Human resource management and intelligent decision-making method based on reinforcement learning

Abstract

The invention discloses a human resource management and intelligent decision-making method based on reinforcement learning. The method comprises the following steps that data are collected to construct an organization hierarchy graph, and node states and constraint vectors are generated; forming a reinforcement learning state by the hierarchical graph and the node state; the method comprises the following steps: constructing a management network, an execution network, a mapping unit and an enabling comparison unit based on FeUdal Networks; a mapping unit performs constraint manifold mapping according to the unconstrained target vector and the constraint vector output by the management network to obtain a constraint consistent target vector; the execution network generates an action according to a node state and a constraint consistent target vector, and an enabling comparison unit calculates an enabling difference reward according to the action and a hypothetical action; according to the business performance, enabling difference rewards, constraining violation construction rewards and updating parameters; network output constraint consistent target vectors and actions are called in the human resource management system, and a scheme is formed. According to the invention, the compliance and coordination of human resource configuration are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and human resource management, and particularly relates to a human resource management and intelligent decision-making method based on reinforcement learning. BACKGROUND

[0002] Human resource management information systems have been widely used in job matching, scheduling management and performance evaluation. The mainstream solutions rely on rule engines, experience parameters and statistical analysis models to complete labor planning and resource allocation. The common practice is to establish a post establishment table and a scheduling template based on historical human data, and then combine a simple scoring model or a linear weight model to perform job matching and scheduling optimization. Some systems introduce machine learning classification models or recommendation algorithms to extract features from resume characteristics and performance indicators to score and sort candidate employees, but the overall system still mainly relies on static rules and one-time decisions, lacks continuous decision-making and long-term feedback mechanisms, and has limited response capabilities to dynamic labor demand in complex organizational structures.

[0003] With the expansion of organizational scale and the richness of business scenarios, human resource decision-making needs to consider multiple factors such as organizational hierarchical structure, job responsibility division, employee ability level, work hour and budget constraints, and labor compliance rules. Existing technologies often only focus on a single aspect in state modeling, such as only matching the relationship between job seekers and positions, or only scheduling the relationship between shifts and employees. Few unified graph structure descriptions are established for organizations, positions, employees and management relationships, and there is also a lack of unified state representation methods that can be processed by learning algorithms to encode business indicator features and constraint information together. This fragmented modeling approach makes it difficult for the model to depict the human resource flow process across departments and positions, which is not conducive to making global coordinated decisions under complex constraints.

[0004] Reinforcement learning methods have been applied in recommendation systems, scheduling optimization and resource allocation, but existing research in the human resource management field often stays at the single-layer strategy or simple multi-step decision-making level. Typically, a single agent constructs a reward function based on immediate performance to learn a strategy, lacks differentiation of different time scale targets, and lacks specialized modeling of compliance constraints and employee empowerment indicators. In complex labor scenarios, relying solely on a single reward function often makes it difficult to balance business performance, organizational compliance and employee development, and easily leads to a situation where the strategy swings between short-term benefits and long-term labor risks. There is a lack of structured constraint mechanisms for work hour limits, budget limits and establishment limits, and a lack of internal incentive mechanisms for measuring employee empowerment differences based on counterfactual comparisons.

[0005] Therefore, how to provide a human resource management and intelligent decision-making method based on reinforcement learning is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to provide a human resource management and intelligent decision-making method based on reinforcement learning, which adopts a hierarchical reinforcement learning structure of FeUdal Networks, models the organization, post and employee as a unified organization hierarchy diagram, generates post allocation, scheduling and training scheme through the coordination of high-level target vector, constraint manifold mapping, empowerment difference reward and multi-scale decision mechanism, realizes dynamic resource allocation and intelligent labor decision-making for complex organizational structure, and has the advantages of strong adaptability to organizational changes, more accurate constraint processing and more reasonable employee empowerment measurement.

[0007] According to the human resource management and intelligent decision-making method based on reinforcement learning, the method comprises the following steps:

[0008] Collecting organization data, post data and employee data, constructing an organization hierarchy diagram, and generating a node state vector and a constraint vector for each node;

[0009] The organization hierarchy diagram and the node state vector form a reinforcement learning state;

[0010] Based on FeUdal Networks, a reinforcement learning network is constructed, which comprises a management network, an execution network, a mapping unit and an empowerment comparison unit, the management network generates an unconstrained high-level target vector according to the reinforcement learning state in a long time scale, and the unconstrained high-level target vector encodes empowerment information and compliance information at the same time;

[0011] The mapping unit performs constraint manifold mapping according to the unconstrained high-level target vector and the constraint vector to obtain a constraint-consistent high-level target vector;

[0012] The execution network determines a human resource management action according to the node state vector and the constraint-consistent high-level target vector in a short time scale, and the empowerment comparison unit calculates an empowerment difference reward according to the human resource management action and the hypothetical action;

[0013] A reward structure is constructed according to the business performance, the empowerment difference reward and the constraint violation, and the parameters of the management network, the execution network, the mapping unit and the empowerment comparison unit are adjusted according to the reward structure;

[0014] The trained reinforcement learning network is applied in a human resource management system, a constraint-consistent high-level target vector and a human resource management action are generated according to the organization hierarchy diagram and the node state vector, and a human resource management scheme is formed.

