and Euclidean distance of the capability distribution sequence due to fluctuations in a single task or instantaneous load spikes, leading to frequent false triggers of evolutionary conditions and wasting the parent system's computing resources. Conversely, too long a duration will delay the detection of the true differentiation of the capability distribution among child nodes, causing the parent system's evolution to lag behind actual demand changes. Experimental tests, with 10 to 50 child nodes and a sampling period of 500 milliseconds, revealed that step response tests with different duration windows showed a false trigger rate exceeding 25% for durations below 3 minutes, and a detection delay exceeding 15 minutes for durations above 8 minutes. A 5-minute window reduced the false trigger rate to below 5% and controlled the detection delay to around 7 minutes. Typically, the preset change determination duration is set between 3 and 8 minutes; in this embodiment, it is set to 5 minutes, effectively filtering out short-term random fluctuations while promptly capturing substantial shifts in capability distribution.

[0099] The preset evolution threshold depends on a trade-off between the system's tolerance for changes in capability distribution and the overhead of parent evolution. It is set based on the historical Euclidean distance changes in the capability distribution sequence during steady-state operation of the system, statistically analyzing the upper limit of normal fluctuations under conditions of no external load changes, and using 1.5 times this upper limit as the minimum amplitude for triggering evolution to avoid unnecessary evolution caused by normal fluctuations. Simultaneously, the average time and computational cost of a complete parent policy update are considered. A threshold that is too low will lead to excessively frequent evolutions and unconverged oscillations in the parent policy; a threshold that is too high will cause the capability index variance to continuously expand without triggering evolution, resulting in excessive differences in service levels among child nodes. Offline simulations, testing the evolution threshold in steps between 0.1 and 0.6, revealed that when the threshold is less than 0.2, the number of evolutions exceeds 12 per hour, and the parent policy exhibits significant oscillations; when the threshold is greater than 0.5, the capability index variance increases to over 0.4 without triggering evolution, and some child nodes remain in a low-capability state for extended periods; when the threshold is 0.3, the evolution frequency drops to 3 to 4 times per hour, and the population variance of the capability index is consistently controlled within 0.25. Typically, the preset evolution threshold is set between 0.2 and 0.5. In this embodiment, it is set to 0.3, which can control the update cost of the parent node while ensuring a relatively balanced distribution of the capabilities of the child node group, thus achieving the optimal selection of the evolution timing.

[0100] The sorting subunit arranges the capability indices of child nodes from smallest to largest to form an ordered capability sequence. The difference calculation subunit calculates the difference between adjacent indices to obtain a capability difference sequence. Then, the normalization subunit weights and compresses the differences based on the resource utilization rate of each node, so that the node with the higher resource utilization rate contributes less to the capability difference in the distribution sequence. The capability distribution sequence constructed in this way can reflect both the absolute difference of capability indices and the inhibitory effect of load pressure on actual service capacity. The parameter extraction unit takes the median of this sequence as the global capability distribution parameter. This median represents the capability of the child node group after load modulation. The force distribution center location is determined by the change determination unit, which uses the relative change of the median within a preset change determination time window as the distribution change range. The underlying logic is that when the ability index of most child nodes in the group systematically decreases or increases due to the lag in parent evolution, the median will shift significantly accordingly. Conversely, in steady state, the median only fluctuates within a small range. The evolution determination unit compares this relative change range with a preset evolution threshold. If the threshold is exceeded, it indicates that the group's ability center has substantially shifted. At this point, the parent's original strategy is no longer suitable for the current distribution state, and the strategy update unit triggers parent evolution to update the global strategy parameters. This allows the timing of parent evolution to precisely align with statistically significant changes in the overall ability level of the child node group, avoiding ineffective evolution caused by random fluctuations and ensuring that the parent can adjust its leading strategy in a timely manner when the group's ability shifts substantially.

[0101] Specifically, the intensity determination module includes:

[0102] A permission determination unit is used to determine a call permission coefficient based on the user level identifier, the permission tag, and the task complexity.

[0103] The gap determination unit is used to determine the capability gap value based on the difference between the capability index of each child node and the global capability distribution parameter;

[0104] The strength calculation unit is used to determine the capability output strength based on the call permission coefficient, the capability gap value, and the global capability distribution parameters.

