Process keep-alive scheduling method and device for software automation RPA workflow
By building a cost consumption optimization problem model and optimizing scheduling solutions using graph neural networks and feedforward neural networks, the problem of large resource overhead in software automation RPA workflows is solved, and the total system cost is minimized and resource efficiency is improved.
Patent Information
- Application Number
- CN202410927676.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-07-11
AI Technical Summary
The task scheduling resource overhead of existing software automation RPA workflows is too high, especially due to inefficiency due to software initialization startup overhead and process maintenance overhead.
By obtaining the execution data of the software automated RPA workflow, a cost-consuming optimization problem model is built, and a graph neural network and a feedforward neural network are used for scheduling and depreciation, optimizing software process maintenance and destruction decisions, and reducing resource overhead.
It minimizes the total system cost and overhead, optimizes the process scheduling process of RPA workflow, reduces software startup overhead and reduces resource waste.
Smart Images

Figure CN118467142B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of automated task scheduling, and in particular to a process keep-alive scheduling method and device for software automated RPA workflows. Background Art
[0002] Robotic Process Automation (RPA) is a modern digital technology that uses robot software to automatically perform regular, tedious and predictable tasks in enterprise processes. The RPA system uses virtual software robots to imitate human operations such as mouse clicks, keyboard input and data processing on computers to perform various tasks such as data entry, document processing, information extraction, etc. These tasks usually involve data processing, interactive communication and other recurring operations in business processes.
[0003] A software automation RPA workflow is a set of RPA jobs that need to use the corresponding software in the executor. There are data and control dependencies between RPA jobs. Different types of software automation RPA jobs need to start and use different software. The initialization and startup overhead of large-scale software applications in enterprises is large, which has a significant impact on the execution efficiency of software automation RPA jobs. If the executor keeps the corresponding software process alive after completing the software automation RPA job, the startup overhead of subsequent executions of the same type of software automation RPA jobs can be reduced. However, keeping the process alive will also cause resource overhead. Summary of the invention
[0004] In view of this, the purpose of the present invention is to provide a process keep-alive scheduling method and device for software automation RPA workflow, which is used to solve the problem of excessive resource overhead in existing task scheduling.
[0005] The present invention provides a process keep-alive scheduling method for software automation RPA workflow, comprising the steps of:
[0006] S1: Obtain the execution data of the software automation RPA workflow and build a cost consumption optimization problem model based on the execution data;
[0007] S2: Construct a directed acyclic graph by executing data, transfer and aggregate features of the directed acyclic graph through a graph neural network, and obtain graph embedding feature representation;
[0008] S3: Use a feedforward neural network to predict the scheduling solution for the graph embedding feature representation and obtain a set of scheduling solutions;
[0009] S4: Substitute the scheduling solution set into the cost consumption optimization problem model to calculate the optimal scheduling solution.
[0010] Preferred:
[0011] The execution data of software automation RPA workflow includes: executor cluster N={1,...,n}, where n is the total number of executors; the hardware resources equipped by executor n are represented by U n ; The job type set K of software automation RPA = {job1,...,job k}, k is the total number of job types; job type job k The hardware resources required are u k ; Executor n runs and completes job type job k The running cost caused is expressed as ;Job type job k The software initialization startup cost on executor n is expressed as ; Executor n pairs of job types per unit time k The software process keepalive overhead used is expressed as .
[0012] Preferably, step S1 specifically comprises:
[0013] S11: Obtain a workflow scheduling task F in the software automation RPA workflow. The system time in the workflow scheduling task is divided into a discrete time step set T={1,...,t}, where t represents the current time step;
[0014] S12: Obtain the total execution cost of the entire system in time step t by executing data calculation and the total process keepalive overhead ;
[0015] S13: The total cost C of the entire system in time step t t It is expressed as:
[0016] ,
[0017] The parameter α is used to adjust and The weight between them, parameter α satisfies ;
[0018] S14: Construct a cost consumption optimization problem model based on execution data and total overhead.