[0015] Optionally, the construction of the organization hierarchy diagram specifically comprises:

[0016] Read organizational data, job data, and employee data from the human resources management system and business system. Register each piece of organizational data as an organizational node, each piece of job data as a job node, and each piece of employee data as an employee node. Organizational data must include at least an organization identifier field and a parent organization identifier field. Job data must include at least a job identifier field and an organization identifier field. Employee data must include at least an employee identifier field and a job identifier field.

[0017] Based on the superior organization identifier field, establish management relationship edges between organization nodes, establish job relationship edges between organization nodes and job nodes based on the organization identifier field, establish employee relationship edges between job nodes and employee nodes based on the job identifier field, and combine organization nodes, job nodes, employee nodes, management relationship edges, job relationship edges and employee relationship edges to form an organizational hierarchy diagram;

[0018] For each organization node, job node, and employee node, select fields related to business scale, job content, personnel capabilities, and historical working hours from the corresponding data records, arrange them in a preset order, perform numerical processing, and generate a node status vector;

[0019] For each organization node, job node, and employee node, select fields related to working hours constraints, budget constraints, staffing constraints, and employment rules from the corresponding data records, and arrange them in a preset order to generate a constraint vector;

[0020] In the organizational hierarchy diagram, each organizational node, job node, and employee node is bound with a node state vector and a constraint vector, resulting in an organizational hierarchy diagram containing node state vectors and constraint vectors.

[0021] Optionally, the generation of the reinforcement learning state specifically includes:

[0022] In an organizational hierarchy diagram containing node state vectors and constraint vectors, a node number is assigned to each organizational node, job node, and employee node. The node number is then associated with the node type, node state vector, and constraint vector. The node number is registered as an integer index in a preset order, and the node type is registered as a category marker that distinguishes organizational nodes, job nodes, and employee nodes. The node state vector and constraint vector are stored in the same index structure according to the node number.

[0023] Based on the management relationship edges, job relationship edges, and employee relationship edges in the organizational hierarchy diagram, the node numbers at both ends of each relationship edge are recorded according to the node number to generate a graph structure record. The graph structure record includes a set of node numbers and a set of edge connection pairs. Each record in the set of edge connection pairs contains the starting node number and the ending node number, which are used to represent the connection method of management relationship, job relationship, and employee relationship in the organizational hierarchy diagram in the number space.

[0024] The node state vector of each node is arranged in a uniform field order. The non-numerical fields in the node state vector are encoded, and the numerical fields in the node state vector are normalized to form a standardized node state vector. The encoding process converts the text field, enumeration field and label field into a fixed-length numerical representation, and the normalization process scales the business volume field, working hour field and performance field to a preset numerical range.

[0025] In terms of feature dimension, the standardized node state vector and constraint vector are concatenated to form a node feature vector. All node feature vectors are arranged in the order of node number to form a feature matrix. The feature matrix is ​​combined with the graph structure record to form a reinforcement learning state. The reinforcement learning state carries both node feature information and connection relationship information, which is used to represent the state corresponding to the organizational hierarchy graph in the human resource management scenario.

[0026] Optionally, the construction of the reinforcement learning network specifically includes:

[0027] Set management time scale and execution time scale on the time axis, divide the time axis into management time step sequence and execution time step sequence. Each management time step covers a preset number of execution time steps. Within each management time step, the reinforcement learning state is regarded as the same batch input on the management network side. Within each execution time step, record the changes of node state vector on the execution network side. Use management time steps to represent the frequency of high-level decision-making and use execution time steps to represent the frequency of human resource management action updates.

[0028] A management network is constructed based on the FeUdal Networks framework, which includes a state encoding layer and a target generation layer. The state encoding layer generates a management hidden representation based on the reinforcement learning state at each management time step. The management hidden representation is a fixed-dimensional vector. The target generation layer generates an unconstrained high-level target vector based on the management hidden representation. The unconstrained high-level target vector is divided into an empowerment information part and a compliance information part in the vector dimension. The empowerment information part is used to represent the target tendency related to employee capability improvement and task authorization, and the compliance information part is used to represent the target tendency related to time constraints, budget constraints, and staffing constraints.

[0029] An execution network is constructed based on the FeUdal Networks framework, which includes a local state processing layer and a target condition processing layer. In each execution time step, the local state processing layer generates an execution hidden representation based on the node state vector. In each execution time step, the target condition processing layer generates an execution feature vector based on the execution hidden representation and the unconstrained high-level target vector. In execution time steps covering the same management time step, the unconstrained high-level target vector is kept unchanged. The unchanging unconstrained high-level target vector guides the execution feature vector to adjust human resource management actions around a unified high-level target in multiple execution time steps.

[0030] In a reinforcement learning network, mapping units and empowerment contrast units are set up. The mapping units are connected to the output of the unconstrained high-level objective vector of the management network, and the empowerment contrast units are connected to the output of the execution feature vector of the execution network. The management network, execution network, mapping units, and empowerment contrast units constitute a reinforcement learning network structure based on FeUdal Networks, which provides high-level objective information and execution feature information for constrained manifold mapping and empowerment difference calculation.

[0031] Optionally, the generation of the constraint-consistent high-level objective vector specifically includes:

[0032] Based on the time limit field, budget limit field, staffing limit field, and employment rule field in the constraint vector, they are arranged in a preset order to form a constraint feature vector. Threshold parameters and weight parameters are set for each field in the constraint feature vector. The threshold parameters include a lower threshold representing the minimum allowable value and an upper threshold representing the maximum allowable value. The weight parameters are used to represent the degree of influence of the corresponding constraint on the adjustment of the high-level target during the constraint manifold mapping process. The threshold parameters and weight parameters are preset based on historical human resource management data and management regulations.

