CN-121979681-A - Multi-agent digital human resource optimization method and system based on dynamic computing power scheduling

CN121979681ACN 121979681 ACN121979681 ACN 121979681ACN-121979681-A

Abstract

The invention discloses a multi-agent digital human resource optimization method and a system based on computational power dynamic scheduling, which relate to the technical field of data processing of computational power resource scheduling and comprise the steps of collecting real-time running state parameters of a multi-agent digital human system, and vectorizing and encoding the real-time running state parameters according to a preset characteristic dimension sequence to generate a load characteristic vector; the method comprises the steps of calculating an calculation force demand index of each digital human intelligent body through a load evaluation model, descending order of each digital human intelligent body according to the calculation force demand index to generate a calculation force distribution priority queue, obtaining available total calculation force capacity, calculating calculation force resource distribution share of each digital human intelligent body according to the calculation force distribution priority queue and the calculation force demand index, and executing dynamic distribution operation of calculation force resources on each digital human intelligent body of the multi-intelligent-body digital human system according to the calculation force resource distribution share to complete collaborative optimization scheduling of the multi-intelligent-body digital human resources. The invention realizes the efficient flow and integration of the computational resources.

Inventors

LI DONGMING
CHENG JIE
WU JIAQIAN

Assignees

南京南数数字产业集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. A multi-agent digital human resource optimization method based on dynamic computing power dispatching is characterized by comprising the following steps of, Collecting real-time running state parameters of each digital human intelligent agent in the multi-intelligent-agent digital human system, and vectorizing and encoding the real-time running state parameters according to a preset characteristic dimension sequence to generate a load characteristic vector; Calculating the calculation power demand index of each digital human intelligent agent through a load evaluation model based on the load characteristic vector; according to the calculation force demand index, descending order arrangement is carried out on each digital human intelligent agent, and a calculation force distribution priority queue is generated; acquiring the available total computing power capacity of the current computing power resource pool, and calculating computing power resource allocation shares of all digital human intelligent agents according to the computing power allocation priority queue and the computing power demand index; and executing dynamic allocation operation of the computing power resources on each digital person of the multi-agent digital person system according to the computing power resource allocation share, and completing collaborative optimization scheduling of the multi-agent digital person resources.
2. The method for optimizing digital human resources of multiple intelligent agents based on dynamic scheduling of computing power according to claim 1, wherein the step of performing dynamic allocation operation of computing power resources to each digital human intelligent agent of the digital human system of multiple intelligent agents according to the share of computing power allocation comprises the steps of: acquiring the allocated computing power resource quantity of each digital human intelligent agent before the current scheduling period starts, and performing difference calculation on the computing power resource allocation share of each digital human intelligent agent and the allocated computing power resource quantity to obtain the computing power resource adjustment quantity of each digital human intelligent agent; Dividing each digital human intelligent agent into an computing power resource release group and a computing power resource acquisition group based on the positive and negative attributes of the computing power resource adjustment quantity; Preferentially sending an computational power resource recovery instruction to the computational power resource release group, waiting for all digital human intelligent agents in the computational power resource release group to return a computational power resource release completion confirmation signal, and updating the available total computational power capacity of a computational power resource pool; And sequentially sending an computing power resource allocation instruction to each digital human intelligent agent in the computing power resource acquisition group according to the arrangement sequence in the computing power allocation priority queue, so as to finish the dynamic allocation operation of the computing power resources.
3. The multi-agent digital human resource optimization method based on dynamic computing power scheduling according to claim 2, wherein the computing method of the computing power resource allocation share is as follows, Sending a resource state query request to an algorithm power resource management module of the multi-agent digital man system to obtain the available total algorithm power capacity; Traversing all queue elements of the computing power distribution priority queue, and accumulating computing power demand indexes stored in each queue element to obtain a computing power demand index sum; calculating a basic distribution proportion value of each digital human intelligent agent according to the ratio of the calculated force demand index to the sum of the calculated force demand indexes; based on the calculation force, distributing priority sequence numbers of all queue elements in the priority queue, and processing through a priority weighting adjustment mechanism to obtain a weighted distribution proportion value; carrying out normalization correction on the weighted distribution proportion value to obtain a normalized distribution proportion value; multiplying the normalized distribution proportion value by the available total calculation force capacity to obtain a preliminary calculation force distribution amount, and performing up-regulation processing on the digital human intelligent agents with the preliminary calculation force distribution amount smaller than a minimum calculation force guarantee threshold value to obtain calculation force resource distribution shares of all the digital human intelligent agents.