[0019] Preferred:
[0020] Total execution cost The calculation formula is:
[0021] ,
[0022] in, Represents the job type job scheduled to executor n in time step t k The number of RPA jobs, Indicates that the job type job is on executor n before time step t k The number of software processes that are kept alive for the RPA jobs corresponding to them;
[0023] Total process keepalive overhead The calculation formula is:
[0024] ,
[0025] in
[0026] ,
[0027] Indicates that at the end of time step t, executor n has a job type job k The number of software processes destroyed corresponding to the RPA jobs;
[0028] The expression of the cost consumption optimization problem model is:
[0029] .
[0030] Preferably, step S2 specifically comprises:
[0031] S21: The running cost of various RPA jobs on each executor , software initialization startup overhead and software process keepalive overhead As the node information set V of the directed acyclic graph, the data transmission and control relationship between RPA jobs in the workflow is taken as the edge information set E of the directed acyclic graph to construct the directed acyclic graph G=<V,E> ;
[0032] S22: The features of the nodes of the directed acyclic graph are transferred and aggregated through the graph neural network, and the node graph embedding feature representation is calculated;
[0033] S23: The features of the edges of the directed acyclic graph are transferred and aggregated through the graph neural network, and the graph embedding feature representation of the edges is calculated;
[0034] S24: The graph embedding feature representation is composed of the node graph embedding feature representation and the edge graph embedding feature representation.
[0035] Preferred:
[0036] Node graph embedding feature representation for:
[0037] ,
[0038] in, represents the neighborhood of node v, u represents the neighboring nodes of node v, W represents the parameters of the graph neural network, and A Represents an aggregate function, C represents the connection operation, l represents the number of layers of the network, t represents the current time step, represents the node graph embedding feature representation of node v in the lth layer of the graph neural network at the current time step t, represents the node graph embedding feature representation of the neighboring node u in the lth layer of the graph neural network at the current time step t, Indicates the node graph embedding feature representation of node v in the l+1th layer of the graph neural network at the current time step t;
[0039] The expression of the graph embedding feature representation of the edge is:
[0040] ,
[0041] Where e is the number of the edge, represents the graph embedding feature representation of edge e at the current time step t, represents the starting node feature of edge e at the current time step t, Represents the endpoint node features of edge e at the current time step t.
[0042] Preferably, step S3 specifically comprises:
[0043] S31: Obtaining a node graph embedding feature representation and an edge graph embedding feature representation through a graph embedding feature representation;
[0044] S32: Use a feedforward neural network to predict the scheduling solution for the node graph embedding feature representation, and calculate the number of software process destruction y on each executor in the executor cluster N. t ;
[0045] S32: The scheduling solution is predicted by the graph embedding feature representation of the edge through the feedforward neural network, and the priority scheduling probability distribution x of each executor in the executor cluster N is calculated. t ;
[0046] S33: Probability distribution x through priority scheduling t Calculate the scheduling set x of each executor S ;
[0047] S34: Destroy the number of software processes y t and the scheduling set x S As a scheduling solution;
[0048] S35: Repeat steps S32-S34 to obtain a total of P scheduling solutions, which constitute a scheduling solution set.
[0049] Preferred:
[0050] Number of software processes destroyed y t The expression is:
[0051] ,
[0052] ,
[0053] ,
[0054] in, and represents the activation function, l represents the number of layers of the network, FFNN l represents the mapping of the lth layer of the feedforward neural network, represents the floor function, represents the node graph embedding feature representation of node v at the current time step t, Represents the job type job scheduled to executor n in time step t k The number of RPA jobs, Indicates that the job type job is on executor n before time step t k The number of software processes that are kept alive for the RPA jobs corresponding to them;
[0055] Executor priority scheduling probability vector x t The expression is:
[0056] ,
[0057] Among them, softmax represents the activation function, represents the graph embedding feature representation of edge e at the current time step t, and E represents the edge information set;
[0058] According to the Euclidean distance, x t Quantized into a binary schedule set x S , x S The i-th job scheduling action of the s-th executor at the current time step t in The expression is:
[0059] ,
[0060] Among them, 1 means executing scheduling, and 0 means not executing scheduling.
[0061] Preferably, step S4 is specifically:
[0062] Destroy the number of software processes y for each scheduling solution in the scheduling solution set t and the scheduling set x S Substitute the cost consumption optimization problem model, calculate the total cost of each scheduling solution, and take the scheduling solution with the smallest total cost as the optimal scheduling solution.