[0033] The unconstrained high-level target vector is divided along the vector dimension to obtain the empowerment information part and the compliance information part. According to the preset mapping relationship, a one-to-one correspondence is established between each dimension of the compliance information part and the working hour limit field, budget limit field, staffing limit field and employment rule field in the constraint feature vector. In the one-to-one correspondence, the type of constraint field that controls each compliance information dimension and the corresponding threshold parameters and weight parameters are recorded, so that the numerical changes of the compliance information part are directly controlled by the constraint feature vector.

[0034] In the mapping unit, constraint manifold mapping is performed based on constraint vectors and constraint feature vectors. The values ​​are adjusted along the constraint directions of each dimension of the compliance information part, and the values ​​of the compliance information part are restricted to the feasible interval determined by the working hour limit field, budget limit field, staffing limit field, and employment rule field. The feasible interval is determined by the threshold parameter. The lower limit value and upper limit value of the feasible interval are calculated for each dimension of compliance information. When the value of the compliance information part is less than the lower limit value, the value of the compliance information part is raised to the lower limit value. When the value of the compliance information part is greater than the upper limit value, the value of the compliance information part is reduced to the upper limit value. The smoothness of the value adjustment is controlled according to the corresponding weight parameter.

[0035] In the mapping unit, the value of the empowerment information part remains unchanged. The compliance information part after constraint manifold mapping is recombined with the empowerment information part to generate a constraint-consistent high-level target vector. The constraint-consistent high-level target vector contains both the empowerment information part and the compliance information part after constraint manifold mapping in the vector dimension, and keeps the dimension of the constraint-consistent high-level target vector consistent with the dimension of the unconsistent high-level target vector.

[0036] Optionally, the calculation of the empowerment difference reward specifically includes:

[0037] At each execution time step, for the job nodes and employee nodes in the organizational hierarchy diagram, select the node state vectors related to the job nodes and employee nodes from the set of node state vectors. Concatenate the node state vectors with the high-level target vectors consistent with the constraints adopted in the current management time step in the feature dimension to form the execution input vector. Input the execution input vector into the local state processing layer and the target condition processing layer in the execution network to generate the execution hidden representation. The execution hidden representation is obtained as a fixed-dimensional numerical vector through linear operation and nonlinear activation, and maintains the mapping relationship with the corresponding job nodes and employee nodes in each execution time step.

[0038] The human resource management action score vector is calculated by action mapping based on the execution hidden representation. Human resource management actions are determined on the human resource management action score vector according to the preset action selection rules. Human resource management actions are recorded at the execution time step. Each dimension of the human resource management action score vector corresponds one-to-one with the preset human resource management actions. The human resource management action with the highest score is selected as the execution result by comparing the scores of each dimension. The execution result is then associated with and stored with the identification information of the job node and the employee node.

[0039] In the empowerment comparison unit, the node state vector, the constraint-consistent high-level target vector, and human resource management actions are used as inputs. According to the preset action combination rules, several actions are selected from the preset action set to form a hypothetical action set. The human resource management actions and hypothetical actions are combined one by one to form a comparison sample. The empowerment index value is calculated based on each comparison sample. The empowerment index value is weighted according to the skill matching degree, ability improvement space, and growth path length. The empowerment difference reward is generated based on the difference between the empowerment index value corresponding to the human resource management action and the empowerment index value corresponding to the hypothetical action, and the empowerment difference reward is registered at the execution time step.

[0040] Optionally, the construction and processing of the reward structure specifically includes:

[0041] Based on the business indicators recorded in the human resources management system, data related to output quantity, service quality and cost consumption are summarized for each management time step. The business performance value is obtained according to the preset calculation function. The business performance value is linearly scaled between the preset lower limit and the preset upper limit and mapped to the external reward range to form the external reward of the management time step. The external reward value is then recorded on the management time step.

[0042] Based on the empowerment difference reward output by the empowerment comparison unit at each execution time step, the empowerment difference reward is weighted and summed according to a preset weight coefficient within the execution time step covered by each management time step to obtain the intrinsic reward of the management time step. At the same time, the empowerment difference reward corresponding to each execution time step is retained as the intrinsic reward of the execution time step. The intrinsic reward of the management time step and the intrinsic reward of the execution time step are associated and stored with the corresponding time step respectively.

[0043] Based on the degree of deviation between the compliance information part of the constraint consistency high-level target vector obtained by the mapping unit and the working hour limit field, budget limit field, staffing limit field and employment rule field in the constraint vector, and based on the working hour limit violation, budget limit violation and staffing limit violation triggered by the execution network in human resource management actions, the constraint violation cost value is calculated. At the management time step level, the constraint violation cost value covering the execution time step is aggregated. At the execution time step level, the constraint violation cost value of the corresponding execution time step is directly recorded. The aggregation result and the recording result are converted into management time step penalty and execution time step penalty, respectively.

[0044] Based on preset coefficients, external rewards, internal rewards, and penalties are linearly combined to obtain the total reward value for each management time step and the total reward value for each execution time step. Based on the total reward value, a reinforcement learning strategy is used to update the rules and simultaneously adjust the parameters of the management network, execution network, mapping unit, and empowerment comparison unit, thus completing a unified parameter update on both the management and execution time scales.