4. The multi-agent digital human resource optimization method based on dynamic computing power scheduling according to claim 3, wherein the method for generating the computing power distribution priority queue, Reading a calculation force demand index record table from a system cache area, and extracting an agent unique identifier and a calculation force demand index of each digital human agent to form an agent demand data pair; Taking the calculated force demand index as a sequencing basis, and executing descending sequencing operation on the intelligent agent demand data pair; Aiming at a plurality of agent demand data pairs with equal calculated force demand index values, carrying out secondary sorting according to the delay deviation degree values in the load characteristic vectors of the corresponding digital human agents; assigning priority sequence numbers to the ordered demand data pairs of each intelligent agent, and dividing each digital human intelligent agent into different demand grade categories according to the comparison result of the calculated force demand index and a preset threshold value; and constructing a force distribution priority queue according to the sequencing result, wherein the force distribution priority queue adopts a first-in first-out queue access mode.
5. The multi-agent digital human resource optimization method based on dynamic power scheduling of claim 4, wherein the power demand index is obtained by the following steps of, The method comprises the steps of obtaining load characteristic vectors of all digital human intelligent agents to form a load characteristic vector set to be evaluated; inputting the load characteristic vector to a load evaluation model, and outputting an initial calculation power demand evaluation value; Performing numerical range constraint processing on the initial calculation force demand evaluation value to obtain a corrected calculation force demand evaluation value; Acquiring an average calculation power consumption record value of each digital human intelligent agent in a history scheduling period, and carrying out weighted average on the average calculation power consumption record value and the corrected calculation power demand evaluation value according to a preset history weight factor and a current weight factor to obtain an initial calculation power demand value; And carrying out normalization mapping processing on the initial calculation force demand value to obtain the calculation force demand index of each digital human intelligent body.
6. The multi-agent digital human resource optimization method based on dynamic computing power scheduling of claim 5, wherein the load assessment model is a neural network model based on a multi-layer perceptron structure, and comprises an input layer, a hidden layer and an output layer, wherein the number of nodes of the input layer is matched with the number of feature dimensions of a load feature vector.
7. The multi-agent digital human resource optimization method based on dynamic computing power scheduling of claim 6, wherein the load feature vector is generated by the method, Establishing data communication connection with each digital personal intelligent agent, and sending a state acquisition instruction carrying a uniform timestamp mark to each digital personal intelligent agent through the data communication connection; Receiving an original running state data packet returned by each digital human intelligent agent according to the state acquisition instruction, and extracting real-time running state parameters, wherein the real-time running state parameters comprise a current task type identifier, a task queue depth value, an occupied computing power resource amount and an interactive response delay time length; Establishing a mapping relation table of task types and numerical codes according to the real-time running state parameters, and converting the current task type identification into a task type code value; performing range normalization processing on the task queue depth value to obtain a normalized task queue depth value; Converting the occupied computing power resource amount into a computing power resource occupation proportion value according to the distributed computing power resource upper limit value of each digital human intelligent agent; Comparing the interaction response delay time length with a preset response delay reference value, and calculating a delay deviation value; and sequentially arranging and combining the task type coding value, the normalized task queue depth value, the computing power resource occupation proportion value and the delay deviation degree value according to a preset characteristic dimension sequence to generate a load characteristic vector.
8. A multi-agent digital human resource optimization system based on dynamic computing power scheduling is characterized by comprising the multi-agent digital human resource optimization method based on dynamic computing power scheduling according to any one of claims 1-7, The state acquisition and feature coding module is used for acquiring real-time running state parameters of each digital human intelligent body in the multi-intelligent-body digital human system, and carrying out vectorization coding on the real-time running state parameters according to a preset feature dimension sequence to generate a load feature vector; the power demand evaluation module is used for calculating power demand indexes of all digital human intelligent agents through a load evaluation model based on the load characteristic vector; The dynamic priority queue generating module is used for descending order arrangement of the digital human intelligent agents according to the calculation force demand index to generate a calculation force distribution priority queue; The computing power resource allocation calculation module is used for acquiring the available total computing power capacity of the current computing power resource pool and calculating computing power resource allocation share of each digital human intelligent agent according to the computing power allocation priority queue and the computing power demand index; And the dynamic scheduling and allocation execution module is used for executing dynamic allocation operation of the computing power resources on each digital person of the multi-agent digital person system according to the computing power resource allocation share to complete collaborative optimization scheduling of the multi-agent digital person resources.
9. A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the computer program is characterized in that the processor realizes the steps of the multi-agent digital human resource optimization method based on dynamic computing power scheduling according to any one of claims 1-7 when executing the computer program.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the multi-agent digital human resource optimization method based on dynamic computing power scheduling as set forth in any one of claims 1 to 7.