[0105] Specifically, the strength calculation unit includes:

[0106] A mapping subunit is used to determine the gap level based on the mapping relationship between the capacity gap value and a preset gap range;

[0107] A constraint subunit is used to constrain the gap level according to the call permission coefficient to obtain the available capability level;

[0108] An adjustment subunit is used to adjust the available capability level according to the global capability distribution parameters to obtain the capability output intensity.

[0109] In this embodiment, the permission determination unit first reads the user level identifier, permission tag, and current task complexity of the user terminal from the acquisition module. The user level identifier is divided into ordinary user and advanced user, and the permission tag is divided into no local operation permission and local operation permission. The permission determination unit sets the baseline permission coefficient to 0.6. If the user level is advanced user, it is increased by 0.2; if the user has local operation permission, it is increased by another 0.2. Then, it is adjusted according to the task complexity: for every 0.1 increase in task complexity, the permission coefficient increases by 0.05, and the final value does not exceed 1.0. The specific formula is: the permission coefficient is equal to 0.6 plus the user level increment plus the permission increment plus the task complexity adjustment amount, and the smaller value is taken by 1.0. The task complexity adjustment amount is the task complexity minus 0.5, divided by 10, and then the smaller value is taken by 0.3.

[0110] The gap determination unit obtains the capability index of each child node from the index determination module and the global capability distribution parameter from the update module. This parameter is defined as the median of the capability distribution sequence. The gap determination unit calculates the capability gap value of each child node as equal to the global capability distribution parameter minus the node's capability index. If the result is negative, it is set to zero. That is, the capability gap value is the larger of the difference between the global distribution parameter and the capability index and zero. The larger this value, the greater the lag of the node relative to the median capability of the group.

[0111] The strength calculation unit internally comprises a mapping subunit, a constraint subunit, and an adjustment subunit. The mapping subunit determines the gap level based on the preset gap range in which the capacity gap value falls. In this embodiment, three gap ranges are preset: a capacity gap value less than 0.1 corresponds to a low gap range and gap level 1; a capacity gap value between 0.1 and 0.3 corresponds to a medium gap range and gap level 2; and a capacity gap value greater than 0.3 corresponds to a high gap range and gap level 3. The mapping subunit maps the capacity gap value of each sub-node to the corresponding level value 1, 2, or 3.

[0112] The constraint subunit reads the call permission coefficient and uses this coefficient to constrain the gap level to obtain the available capability level. The constraint method is to multiply the gap level by the call permission coefficient, round down, and then take the smaller value between the result and the gap level itself, while ensuring that the available capability level is not lower than 1. For example, if the gap level is 3 and the call permission coefficient is 0.6, then 3 multiplied by 0.6 equals 1.8, rounded down to 1, and the available capability level is 1; if the call permission coefficient is 1.0, then the available capability level is equal to the gap level. This constraint process ensures that even if the gap in a child node is large, a lower capability output level can only be obtained when the user's permissions are insufficient.

[0113] The adjustment subunit adjusts the available capability level based on the global capability distribution parameter to obtain the final capability output strength. In this embodiment, the global capability distribution parameter is the median of the capability distribution sequence, with a value ranging from 0.5 to 1.5. The adjustment method involves multiplying the available capability level by the global capability distribution parameter, dividing by the maximum possible gap level of 3, and then taking the smaller value between 3 and 1.0 to normalize the output strength to between 0 and 1. For example, if the available capability level is 2 and the global capability distribution parameter is 1.2, then the output strength equals 2 multiplied by 1.2 divided by 3, which equals 0.8. This adjustment amplifies the output strength corresponding to the same available capability level when the overall capability level of the group is high, and compresses it when it is low, thereby achieving dynamic adjustment of the actual output based on the group baseline.

[0114] The strength calculation unit sends the final capability output strength to the generation module, which then decides which parameters to distribute and whether to incubate new child nodes. Through the above three-level serial processing, the capability output strength is simultaneously and finely controlled by three factors: the size of the gap in the child node itself, the level of user permission, and the overall capability level of the group.