[0045] Optionally, the generation of the human resource management plan specifically includes:

[0046] The trained management network, execution network, mapping unit, and empowerment comparison unit are loaded into the human resource management system. The management network, execution network, mapping unit, and empowerment comparison unit are combined into a reinforcement learning network, and a data interface with organizational data, job data, and employee data is established. The parameters of the management network, execution network, mapping unit, and empowerment comparison unit are written into the system's running memory. The mapping relationship between the input end of the reinforcement learning network and the fields of organizational data, job data, and employee data, as well as the mapping relationship between the output end of the reinforcement learning network and the fields of job configuration results, scheduling results, and training results are completed.

[0047] The organizational hierarchy diagram, node state vector, and constraint vector are updated based on the current organizational data, job data, and employee data. The organizational hierarchy diagram, node state vector, and constraint vector are combined to form the reinforcement learning state in the operation phase. The reinforcement learning state in the operation phase is refreshed once at the beginning of each management time step based on the latest organizational data, job data, and employee data. The organizational hierarchy diagram is updated by parsing the organizational identifier field, job identifier field, and employee identifier field. The node state vector is updated by parsing the business field and personnel field. The constraint vector is updated by parsing the restriction field and rule field.

[0048] Within each management time step, the reinforcement learning state of the running phase is input into the management network to generate an unconstrained high-level objective vector. The constrained manifold mapping is performed through the mapping unit to generate a constrained consistent high-level objective vector. Within the execution time step covering the management time step, the node state vector and the constrained consistent high-level objective vector are input into the execution network to generate human resource management actions. Human resource management actions are calculated one by one according to job nodes and employee nodes in the execution time step dimension. The constrained consistent high-level objective vector remains unchanged within the same management time step, so that the execution network follows a unified high-level objective in the sequence of human resource management actions.

[0049] Based on the constraint-consistent high-level objective vector and human resource management actions, job allocation arrangements, shift schedules, and training arrangements are generated. These arrangements are then combined into a human resource management plan, which is written into the job allocation table, shift schedule, and training plan table of the human resource management system. At the end of the execution cycle, the execution records and actual results corresponding to the human resource management plan are recorded for subsequent business performance calculations and incremental training of the reinforcement learning network.

[0050] The beneficial effects of this invention are:

[0051] This invention models organizational data, job data, and employee data into a unified organizational hierarchy graph. It simultaneously encodes business scale, personnel capabilities, and historical working hours in the node state vectors, and centrally expresses working hour constraints, budget constraints, staffing constraints, and employment rules in the constraint vectors. This provides a structured and computable state representation for human resource management scenarios in reinforcement learning. Compared to traditional methods based solely on local job features or static rules, this graph-structured state modeling approach can reflect organizational hierarchy and cross-departmental resource relationships within a unified framework. This provides a complete global information foundation for subsequent strategy learning, facilitating the generation of human resource management solutions that coordinate the needs of different organizational levels and job positions.

[0052] This invention constructs management and execution networks based on FeUdal Networks, introducing management and execution time scales to separate long-term human resource allocation goals from short-term human resource management actions. The management network generates unconstrained high-level goal vectors encoding empowerment and compliance information. Then, a mapping unit performs constrained manifold mapping to obtain constrained consistent high-level goal vectors under constraints of working hours, budget, staffing, and employment rules. This shifts compliance requirements to the high-level goal space for constraint during the decision-making process. Simultaneously, the execution network determines human resource management actions based on node state vectors and constrained consistent high-level goal vectors at a short time scale. In the empowerment comparison unit, counterfactual comparisons between human resource management actions and hypothetical actions generate empowerment difference rewards. This ensures that strategy optimization no longer relies solely on business performance but also explicitly considers employee empowerment and development opportunities, facilitating a balance between business performance and employee growth.

[0053] This invention simultaneously incorporates business performance, empowerment-based differential rewards, and constraint violation penalties into its reward structure. Through unified reinforcement learning update rules, it jointly adjusts the parameters of the management network, execution network, mapping unit, and empowerment comparison unit. This allows the learning process to simultaneously consider performance goals, compliance constraints, and empowerment objectives on a global scale, improving the stability and interpretability of strategy convergence. During the deployment phase, relying on the trained reinforcement learning network, it automatically outputs job allocation arrangements, shift schedules, and training arrangements in the human resource management system. This enables it to continuously provide highly adaptable human resource management solutions in the face of organizational restructuring, business fluctuations, and changes in constraints. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of a reinforcement learning-based human resource management and intelligent decision-making method proposed in this invention;

[0056] Figure 2 This is a schematic diagram of the reinforcement learning network structure for a reinforcement learning-based human resource management and intelligent decision-making method proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figures 1-2A reinforcement learning-based method for human resource management and intelligent decision-making includes the following steps:

[0059] Collect organizational data, job data, and employee data; construct an organizational hierarchy diagram; and generate a node state vector and constraint vector for each node.

[0060] The organizational hierarchy diagram and node state vectors are used to form the reinforcement learning state;

[0061] A reinforcement learning network is constructed based on FeUdal Networks. The reinforcement learning network includes a management network, an execution network, a mapping unit, and an empowerment-comparison unit. The management network generates unconstrained high-level target vectors based on the reinforcement learning state over a long time scale. The unconstrained high-level target vectors encode empowerment information and compliance information at the same time.