[0063] A process keep-alive scheduling device for software automation RPA workflow, comprising: a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the process keep-alive scheduling method for software automation RPA workflow.
[0064] The present invention has the following beneficial effects:
[0065] A cost optimization problem model is constructed through the total execution overhead and total process keep-alive overhead of the entire system. This model not only considers the dispatching of software automation RPA jobs to appropriate executors to reduce software startup overhead, but also considers making reasonable keep-alive decisions for software processes on the executor cluster to reduce the waste of computing resource overhead. By constructing a graph embedding feature representation to obtain a set of scheduling solutions, the cost optimization problem model is solved. Through the optimal scheduling solution, the executor is reasonably selected to schedule software automation RPA jobs and determine the number of software processes to be kept alive and destroyed, so as to minimize the total cost of the system and optimize the process keep-alive scheduling process of the RPA workflow. BRIEF DESCRIPTION OF THE DRAWINGS
[0066] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;
[0067] Figure 2 This is a structural diagram of the device according to an embodiment of the present invention;
[0068] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0069] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0070] Reference Figure 1 The present application provides a process keep-alive scheduling method for software automation RPA workflow, comprising the steps of:
[0071] S1: Obtain the execution data of the software automation RPA workflow and build a cost consumption optimization problem model based on the execution data;
[0072] Furthermore, a software automated RPA workflow is a set of RPA jobs that require the use of corresponding software in the executor, and there are data and control dependencies between the RPA jobs.
[0073] The execution data of software automation RPA workflow includes: executor cluster N={1,...,n}, where n is the total number of executors; the hardware resources equipped by executor n are represented by U n ; The job type set K of software automation RPA = {job1,...,job k}, k is the total number of job types; job type job k The hardware resources required are u k ; Executor n runs and completes job type job k The running cost caused is expressed as ;Job type job k The software initialization startup cost on executor n is expressed as ; Executor n pairs of job types per unit time k The software process keepalive overhead used is expressed as .
[0074] Specifically, if executor n completes job k Then keep the corresponding software process alive, and then execute the same type of software automation RPA job k This part of the initialization overhead can be saved, but the process keep alive needs to occupy resources, so the process keep alive is set to .
[0075] Furthermore, step S1 is specifically as follows:
[0076] S11: Obtain a workflow scheduling task F in the software automation RPA workflow. The system time in the workflow scheduling task is divided into a discrete time step set T={1,...,t}, where t represents the current time step;
[0077] Specifically, in a software automation RPA workflow scheduling process, the system time is divided into discrete unit time steps T={1,...,t}, and the workflow scheduling tasks during this period are represented as F={job1,job2,...,job m}, job i ∈K, the same type of software automation RPA job in a workflow scheduling task F may need to be scheduled for execution multiple times.
[0078] S12: Obtain the total execution cost of the entire system in time step t by executing data calculation and the total process keepalive overhead ;
[0079] S13: The total cost C of the entire system in time step t t It is expressed as:
[0080] ,
[0081] The parameter α is used to adjust and The weight between them, parameter α satisfies ;
[0082] Specifically, the parameter α is used to adjust the trade-off between the two overheads, and the parameter α should always comply with the constraint Otherwise, process keepalive cannot alleviate the software initialization startup overhead problem
[0083] Furthermore, the total execution cost The calculation formula is:
[0084] ,
[0085] in, Represents the job type job scheduled to executor n in time step t k The number of RPA jobs, Indicates that the job type job is on executor n before time step t k The number of software processes that are kept alive for the RPA jobs corresponding to them;
[0086] Total process keepalive overhead The calculation formula is:
[0087] ,
[0088] in
[0089] ,
[0090] Indicates that at the end of time step t, executor n has a job type job k The number of software processes destroyed corresponding to the RPA jobs;
[0091] S14: Construct a cost consumption optimization problem model based on execution data and total overhead.
[0092] Furthermore, the process keep-alive scheduling method needs to distribute and schedule the software automation RPA jobs to the appropriate executors in each time step and determine the number of software processes of various types to be destroyed on each executor. The goal is to reduce the total overhead of the system to To minimize the problem, model it as an integer linear programming problem representation of convex optimization.