[0115] The permission determination unit calculates a call permission coefficient based on user level identifiers, permission tags, and task complexity. The value of this coefficient directly reflects the upper limit of the capabilities a user is allowed to acquire from the parent system; higher levels and more complex tasks result in a larger upper limit. The gap determination unit calculates the capability gap value based on the median capability of the group. A larger value indicates a more severe lag of the child node relative to the group's average level, and a more urgent need to supplement the parent system's capabilities. The mapping subunit discretizes continuous capability gap values ​​into finite gap levels. The constraint subunit uses the call permission coefficient to multiply and truncate these levels, ensuring that even with a large gap, a child node cannot obtain a high-level output if user permissions are insufficient. User authorization is a prerequisite for capability output; even a large gap cannot exceed the permission limit. The adjustment subunit then scales the available capability levels using global capability distribution parameters. When the overall group level is high, the output intensity corresponding to the same available level is amplified; when the overall group level is low, it is compressed. This ensures that the output intensity considers not only individual gaps and user permissions but also the current baseline state of the cluster. The resulting capability output strength is highest when user permissions are sufficient, child node gaps are significant, and the group is not at an overall low level. Conversely, the strength automatically decreases when any of these conditions are not met, thus achieving coordinated constraints among user-side authorization, individual-side needs, and cluster-side benchmarks.

[0116] Please see Figure 3 As shown, this is a logic diagram of the distribution unit's distribution determination in this embodiment. In this embodiment, the generation module includes:

[0117] A distribution unit is used to distribute the global capability distribution parameters based on a comparison result where the capability output strength is less than a preset strength threshold, and to distribute the global strategy parameters based on a comparison result where the capability output strength is greater than or equal to the preset strength threshold.

[0118] The frequency determination unit is used to obtain the task request frequency of various tasks within a preset statistical time period in real time, and to determine the high-frequency task set based on the processing tasks corresponding to the task request frequency greater than the preset frequency threshold.

[0119] A gap determination unit is used to determine the capability gap value based on the difference between the capability index corresponding to the high-frequency task set and the global capability distribution parameter;

[0120] A generation unit is used to generate new child nodes based on the comparison result between the capability gap value and the preset gap threshold, wherein the new child nodes are initialized based on the global policy parameters.

[0121] In this embodiment, the distribution unit receives the capability output strength from the strength determination module and compares it with a preset strength threshold. If the capability output strength is less than the preset strength threshold, the global capability distribution parameter, i.e., the median of the capability distribution sequence, is distributed to the corresponding child node for lightweight adjustment of the node's capability baseline. If the capability output strength is greater than or equal to the preset strength threshold, the updated global policy parameters of the master node, including the weights of the deep reinforcement learning policy network or the thresholds of decision tree nodes, are distributed for deep capability upgrades of the node. The frequency determination unit uses a preset statistical duration as a window to count the occurrence frequency of various processing tasks in real time and calculates the request frequency per minute. Task types with frequencies exceeding a preset frequency threshold are classified into a high-frequency task set. The gap determination unit queries the capability index of the child nodes processing that type of task for each type of task in the high-frequency task set, calculates the difference between the capability index and the global capability distribution parameter, and takes the maximum value of the difference as the capability gap value for that type of task. The larger the value, the more insufficient the existing child node's capability in processing that type of high-frequency task is relative to the median level of the group. The generation unit compares the capability gap value with a preset gap threshold. If the capability gap value is greater than the preset gap threshold, it determines that a new child node needs to be incubated for the corresponding high-frequency task. The generation unit reads the current global policy parameters of the master node from the public virtual environment, uses the parameters to assign values ​​to the initial policy network of the new child node, and registers the new node to the public virtual environment to start service. It can automatically identify high-frequency task types with insufficient capabilities and dynamically expand child nodes as needed. At the same time, it selects lightweight parameter distribution or heavyweight policy distribution according to different capability output strengths to achieve differentiated capability feedback.

[0122] The preset strength threshold depends on the physical meaning of the capability output strength and its impact on system resource consumption. The setting is based on experimental testing of the differences in system performance between distributing global policy parameters and global capability distribution parameters under different capability output strengths. Statistical analysis revealed that when the capability output strength is below 0.4, the computational load and network transmission overhead caused by distributing global policy parameters far outweigh the performance gains, while distributing global capability distribution parameters is sufficient to meet the capability supplementation needs of child nodes. When the capability output strength is above 0.6, distributing only global capability distribution parameters cannot effectively reduce the capability gap; global policy parameters must be distributed to significantly improve the service level of child nodes. Step tests between 0.4 and 0.6 showed that when 0.5 is used as the dividing point, the mismatch rate of both types of distribution is below 8%, and the overall resource consumption is minimized. Typically, the preset strength threshold is set between 0.4 and 0.6; in this embodiment, it is set to 0.5, which achieves the optimal switching point for using lightweight parameter distribution to save resources when the capability output strength is below this value and heavyweight policy distribution to ensure performance improvement when it is above this value.