[0062] The mapping unit performs constrained manifold mapping based on the unconstrained high-level objective vector and the constrained vector to obtain a constrained high-level objective vector;

[0063] The execution network determines human resource management actions based on the node state vector and the constraint-consistent high-level objective vector in a short time scale, and the empowerment comparison unit calculates the empowerment difference reward based on the human resource management actions and the assumed actions.

[0064] Based on business performance, empowerment differentiation rewards, and constraint violations, a reward structure is constructed, and the parameters of the management network, execution network, mapping unit, and empowerment comparison unit are jointly adjusted according to the reward structure.

[0065] In the human resource management system, a trained reinforcement learning network is applied to generate a constrained, consistent high-level objective vector and human resource management actions based on the organizational hierarchy diagram and node state vectors, thereby forming a human resource management plan.

[0066] In this embodiment, the construction of the organizational hierarchy diagram specifically includes:

[0067] Read organizational data, job data, and employee data from the human resources management system and business system. Register each piece of organizational data as an organizational node, each piece of job data as a job node, and each piece of employee data as an employee node. Organizational data must include at least an organization identifier field and a parent organization identifier field. Job data must include at least a job identifier field and an organization identifier field. Employee data must include at least an employee identifier field and a job identifier field.

[0068] Based on the superior organization identifier field, establish management relationship edges between organization nodes, establish job relationship edges between organization nodes and job nodes based on the organization identifier field, establish employee relationship edges between job nodes and employee nodes based on the job identifier field, and combine organization nodes, job nodes, employee nodes, management relationship edges, job relationship edges and employee relationship edges to form an organizational hierarchy diagram;

[0069] For each organization node, job node, and employee node, select fields related to business scale, job content, personnel capabilities, and historical working hours from the corresponding data records, arrange them in a preset order, perform numerical processing, and generate a node status vector;

[0070] For each organization node, job node, and employee node, select fields related to working hours constraints, budget constraints, staffing constraints, and employment rules from the corresponding data records, and arrange them in a preset order to generate a constraint vector;

[0071] In the organizational hierarchy diagram, each organizational node, job node, and employee node is bound with a node state vector and a constraint vector, resulting in an organizational hierarchy diagram containing node state vectors and constraint vectors.

[0072] In this embodiment, the generation of the reinforcement learning state specifically includes:

[0073] In an organizational hierarchy diagram containing node state vectors and constraint vectors, a node number is assigned to each organizational node, job node, and employee node. The node number is then associated with the node type, node state vector, and constraint vector. The node number is registered as an integer index in a preset order, and the node type is registered as a category marker that distinguishes organizational nodes, job nodes, and employee nodes. The node state vector and constraint vector are stored in the same index structure according to the node number.

[0074] Based on the management relationship edges, job relationship edges, and employee relationship edges in the organizational hierarchy diagram, the node numbers at both ends of each relationship edge are recorded according to the node number to generate a graph structure record. The graph structure record includes a set of node numbers and a set of edge connection pairs. Each record in the set of edge connection pairs contains the starting node number and the ending node number, which are used to represent the connection method of management relationship, job relationship, and employee relationship in the organizational hierarchy diagram in the number space.

[0075] The node state vector of each node is arranged in a uniform field order. The non-numerical fields in the node state vector are encoded, and the numerical fields in the node state vector are normalized to form a standardized node state vector. The encoding process converts the text field, enumeration field and label field into a fixed-length numerical representation, and the normalization process scales the business volume field, working hour field and performance field to a preset numerical range.

[0076] In terms of feature dimension, the standardized node state vector and constraint vector are concatenated to form a node feature vector. All node feature vectors are arranged in the order of node number to form a feature matrix. The feature matrix is ​​combined with the graph structure record to form a reinforcement learning state. The reinforcement learning state carries both node feature information and connection relationship information, which is used to represent the state corresponding to the organizational hierarchy graph in the human resource management scenario.

[0077] This invention establishes a unified node numbering, graph structure record, and node feature matrix in the organizational hierarchy graph. It encodes information such as the business status, historical performance, working hour constraints, and budget constraints of organizational nodes, job nodes, and employee nodes into standardized node features within the same index space. These features, along with the connection methods of management relationships, job relationships, and employee relationships, are used to construct a reinforcement learning state. This enables human resource management problems to have a clear, dimensionally unified, and constraint-explicit graph representation on the input side. This facilitates the subsequent learning of cross-level and cross-job human resource allocation strategies by management and execution networks based on FeUdal Networks from a global perspective, reducing training instability and strategy bias caused by feature fragmentation and inconsistent encoding.

[0078] In this embodiment, the construction of the reinforcement learning network specifically includes:

[0079] Set management time scale and execution time scale on the time axis, divide the time axis into management time step sequence and execution time step sequence. Each management time step covers a preset number of execution time steps. Within each management time step, the reinforcement learning state is regarded as the same batch input on the management network side. Within each execution time step, record the changes of node state vector on the execution network side. Use management time steps to represent the frequency of high-level decision-making and use execution time steps to represent the frequency of human resource management action updates.

[0080] A management network is constructed based on the FeUdal Networks framework, which includes a state encoding layer and a target generation layer. The state encoding layer generates a management hidden representation based on the reinforcement learning state at each management time step. The management hidden representation is a fixed-dimensional vector. The target generation layer generates an unconstrained high-level target vector based on the management hidden representation. The unconstrained high-level target vector is divided into an empowerment information part and a compliance information part in the vector dimension. The empowerment information part is used to represent the target tendency related to employee capability improvement and task authorization, and the compliance information part is used to represent the target tendency related to time constraints, budget constraints, and staffing constraints.