Description

Multi-agent digital human resource optimization method and system based on dynamic computing power scheduling Technical Field The invention relates to the technical field of data processing of computational power resource scheduling, in particular to a multi-agent digital human resource optimization method and system based on computational power dynamic scheduling. Background Along with the rapid development of digital man-made technology, intelligent collaborative computing and cloud edge end fusion computing systems, the multi-intelligent digital man-made system is gradually applied to complex business scenes such as virtual customer service, digital government affairs, intelligent exhibition, meta-universe interaction, industrial digital twin and the like. Such systems are typically operated in concert by multiple digital human agents with sensing, decision making and interaction capabilities, and the different agents have significant heterogeneous and dynamic fluctuation characteristics in terms of demand for computational resources in speech synthesis, action driving, visual rendering, semantic understanding, and the like. The existing computing power scheduling technology takes tasks or clusters as core objects, focuses on overall load balancing and throughput rate improvement of computing power resources, and is difficult to finely describe load changes of single digital human intelligent agents in a real-time operation process and influences of the load changes on overall collaborative efficiency of a system. Under the multi-agent concurrent operation scene, if a dynamic calculation power evaluation and priority scheduling mechanism for digital person individuals is lacked, the problems of calculation power distribution lag, insufficient calculation power of key agents, non-key agent resource redundancy and the like are easily caused, so that the real-time performance, consistency and service stability of digital person interaction are affected, and the large-scale deployment and high-quality operation of the multi-agent digital person system are restricted. CN119248490B discloses an intelligent power dispatching method and system based on dynamic programming, which calculates the resource difference value between clusters by monitoring the available resource information of a plurality of power clusters in real time, dynamically optimizes the power grouping by combining the historical task type and the resource consumption characteristics, and then scores and recommends the power service according to the service level agreement. According to the scheme, balanced scheduling and utilization rate improvement of the computational power resources are realized at the cluster level, but scheduling objects of the scheme are mainly oriented to computational power clusters and task types, fine granularity modeling of the running states of single digital persons in a multi-agent system is lacked, and the instant computational power demand difference cannot be described from the loading characteristics of the agents, so that the scheme is difficult to adapt to the application scene of rapid change of the individual computational power demands when the multi-digital persons run concurrently. CN119311428B proposes an intelligent computing power resource scheduling and optimizing system, which obtains computing power resource operation data through a monitoring acquisition module, and predicts task resource requirements in combination with historical data, so as to dynamically adjust computing power resources in advance, thereby improving the operation efficiency of resource nodes. The scheme has certain advantages in the aspects of task demand prediction and resource pre-scheduling, but the core of the scheme still takes tasks as centers to perform calculation power distribution, the scheduling strategy focuses on the optimization of a resource node level, a corresponding calculation power demand evaluation model is not established aiming at the running states, interaction strength and real-time load differences of different digital human intelligent agents in the multi-intelligent digital human system, a dynamic calculation power distribution mechanism based on intelligent agent priority is not formed, and calculation power continuity and interaction experience of key digital human intelligent agents are difficult to guarantee. In summary, the existing calculation power scheduling technology still commonly has the problems of coarse scheduling granularity, lack of calculation power demand quantification of an agent level, insufficient collaborative optimization capability and the like in multi-agent digital person application. Aiming at the problem that the dynamic change of the computational power demand is difficult to match accurately in the running process of the multi-agent digital human system, the invention provides the multi-agent digital human resource optimization method and system based on the computa