[0123] The preset frequency threshold depends on the balance between the system's sensitivity to hot tasks and the incubation overhead. The setting is based on collecting the request frequency distribution of various tasks during normal system operation, statistically analyzing the long-term mean and standard deviation of task request frequencies within a natural week, and using the mean plus one standard deviation as the initial boundary to distinguish between regular and high-frequency tasks. Offline simulations were conducted to test the benefits of incubating new child nodes at different frequency thresholds ranging from 2 to 10 times per minute. It was found that when the threshold was below 3 times per minute, the excessive number of incubated nodes led to a resource occupancy rate increase of over 30%, and most new nodes remained idle for extended periods. When the threshold was above 8 times per minute, some high-frequency tasks experienced excessive response latency due to delayed incubation, resulting in decreased user satisfaction. Typically, the preset frequency threshold is set between 3 and 8 times per minute; in this embodiment, it is set to 5 times per minute, enabling the system to identify and selectively incubate hot tasks accounting for approximately 20% of the total requests, while avoiding resource waste caused by over-incubation.

[0124] The preset gap threshold depends on a trade-off between tolerance for insufficient capabilities of child nodes and incubation costs. It is set based on analyzing the statistical distribution of capability gap values, using the critical gap value at which the success rate of child nodes in handling tasks begins to decline significantly as a reference. Experiments show that when the capability gap value is below 0.2, child nodes, although slightly below the median level of the group, can still maintain a task success rate of over 90%; when the capability gap value exceeds 0.3, the task success rate drops below 80%, and the user complaint rate increases significantly. Comparing the performance improvement after incubation at different gap thresholds, when the threshold is below 0.2, the incubation benefit is small while the cost remains unchanged, resulting in low cost-effectiveness; when the threshold is above 0.35, high-frequency tasks have been at a low service level for a long time before incubation. Typically, the preset gap threshold is set between 0.2 and 0.35; in this embodiment, it is set to 0.25, which ensures that new nodes are incubated in a timely manner when child nodes lag behind the median level of the group to a certain extent, and the capability improvement effect of the incubated nodes is the most significant.

[0125] The preset statistical duration depends on the balance between the system's stability in calculating task request frequency and its sensitivity to hotspot responses. The setting is based on the following: if the statistical duration is too short, for example, less than 30 minutes, random bursts of requests within a single time period will be misclassified as high-frequency tasks, leading to frequent incubation of new child nodes, wasting resources, and new nodes remaining idle for extended periods after the bursts. If the statistical duration is too long, for example, more than 2 hours, morning peak tasks will be diluted by the statistical window including afternoon off-peak periods, failing to be identified as high-frequency tasks, resulting in the incubated hotspot tasks not receiving timely capacity replenishment. By comparing the accuracy of task frequency identification and the effective utilization rate of incubated nodes under different durations, tests showed that the misclassification rate was approximately 20% under a 30-minute window, the missed detection rate was approximately 25% under a 2-hour window, while a 1-hour window can cover at least one complete small business peak cycle, keeping both the misclassification and missed detection rates below 12%. Typically, the preset statistical duration is set between 30 minutes and 2 hours. In this embodiment, it is set to 1 hour, which can obtain stable frequency statistical results and capture substantial changes in the popularity of task requests in a timely manner, ensuring that the identification of high-frequency task sets is neither affected by short-term fluctuations nor fails due to averaging.

[0126] The distribution unit selects to distribute lightweight global capability distribution parameters or heavyweight global strategy parameters based on the comparison between the capability output strength and the preset strength threshold. When the strength is below the threshold, the child nodes only need to be adjusted moderately, and the cost of distributing the complete strategy exceeds the benefits. When the strength is above the threshold, only strategy-level updates can effectively make up for the capability gap. The frequency determination unit selects a set of high-frequency tasks based on the preset frequency threshold. If the threshold is too low, it will lead to excessive incubation and waste of resources. If the threshold is too high, it will lead to the accumulation of hot tasks and a decline in service quality. The gap determination unit calculates the difference between the existing child node capability index and the global median capability for high-frequency tasks as the capability gap value. This difference directly quantifies the degree of collective capability deficiency of the group for this type of task. The generation unit compares this gap value with the preset gap threshold. Only when the gap exceeds the threshold will a new node be incubated and initialized with global strategy parameters, so that the capability starting point of the new node is higher than the current group median level. The three thresholds control resource consumption, hotspot identification sensitivity, and capability insufficiency tolerance, respectively. Through their synergistic effect, the system avoids blind expansion when the task is not popular or the capability gap is not large. When high-frequency tasks appear and the existing node capabilities are significantly behind the group benchmark, targeted nodes are incubated in a timely manner, achieving a precise match between expansion timing and task demand intensity and capability defect depth.