[0093] The expression of the cost consumption optimization problem model is:
[0094]
[0095] Specifically, the first constraint ensures that each executor complies with the hardware resource limit, and the second constraint limits the number of software processes destroyed to no more than the number of software processes kept alive.
[0096] S2: Construct a directed acyclic graph by executing data, transfer and aggregate features of the directed acyclic graph through a graph neural network, and obtain graph embedding feature representation;
[0097] Furthermore, step S2 is specifically as follows:
[0098] S21: The running cost of various RPA jobs on each executor , software initialization startup overhead and software process keepalive overhead As the node information set V of the directed acyclic graph, the data transmission and control relationship between RPA jobs in the workflow is taken as the edge information set E of the directed acyclic graph to construct the directed acyclic graph G=<V,E> ;
[0099] Specifically, the node information set V in the graph G contains the execution data of the software automation RPA job on each executor, and the edge information set E in the graph G represents the data transmission and control relationship between the software automation RPA jobs in the RPA workflow.
[0100] S22: The features of the nodes of the directed acyclic graph are transferred and aggregated through the graph neural network, and the embedded feature representation of the node graph is calculated;
[0101] Specifically, the RPA workflow is transmitted and aggregated based on the graph neural network GCN. First, for the time step t, the set of software automation RPA jobs in the time step is represented as F t , aggregate RPA workflow graph information, job set and cluster information to obtain the input data h required by GCN 0 :
[0102] ,
[0103] GCN learns the embedding of graph nodes in the next state by aggregating features of other nodes in the neighborhood.
[0104] Furthermore, the second-order neighborhood of the node in the graph includes the RPA job node and other RPA job nodes that have a control relationship with the node. The execution environment of all RPA jobs that have a control relationship with each other in the cluster at the current time step t can be understood by aggregating the information in the second-order neighborhood of each node. Therefore, a process keep-alive scheduling strategy can be made based on the second-order domain feature information of the node.
[0105] The iterative update expression of the node graph embedding feature representation is:
[0106] ,
[0107] in, represents the neighborhood of node v, u represents the neighboring nodes of node v, W represents the parameters of the graph neural network, and A Represents an aggregate function, C represents the connection operation, l represents the number of layers of the network, t represents the current time step, represents the node graph embedding feature representation of node v in the lth layer of the graph neural network at the current time step t, represents the node graph embedding feature representation of the neighboring node u in the lth layer of the graph neural network at the current time step t, Indicates the node graph embedding feature representation of node v in the l+1th layer of the graph neural network at the current time step t;
[0108] S23: The features of the edges of the directed acyclic graph are transferred and aggregated through the graph neural network, and the graph embedding feature representation of the edges is calculated;
[0109] Furthermore, based on the graph embedding feature representation of the second-order neighborhood of the node, the starting node features and the end node features of each edge are combined to obtain the graph embedding feature representation of each edge.
[0110] The expression of the graph embedding feature representation of the edge is:
[0111] ,
[0112] Where e is the number of the edge, represents the graph embedding feature representation of edge e at the current time step t, represents the starting node feature of edge e at the current time step t, Represents the endpoint node features of edge e at the current time step t.
[0113] S24: The graph embedding feature representation is composed of the node graph embedding feature representation and the edge graph embedding feature representation.
[0114] S3: Use a feedforward neural network to predict the scheduling solution for the graph embedding feature representation and obtain a set of scheduling solutions;
[0115] Furthermore, step S3 is specifically as follows:
[0116] S31: Obtaining a node graph embedding feature representation and an edge graph embedding feature representation through a graph embedding feature representation;
[0117] S32: Use a feedforward neural network to predict the scheduling solution for the node graph embedding feature representation, and calculate the number of software process destruction y on each executor in the executor cluster N. t ;
[0118] Furthermore, the number of software process destruction y t The expression is:
[0119] ,
[0120] ,
[0121] ,
[0122] in, and represents the activation function, l represents the number of layers of the network, FFNN l represents the mapping of the lth layer of the feedforward neural network, represents the floor function, represents the node graph embedding feature representation of node v at the current time step t, Represents the job type job scheduled to executor n in time step t k The number of RPA jobs, Indicates that the job type job is on executor n before time step t k The number of software processes that are kept alive for the RPA jobs corresponding to them;
[0123] S32: The scheduling solution is predicted by the graph embedding feature representation of the edge through the feedforward neural network, and the priority scheduling probability distribution x of each executor in the executor cluster N is calculated. t ;
[0124] Furthermore, the executor priority scheduling probability vector x t The expression is:
[0125] ,
[0126] Among them, softmax represents the activation function, represents the graph embedding feature representation of edge e at the current time step t, and E represents the edge information set;
[0127] S33: Probability distribution x through priority scheduling t Calculate the scheduling set x of each executor S ;
[0128] Furthermore, according to the Euclidean distance, x t Quantized into a binary schedule set x S , x S The i-th job scheduling action of the s-th executor at the current time step t in The expression is:
[0129] ,
[0130] Among them, 1 means executing scheduling, and 0 means not executing scheduling.