[0081] An execution network is constructed based on the FeUdal Networks framework, which includes a local state processing layer and a target condition processing layer. In each execution time step, the local state processing layer generates an execution hidden representation based on the node state vector. In each execution time step, the target condition processing layer generates an execution feature vector based on the execution hidden representation and the unconstrained high-level target vector. In execution time steps covering the same management time step, the unconstrained high-level target vector is kept unchanged. The unchanging unconstrained high-level target vector guides the execution feature vector to adjust human resource management actions around a unified high-level target in multiple execution time steps.

[0082] In a reinforcement learning network, mapping units and empowerment contrast units are set up. The mapping units are connected to the output of the unconstrained high-level objective vector of the management network, and the empowerment contrast units are connected to the output of the execution feature vector of the execution network. The management network, execution network, mapping units, and empowerment contrast units constitute a reinforcement learning network structure based on FeUdal Networks, which provides high-level objective information and execution feature information for constrained manifold mapping and empowerment difference calculation.

[0083] This invention introduces both management and execution time scales on the time axis and constructs a hierarchical structure based on FeUdalNetworks that distinguishes between management and execution networks. It maps reinforcement learning states at a high level to unconstrained high-level target vectors that simultaneously encode empowerment and compliance information. The execution network then adjusts the node state vectors around stable, unchanging high-level targets within a short timescale. Simultaneously, the mapping unit and empowerment comparison unit provide the target information required for constrained manifold mapping and the execution characteristic information required for human resource management actions, respectively. This decouples high-level long-term goal planning from low-level frequent scheduling, allocation, and training decisions while maintaining synergy. This improves the stability and convergence efficiency of strategy learning and enhances fine-grained control over compliance requirements and employee empowerment in complex constraint and business fluctuation scenarios.

[0084] In this embodiment, the generation of the constraint-consistent high-level objective vector specifically includes:

[0085] Based on the time limit field, budget limit field, staffing limit field, and employment rule field in the constraint vector, they are arranged in a preset order to form a constraint feature vector. Threshold parameters and weight parameters are set for each field in the constraint feature vector. The threshold parameters include a lower threshold representing the minimum allowable value and an upper threshold representing the maximum allowable value. The weight parameters are used to represent the degree of influence of the corresponding constraint on the adjustment of the high-level target during the constraint manifold mapping process. The threshold parameters and weight parameters are preset based on historical human resource management data and management regulations.

[0086] The unconstrained high-level target vector is divided along the vector dimension to obtain the empowerment information part and the compliance information part. According to the preset mapping relationship, a one-to-one correspondence is established between each dimension of the compliance information part and the working hour limit field, budget limit field, staffing limit field and employment rule field in the constraint feature vector. In the one-to-one correspondence, the type of constraint field that controls each compliance information dimension and the corresponding threshold parameters and weight parameters are recorded, so that the numerical changes of the compliance information part are directly controlled by the constraint feature vector.

[0087] In the mapping unit, constraint manifold mapping is performed based on constraint vectors and constraint feature vectors. The values ​​are adjusted along the constraint directions of each dimension of the compliance information part, and the values ​​of the compliance information part are restricted to the feasible interval determined by the working hour limit field, budget limit field, staffing limit field, and employment rule field. The feasible interval is determined by the threshold parameter. The lower limit value and upper limit value of the feasible interval are calculated for each dimension of compliance information. When the value of the compliance information part is less than the lower limit value, the value of the compliance information part is raised to the lower limit value. When the value of the compliance information part is greater than the upper limit value, the value of the compliance information part is reduced to the upper limit value. The smoothness of the value adjustment is controlled according to the corresponding weight parameter.

[0088] In the mapping unit, the value of the empowerment information part remains unchanged. The compliance information part after constraint manifold mapping is recombined with the empowerment information part to generate a constraint-consistent high-level target vector. The constraint-consistent high-level target vector contains both the empowerment information part and the compliance information part after constraint manifold mapping in the vector dimension, and keeps the dimension of the constraint-consistent high-level target vector consistent with the dimension of the unconsistent high-level target vector.

[0089] This invention explicitly embeds time constraints, budget constraints, staffing constraints, and employment rules into a high-level target space in the form of constraint feature vectors, thresholds, and weight parameters. It performs constraint manifold mapping on the compliance information portion of the unconstrained high-level target vector, smoothly pruning and adjusting it according to feasible intervals and weight parameters in each dimension, while keeping the empowerment information portion from being directly weakened. This achieves role-based control of compliance and empowerment objectives within the same high-level target vector, enabling subsequent human resource management actions to inherently meet employment constraints and preserve employee development space during the strategy learning stage, reducing reliance on ex-post rule corrections that lead to strategy distortions and scheduling rework.

[0090] In this embodiment, the calculation of the empowerment difference reward specifically includes:

[0091] At each execution time step, for the job nodes and employee nodes in the organizational hierarchy diagram, select the node state vectors related to the job nodes and employee nodes from the set of node state vectors. Concatenate the node state vectors with the high-level target vectors consistent with the constraints adopted in the current management time step in the feature dimension to form the execution input vector. Input the execution input vector into the local state processing layer and the target condition processing layer in the execution network to generate the execution hidden representation. The execution hidden representation is obtained as a fixed-dimensional numerical vector through linear operation and nonlinear activation, and maintains the mapping relationship with the corresponding job nodes and employee nodes in each execution time step.