[0127] Please see Figure 4 As shown, this is a schematic diagram of the adjustment module in this embodiment. In this embodiment, the adjustment module includes:

[0128] The preprocessing unit is used to sample the response delay and task completion status after capacity distribution within a preset observation period, and to standardize the sampling results to obtain a standardized delay sequence and a standardized completion sequence, respectively.

[0129] The observation and determination unit is used to determine the time delay fluctuation characteristic value based on the standardized time delay sequence, and to determine the completion stability characteristic value based on the standardized completion sequence.

[0130] The indicator calculation unit is used to determine the short-cycle performance indicator value based on the time delay fluctuation characteristic value and the completion stability characteristic value, and to determine the comprehensive performance indicator value according to the changing trend of all cycle performance indicator values ​​within the preset adjustment period.

[0131] An adjustment unit is used to adjust the preset complexity threshold or the call permission coefficient based on the comprehensive performance index value.

[0132] Specifically, the adjustment unit includes:

[0133] The interval determination subunit is used to divide the comprehensive performance index value into a performance insufficiency interval, a performance stability interval, or a performance excess interval based on the comparison results between the comprehensive performance index value and the preset first index threshold and the preset second index threshold.

[0134] A threshold adjustment subunit is used to reduce the preset complexity threshold when the comprehensive performance index value is in the performance insufficiency range, and to increase the preset complexity threshold when the comprehensive performance index value is in the performance excess range.

[0135] The license adjustment subunit is configured to increase the call license coefficient when the overall performance index value is in the underperformance range, and to decrease the call license coefficient when the overall performance index value is in the overperformance range.

[0136] In this embodiment, the preprocessing unit uses a preset observation period as a time window, within which it samples the response latency and task completion status of child nodes after each capability distribution. Response latency is defined as the time interval from the completion of distribution to the child node returning the task execution result. Task completion status is represented by a value of 1 for success and 0 for failure. After collecting the raw data of all sampling points within the window, the preprocessing unit uses the Z-score standardization method to process the latency data and completion data respectively: for latency data, the mean and standard deviation of all response latency within the window are calculated, and the latency of each sampling point is subtracted from the mean and then divided by the standard deviation to obtain a standardized latency sequence; for completion data, the mean and standard deviation are calculated similarly and then standardized to obtain a standardized completion sequence.

[0137] The observation and determination unit calculates the time delay fluctuation characteristic value based on the standardized time delay sequence, specifically using the standard deviation of the sequence as the fluctuation characteristic value. The larger the standard deviation, the more unstable the response time delay. The completion stability characteristic value is calculated based on the standardized completion sequence, specifically by dividing the sum of the absolute values ​​of the differences between adjacent sampling points in the sequence by the number of sampling points. The larger this value, the more violent the oscillation of task completion over time, i.e., the worse the stability.

[0138] The performance indicator calculation unit calculates short-cycle performance indicator values ​​based on latency fluctuation characteristic values ​​and completion stability characteristic values. The calculation formula is that the short-cycle performance indicator value equals the weighted sum of the latency fluctuation characteristic value and the completion stability characteristic value, where the latency fluctuation characteristic value has a weight of 0.6 and the completion stability characteristic value has a weight of 0.4, to highlight the dominant impact of response latency fluctuations on user experience. The performance indicator calculation unit stores the short-cycle performance indicator values ​​calculated for each observation period in the historical record, and collects all short-cycle performance indicator values ​​within a preset adjustment period. The average of these indicator values ​​is calculated as the comprehensive performance indicator value, which comprehensively reflects the overall service quality of the system over the past thirty minutes.

[0139] The interval determination subunit in the adjustment unit reads the preset first indicator threshold and the preset second indicator threshold, and compares the comprehensive performance index value with these two thresholds: if the comprehensive performance index value is less than the preset first indicator threshold, it is determined to be in the performance excess interval, indicating that the system resources are sufficient and the service quality far exceeds the demand; if the comprehensive performance index value is between the preset first indicator threshold and the preset second indicator threshold, it is determined to be in the performance stable interval, indicating that the service quality is within the ideal range; if the comprehensive performance index value is greater than the preset second indicator threshold, it is determined to be in the performance insufficient interval, indicating that the service quality has deteriorated and cannot meet the demand.