[0131] S34: Destroy the number of software processes y t and the scheduling set x S As a scheduling solution;
[0132] S35: Repeat steps S32-S34 to obtain a total of P scheduling solutions, which constitute a scheduling solution set.
[0133] S4: Substitute the scheduling solution set into the cost consumption optimization problem model to calculate the optimal scheduling solution.
[0134] Furthermore, step S4 is specifically as follows:
[0135] Destroy the number of software processes y for each scheduling solution in the scheduling solution set t and the scheduling set x S Substitute the cost consumption optimization problem model, calculate the total cost of each scheduling solution, and take the scheduling solution with the smallest total cost as the optimal scheduling solution.
[0136] Specifically, according to the optimal scheduling solution, at the current time step t, the RPA job in the workflow is scheduled to the corresponding executor for execution, and after the execution is completed, the corresponding number of RPA job software processes are destroyed on each executor.
[0137] See also Figure 2 , Figure 2 It is a working diagram of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a process keep-alive scheduling device 401 for software automated RPA workflow, a processor 402 and a storage medium 403.
[0138] A process keep-alive scheduling device 401 for software automation RPA workflow: The process keep-alive scheduling device 401 for software automation RPA workflow implements the process keep-alive scheduling method for software automation RPA workflow.
[0139] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the process keep-alive scheduling method for software automation RPA workflow.
[0140] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the process keep-alive scheduling method for software automation RPA workflow.
[0141] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0142] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.
[0143] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A process keep-alive scheduling method for software automation RPA workflow, It is characterized in that Includes steps: S1: Obtain the execution data of the software automation RPA workflow and build a cost consumption optimization problem model based on the execution data; S2: Construct a directed acyclic graph by executing data, transfer and aggregate features of the directed acyclic graph through a graph neural network, and obtain graph embedding feature representation; S3: Use a feedforward neural network to predict the scheduling solution for the graph embedding feature representation and obtain a set of scheduling solutions; S4: Substitute the scheduling solution set into the cost consumption optimization problem model to calculate and obtain the optimal scheduling solution; The execution data of software automation RPA workflow includes: executor cluster N={1,...,n}, where n is the total number of executors; the hardware resources equipped by executor n are represented by U n ; The job type set K of software automation RPA = {job 1 ,...,job k }, k is the total number of job types; job type job k The hardware resources required are u k ; Executor n runs and completes job type job k The running cost caused is expressed as ;Job type job k The software initialization startup cost on executor n is expressed as ; Executor n pairs of job types per unit time k The software process keepalive overhead used is expressed as ; Step S1 is specifically as follows: S11: Obtain a workflow scheduling task F in the software automation RPA workflow. The system time in the workflow scheduling task is divided into a discrete time step set T={1,...,t}, where t represents the current time step; S12: Obtain the total execution cost of the entire system in time step t by executing data calculation and the total process keepalive overhead ; S13: The total cost C of the entire system in time step t t It is expressed as: , The parameter α is used to adjust and The weight between them, parameter α satisfies ; S14: Construct a cost consumption optimization problem model through execution data and total overhead; Total execution cost The calculation formula is: , in, Represents the job type job scheduled to executor n in time step t k The number of RPA jobs, Indicates that the job type job is on executor n before time step t k The number of software processes that are kept alive for the RPA jobs corresponding to them; Total process keepalive overhead The calculation formula is: , in , Indicates that at the end of time step t, executor n has a job type job k The number of software processes destroyed corresponding to the RPA jobs; The expression of the cost consumption optimization problem model is: ; Step S4 is specifically as follows: Destroy the number of software processes y for each scheduling solution in the scheduling solution set t and the scheduling set x S Substitute the cost consumption optimization problem model, calculate the total cost of each scheduling solution, and take the scheduling solution with the smallest total cost as the optimal scheduling solution.