[0092] The human resource management action score vector is calculated by action mapping based on the execution hidden representation. Human resource management actions are determined on the human resource management action score vector according to the preset action selection rules. Human resource management actions are recorded at the execution time step. Each dimension of the human resource management action score vector corresponds one-to-one with the preset human resource management actions. The human resource management action with the highest score is selected as the execution result by comparing the scores of each dimension. The execution result is then associated with and stored with the identification information of the job node and the employee node.

[0093] In the empowerment comparison unit, the node state vector, the constraint-consistent high-level target vector, and human resource management actions are used as inputs. According to the preset action combination rules, several actions are selected from the preset action set to form a hypothetical action set. The human resource management actions and hypothetical actions are combined one by one to form a comparison sample. The empowerment index value is calculated based on each comparison sample. The empowerment index value is weighted according to the skill matching degree, ability improvement space, and growth path length. The empowerment difference reward is generated based on the difference between the empowerment index value corresponding to the human resource management action and the empowerment index value corresponding to the hypothetical action, and the empowerment difference reward is registered at the execution time step.

[0094] In this embodiment, the construction and processing of the reward structure specifically includes:

[0095] Based on the business indicators recorded in the human resources management system, data related to output quantity, service quality and cost consumption are summarized for each management time step. The business performance value is obtained according to the preset calculation function. The business performance value is linearly scaled between the preset lower limit and the preset upper limit and mapped to the external reward range to form the external reward of the management time step. The external reward value is then recorded on the management time step.

[0096] Based on the empowerment difference reward output by the empowerment comparison unit at each execution time step, the empowerment difference reward is weighted and summed according to a preset weight coefficient within the execution time step covered by each management time step to obtain the intrinsic reward of the management time step. At the same time, the empowerment difference reward corresponding to each execution time step is retained as the intrinsic reward of the execution time step. The intrinsic reward of the management time step and the intrinsic reward of the execution time step are associated and stored with the corresponding time step respectively.

[0097] Based on the degree of deviation between the compliance information part of the constraint consistency high-level target vector obtained by the mapping unit and the working hour limit field, budget limit field, staffing limit field and employment rule field in the constraint vector, and based on the working hour limit violation, budget limit violation and staffing limit violation triggered by the execution network in human resource management actions, the constraint violation cost value is calculated. At the management time step level, the constraint violation cost value covering the execution time step is aggregated. At the execution time step level, the constraint violation cost value of the corresponding execution time step is directly recorded. The aggregation result and the recording result are converted into management time step penalty and execution time step penalty, respectively.

[0098] Based on preset coefficients, external rewards, internal rewards, and penalties are linearly combined to obtain the total reward value for each management time step and the total reward value for each execution time step. Based on the total reward value, a reinforcement learning strategy is used to update the rules and simultaneously adjust the parameters of the management network, execution network, mapping unit, and empowerment comparison unit, thus completing a unified parameter update on both the management and execution time scales.

[0099] In this embodiment, the generation of the human resource management plan specifically includes:

[0100] The trained management network, execution network, mapping unit, and empowerment comparison unit are loaded into the human resource management system. The management network, execution network, mapping unit, and empowerment comparison unit are combined into a reinforcement learning network, and a data interface with organizational data, job data, and employee data is established. The parameters of the management network, execution network, mapping unit, and empowerment comparison unit are written into the system's running memory. The mapping relationship between the input end of the reinforcement learning network and the fields of organizational data, job data, and employee data, as well as the mapping relationship between the output end of the reinforcement learning network and the fields of job configuration results, scheduling results, and training results are completed.

[0101] The organizational hierarchy diagram, node state vector, and constraint vector are updated based on the current organizational data, job data, and employee data. The organizational hierarchy diagram, node state vector, and constraint vector are combined to form the reinforcement learning state in the operation phase. The reinforcement learning state in the operation phase is refreshed once at the beginning of each management time step based on the latest organizational data, job data, and employee data. The organizational hierarchy diagram is updated by parsing the organizational identifier field, job identifier field, and employee identifier field. The node state vector is updated by parsing the business field and personnel field. The constraint vector is updated by parsing the restriction field and rule field.

[0102] Within each management time step, the reinforcement learning state of the running phase is input into the management network to generate an unconstrained high-level objective vector. The constrained manifold mapping is performed through the mapping unit to generate a constrained consistent high-level objective vector. Within the execution time step covering the management time step, the node state vector and the constrained consistent high-level objective vector are input into the execution network to generate human resource management actions. Human resource management actions are calculated one by one according to job nodes and employee nodes in the execution time step dimension. The constrained consistent high-level objective vector remains unchanged within the same management time step, so that the execution network follows a unified high-level objective in the sequence of human resource management actions.

[0103] Based on the constraint-consistent high-level objective vector and human resource management actions, job allocation arrangements, shift schedules, and training arrangements are generated. These arrangements are then combined into a human resource management plan, which is written into the job allocation table, shift schedule, and training plan table of the human resource management system. At the end of the execution cycle, the execution records and actual results corresponding to the human resource management plan are recorded for subsequent business performance calculations and incremental training of the reinforcement learning network.