[0140] When the overall performance index value is in the underperformance range, the threshold adjustment subunit will reduce the preset complexity threshold by 10%, but not lower than the lower limit of 0.6, so that more tasks trigger the main node to intervene to alleviate the pressure on the child nodes; when the overall performance index value is in the performance excess range, the preset complexity threshold will be increased by 10%, but not exceeding the upper limit of 0.9, so that more tasks are handled independently by the child nodes to release the parent resources; no adjustment is made when the performance is in the stable range.

[0141] When the overall performance index value is in the underperformance range, the permission adjustment subunit will increase the call permission coefficient by 10% but not exceed the upper limit of 1.0, so that users can obtain higher access rights to the parent capability to improve service quality; when the overall performance index value is in the overperformance range, the call permission coefficient will be reduced by 10% but not lower than the lower limit of 0.3 to avoid unnecessary output of the parent capability and waste of resources; no adjustment will be made when the performance is in the stable range.

[0142] The preset observation period depends on the balance between the system's sampling resolution for short-term performance changes and computational overhead. The rationale is as follows: if the observation period is too short, for example, less than 1 minute, the insufficient number of sampling points leads to unstable statistical characteristics of the standardized latency sequence, and the calculation of latency fluctuation characteristics and completion stability characteristics is severely affected by random noise, easily resulting in misjudgments; if the observation period is too long, for example, more than 10 minutes, it is impossible to detect short-term performance fluctuations in a timely manner, leading to delayed adjustment responses. Through testing the performance of observation periods ranging from 2 to 8 minutes under different load conditions, it was found that a period of 5 minutes ensures that each window contains enough sampling points to make Z-score standardization statistically meaningful, while also keeping the performance detection latency within an acceptable range. Typically, the preset observation period is set between 3 and 7 minutes; in this embodiment, it is set to 5 minutes, which can stably capture short-term fluctuations in response latency and task completion with a suitable sampling granularity.

[0143] The preset adjustment period depends on a trade-off between the system's sensitivity to tracking long-term performance trends and the stability of parameter adjustments. The rationale is as follows: a too-short adjustment period, such as 10 minutes, will cause the overall performance index to be excessively affected by occasional fluctuations in a single observation period, resulting in frequent oscillations of the preset complexity threshold and call permission coefficients, preventing the system from converging to a stable state; a too-long adjustment period, such as 60 minutes, will result in a slow response when performance continues to deteriorate, with users experiencing an excessively long period of service degradation. Simulation tests revealed that a 20-minute adjustment period resulted in the smallest parameter oscillation amplitude, while a 40-minute adjustment period resulted in an excessively long response time for deterioration. A 30-minute period can trigger adjustments within two observation periods after performance degradation and effectively smooth out short-term fluctuations. Typically, the preset adjustment period is set between 20 and 40 minutes; in this embodiment, it is set to 30 minutes, allowing the overall performance index to smooth out long-term trends based on the average of six observation periods, ensuring that the adjustment of thresholds and coefficients is neither too sensitive nor too sluggish.

[0144] The preset first performance threshold depends on the balance between the system's definition of performance excess and the economic efficiency of resource utilization. The setting is based on: measuring the distribution of comprehensive performance index values ​​in a simulated environment under ideal operating conditions, using the 5th percentile of the steady-state performance index values ​​as the trigger boundary for performance excess. Experiments show that when the comprehensive performance index value is below 0.4, the average system resource utilization is less than 60%, the response latency is far below the design limit, and the task completion rate fluctuation is almost zero, indicating that the system has a large amount of idle resources, which is a clear performance excess state; when the index value is between 0.4 and 0.8, the resource utilization is between 60% and 85%, and the response latency and task completion rate are both within acceptable ranges, representing a stable performance range; when the index value is above 0.8, it enters the performance deficiency range. Therefore, using 0.4 as the first performance threshold can clearly distinguish the boundary between excessive system resources and normal operation. Typically, the first indicator threshold is set between 0.3 and 0.5. In this embodiment, it is set to 0.4, which ensures that the complexity threshold is increased in time and the call permission coefficient is reduced when there is excess performance. This allows redundant tasks to be handled independently by child nodes to release parent resources and avoid unnecessary waste of computing power caused by parent intervention.