2. According to claim 1, the process keep-alive scheduling method for software automation RPA workflow, It is characterized in that Step S2 is specifically as follows: S21: The running cost of various RPA jobs on each executor , software initialization startup overhead and software process keepalive overhead As the node information set V of the directed acyclic graph, the data transmission and control relationship between RPA jobs in the workflow is taken as the edge information set E of the directed acyclic graph to construct the directed acyclic graph G=<V,E> ; S22: The features of the nodes of the directed acyclic graph are transferred and aggregated through the graph neural network, and the embedded feature representation of the node graph is calculated; S23: The features of the edges of the directed acyclic graph are transferred and aggregated through the graph neural network, and the graph embedding feature representation of the edges is calculated; S24: The graph embedding feature representation is composed of the node graph embedding feature representation and the edge graph embedding feature representation.
3. According to claim 2, the process keep-alive scheduling method for software automation RPA workflow, Features: The iterative update expression of the node graph embedding feature representation is: , in, represents the neighborhood of node v, u represents the neighboring nodes of node v, W represents the parameters of the graph neural network, and A Represents an aggregate function, C represents the connection operation, l represents the number of layers of the network, t represents the current time step, represents the node graph embedding feature representation of node v in the lth layer of the graph neural network at the current time step t, represents the node graph embedding feature representation of the neighboring node u in the lth layer of the graph neural network at the current time step t, Indicates the node graph embedding feature representation of node v in the l+1th layer of the graph neural network at the current time step t; The expression of the graph embedding feature representation of the edge is: , Where e is the number of the edge, represents the graph embedding feature representation of edge e at the current time step t, represents the starting node feature of edge e at the current time step t, Represents the endpoint node features of edge e at the current time step t.
4. According to claim 1, the process keep-alive scheduling method for software automation RPA workflow, It is characterized in that Step S3 is specifically as follows: S31: Obtaining a node graph embedding feature representation and an edge graph embedding feature representation through a graph embedding feature representation; S32: Use a feedforward neural network to predict the scheduling solution for the node graph embedding feature representation, and calculate the number of software process destruction y on each executor in the executor cluster N. t ; S32: The scheduling solution is predicted by the graph embedding feature representation of the edge through the feedforward neural network, and the priority scheduling probability distribution x of each executor in the executor cluster N is calculated. t ; S33: Probability distribution x through priority scheduling t Calculate the scheduling set x of each executor S ; S34: Destroy the number of software processes y t and the scheduling set x S As a scheduling solution; S35: Repeat steps S32-S34 to obtain a total of P scheduling solutions, which constitute a scheduling solution set.
5. According to claim 4, the process keep-alive scheduling method for software automation RPA workflow, Features: Number of software processes destroyed y t The expression is: , , , in, and represents the activation function, l represents the number of layers of the network, FFNN l represents the mapping of the lth layer of the feedforward neural network, represents the floor function, represents the node graph embedding feature representation of node v at the current time step t, Represents the job type job scheduled to executor n in time step t k The number of RPA jobs, Indicates that the job type job is on executor n before time step t k The number of software processes that are kept alive for the RPA jobs corresponding to them; Executor priority scheduling probability vector x t The expression is: , Among them, softmax represents the activation function, represents the graph embedding feature representation of edge e at the current time step t, and E represents the edge information set; According to the Euclidean distance, x t Quantized into a binary schedule set x S , x S The i-th job scheduling action of the s-th executor at the current time step t in The expression is: , Among them, 1 means executing scheduling, and 0 means not executing scheduling.
6. A process keep-alive scheduling device for software automation RPA workflow, Features: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a process keep-alive scheduling method for software automated RPA workflow as described in any one of claims 1 to 5.