[0145] The preset second indicator threshold depends on the system's tolerance boundary for insufficient service quality and the need to differentiate between performance excess and insufficient performance. The setting is based on: measuring the distribution of comprehensive performance indicator values ​​in a simulated environment under ideal operating conditions, and taking the 95th percentile of the indicator values ​​under steady-state operation as the trigger boundary for performance deficiency. Experiments show that when the comprehensive performance indicator value is below 0.4, the system resource utilization is less than 60% and the user response latency is far below the indicator upper limit, indicating performance excess; when the indicator value is between 0.4 and 0.8, both resource utilization and response latency are within the design target range, representing a stable performance range; when the indicator value exceeds 0.8, the number of response latency exceedances increases and the task failure rate rises significantly, and users begin to perceive a decline in service quality. Therefore, using 0.8 as the second indicator threshold can clearly distinguish between tolerable fluctuations and performance deficiencies requiring intervention. Typically, the preset second indicator threshold is set between 0.7 and 0.9; in this embodiment, it is set to 0.8, ensuring that the complexity threshold is reduced and the permission coefficient is increased before the service quality deteriorates to a user-perceptible critical point, thereby promptly invoking the parent system's capabilities to restore service quality.

[0146] The adjustment range of the threshold adjustment subunit and the permission adjustment subunit is set based on sensitivity tests to observe the system's convergence speed and stability at different adjustment step sizes: when the step size is less than 5%, it takes more than five adjustment cycles to adjust from the underperforming range to the stable range, and the slow response leads to long-term deterioration of service quality; when the step size is greater than 15%, a single adjustment is too large and the system may skip the stable range and directly enter the opposite performance range, causing the preset complexity threshold and call permission coefficient to oscillate back and forth. Tests with step sizes between 5% and 15% revealed that a step size of 10% can bring the comprehensive performance index value back from the underperforming range to the stable range within two to three adjustment cycles, with an overshoot probability of less than 5%. It also takes into account the synergistic effect of threshold adjustment and permission coefficient adjustment, meaning that when both are adjusted synchronously in the same direction, the system response speed is moderate and the convergence is smooth. In this embodiment, the adjustment range is set to 10%, which ensures rapid relief of bottlenecks when performance is insufficient and gentle resource recovery when performance is excessive, achieving an adaptive balance of system performance.

[0147] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A master-slave hierarchical OpenClaw terminal access and collaborative management system, characterized in that, include: The acquisition module is used to acquire in real time the interaction requests, task complexity, response latency, resource utilization, user level identifiers and permission tags of each child node in the public virtual environment, and to acquire the capability call success rate of the master node. The determination module is used to determine whether to trigger the main node intervention based on the task complexity and the preset complexity threshold, so as to obtain the task classification result. The index determination module is used to determine the capability index of each sub-node based on the task classification results, the response latency, the resource utilization rate, and the capability call success rate. The update module is used to determine the global capability distribution parameters based on each capability index and resource utilization rate, and to determine whether the evolution conditions of the master node are met based on the change range of the global capability distribution parameters, so as to update the global strategy parameters of the master node based on the determination result of meeting the evolution conditions. The strength determination module is used to determine the capability output strength based on the call permission coefficient and the global capability distribution parameters, wherein the call permission coefficient is calculated based on the user level identifier, the permission tag and the task complexity; The generation module is used to distribute global capability distribution parameters or global strategy parameters to the corresponding child nodes according to the capability output strength, and to generate new child nodes based on the real-time acquired task request frequency and capability index. The adjustment module is used to adjust the preset complexity threshold or the call permission coefficient based on the response latency and task completion status after distribution.

2. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 1, characterized in that, The determination module includes: An offset calculation unit is used to calculate a complexity offset based on the difference between the task complexity and a preset complexity threshold, and to determine the task classification result based on the mapping relationship between the complexity offset and a preset classification interval. The determination unit is used to determine whether to trigger the main node to intervene based on the comparison result that the complexity offset is greater than a preset trigger threshold.

3. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 2, characterized in that, The index determination module includes: The weight determination unit is used to determine the corresponding task weight coefficient based on the task classification result when determining the triggering of the master node intervention; The delay determination unit is used to determine the delay deviation value based on the response delay and the preset delay reference value when it is determined that the master node intervention is triggered; The resource determination unit is used to determine the resource deviation value based on the resource occupancy rate and the preset resource occupancy benchmark value when it is determined that the master node intervention is triggered; The index determination unit is used to determine the basic capability value based on the capability call success rate, the latency deviation value, and the resource deviation value, and modulate the basic capability value in combination with the task weight coefficient to obtain the capability index of each sub-node.

4. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 3, characterized in that, The update module includes: A distributed building unit, which is used to construct a capability distribution sequence based on the capability index and resource utilization of each child node; A parameter extraction unit is used to extract global capability distribution parameters from the capability distribution sequence; The change determination unit is used to determine the distribution change amplitude based on the change trend of the global capability distribution parameters within a preset change determination time period; An evolution determination unit is used to determine whether the evolution conditions of the master node are met based on the comparison between the distribution change amplitude and the preset evolution threshold. The strategy update unit is used to update the global strategy parameters of the master node when the evolution conditions are met.

5. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 4, characterized in that, The distributed construction unit includes: The sorting subunit is used to sort the capability indices of each child node according to their numerical values ​​to form an ordered capability sequence. A difference calculation subunit is used to calculate a capability difference sequence based on the difference between adjacent capability indices in the ordered capability sequence. A normalization subunit is used to normalize the capability difference sequence based on the resource occupancy rate of each child node to obtain the capability distribution sequence.

6. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 5, characterized in that, The intensity determination module includes: A permission determination unit is used to determine a call permission coefficient based on the user level identifier, the permission tag, and the task complexity. The gap determination unit is used to determine the capability gap value based on the difference between the capability index of each child node and the global capability distribution parameter; The strength calculation unit is used to determine the capability output strength based on the call permission coefficient, the capability gap value, and the global capability distribution parameters.

7. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 6, characterized in that, The strength calculation unit includes: A mapping subunit is used to determine the gap level based on the mapping relationship between the capacity gap value and a preset gap range; A constraint subunit is used to constrain the gap level according to the call permission coefficient to obtain the available capability level; An adjustment subunit is used to adjust the available capability level according to the global capability distribution parameters to obtain the capability output intensity.

8. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 7, characterized in that, The generation module includes: A distribution unit is used to distribute the global capability distribution parameters based on a comparison result where the capability output strength is less than a preset strength threshold, and to distribute the global strategy parameters based on a comparison result where the capability output strength is greater than or equal to the preset strength threshold. The frequency determination unit is used to obtain the task request frequency of various tasks within a preset statistical time period in real time, and to determine the high-frequency task set based on the processing tasks corresponding to the task request frequency greater than the preset frequency threshold. A gap determination unit is used to determine the capability gap value based on the difference between the capability index corresponding to the high-frequency task set and the global capability distribution parameter; A generation unit is used to generate new child nodes based on the comparison result between the capability gap value and the preset gap threshold, wherein the new child nodes are initialized based on the global policy parameters.

9. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 8, characterized in that, The adjustment module includes: The preprocessing unit is used to sample the response delay and task completion status after capacity distribution within a preset observation period, and to standardize the sampling results to obtain a standardized delay sequence and a standardized completion sequence, respectively. The observation and determination unit is used to determine the time delay fluctuation characteristic value based on the standardized time delay sequence, and to determine the completion stability characteristic value based on the standardized completion sequence. The indicator calculation unit is used to determine the short-cycle performance indicator value based on the time delay fluctuation characteristic value and the completion stability characteristic value, and to determine the comprehensive performance indicator value according to the changing trend of all cycle performance indicator values ​​within the preset adjustment period. An adjustment unit is used to adjust the preset complexity threshold or the call permission coefficient based on the comprehensive performance index value.

10. The master-slave hierarchical OpenClaw terminal access and collaborative management system according to claim 9, characterized in that, The adjustment unit includes: The interval determination subunit is used to divide the comprehensive performance index value into a performance insufficiency interval, a performance stability interval, or a performance excess interval based on the comparison results between the comprehensive performance index value and the preset first index threshold and the preset second index threshold. A threshold adjustment subunit is used to reduce the preset complexity threshold when the comprehensive performance index value is in the performance insufficiency range, and to increase the preset complexity threshold when the comprehensive performance index value is in the performance excess range. A license adjustment subunit is configured to increase the call license coefficient when the overall performance index value is in the underperformance range, and to decrease the call license coefficient when the overall performance index value is in the overperformance range.

Images