[0104] Example 1:

[0105] To verify the feasibility of this invention in practice, it was applied to the human resource management system of a human resource-intensive enterprise with a group headquarters, business units, and regional service centers. The enterprise has approximately 3,000 employees and a multi-layered organizational structure, including headquarters positions responsible for overall management, frontline customer service positions, and back-office support positions. Most positions operate on a shift system, and staffing needs fluctuate significantly with business volume. Before introducing this invention, the enterprise relied on a traditional rule-based scheduling scheme. The HR department ranked employees based on a fixed staffing table and scheduling template, combined with historical working hours, performance scores, and simple weighting rules, with business supervisors making manual adjustments. On the surface, this scheme could roughly meet staffing needs, but during peak order periods or periods of significant regional business fluctuations, issues frequently arose such as concentrated overtime for frontline staff, persistent staff shortages on some shifts, budget execution deviating from the annual plan, and job matching relying heavily on supervisors' personal experience. Employee satisfaction surveys showed a high proportion of feedback regarding "uneven scheduling" and "lack of transparency in training opportunities." The HR department also generally reported a significant increase in scheduling and job adjustments before and after major business promotions or regional adjustments.

[0106] In this scenario, the organizational data, job data, and employee data from the company's existing human resource management system are first loaded into the method of this invention according to organizational hierarchy. The group headquarters, various business units, and regional service centers are modeled as organizational nodes, customer service specialists, operations specialists, scheduling administrators, quality inspectors, etc., are modeled as job nodes, and on-the-job employees are modeled as employee nodes. Management relationship edges, job relationship edges, and employee relationship edges are established through the superior organization field, the subordinate organization field, and the subordinate job field, forming a human resource organizational hierarchy diagram covering the entire organization. Around this hierarchy diagram, a node state vector is constructed for each node, encoding information such as business volume, number of work orders, service duration, performance rating, skill tags, and historical working hours in a fixed order. At the same time, working hour limits, rest intervals, departmental budgets, job positions, and employment rules are extracted from the system configuration to construct a constraint vector, so that each node has information reflecting business and personnel status as well as clear constraint information. After generating reinforcement learning states based on node state vectors and constraint vectors, FeUdal Networks are used to construct management and execution networks. The management network outputs an unconstrained high-level target vector based on the reinforcement learning states over a longer timescale. Within this target vector, empowerment information and compliance information are distinguished. A mapping unit then performs constraint manifold mapping using the constraint vectors, "pushing back" the compliance information to the feasible ranges of work hours, budgets, and staffing, thus obtaining a constraint-consistent high-level target vector while preserving the expression space of empowerment information. The execution network, on a shorter timescale, uses node state vectors and constraint-consistent high-level target vectors as input to generate human resource management actions such as job allocation, shift scheduling, and training arrangements. An empowerment comparison unit designs a set of reasonable hypothetical actions around key positions. It performs a counterfactual comparison between actual and hypothetical actions in terms of skill matching, career path alignment, and training coverage, obtaining empowerment difference rewards. These rewards are then combined with business indicators such as order completion rate, service timeout rate, customer satisfaction, and labor costs to form external rewards, and constraint violation penalties are superimposed to form a complete reward structure used for iterative updates to the management network, execution network, mapping unit, and empowerment comparison unit.

[0107] After completing offline training based on historical data, the converged reinforcement learning network was deployed in parallel with the existing human resource management system. During a continuous business cycle, both the traditional rule-engine scheduling scheme and the reinforcement learning-based human resource management and intelligent decision-making method of this invention were run simultaneously. For ease of comparison, a "dual-track operation" mode was implemented in the call center business unit and two regional customer service centers. Some teams continued to use the traditional rule-engine scheduling scheme, while others used the human resource management scheme generated by the method of this invention, maintaining consistency in business volume, product structure, and policy constraints. Observational indicators included average job vacancy rate, overall overtime rate, budget execution deviation, quarterly training coverage rate, and employee job transfer request fulfillment rate. The statistical results are summarized in Table 1.

[0108] Table 1. Comparison between traditional rule-based scheduling schemes and the reinforcement learning-based scheme of this invention.

[0109] Indicator Traditional rule engine scheduling scheme The present application is based on a reinforcement learning scheme Improvement range Average job gap rate 7.8% 3.1% Reduced by 4.7 percentage points Overall overtime rate 18.5% 10.2% Reduced by 8.3 percentage points Budget execution deviation (absolute value) 9.6% 4.2% Reduced by 5.4 percentage points Quarterly training coverage rate 61.4% 82.7% Increased by 21.3 percentage points Employee job transfer demand satisfaction rate 46.2% 71.9% Increased by 25.7 percentage points

[0110] As can be seen from the table, under the premise of unchanged organizational structure, employment policies, and business scale, the human resource management solution output by the method of this invention is superior to the traditional rule engine scheduling solution in terms of job vacancy control, overtime distribution, budget execution stability, and employee development-related indicators: the average job vacancy rate has decreased significantly, indicating that cross-departmental and cross-shift collaborative allocation is smoother through organizational hierarchy modeling and high-level goal guidance; the overall overtime rate has decreased significantly, reflecting that constrained manifold mapping and compliance information modeling make it easier for the system to avoid overtime arrangements during the scheduling stage, no longer relying too much on manual adjustments afterward; budget execution deviation has decreased, and fluctuations in human resource costs with changes in business are more controllable; the training coverage rate and job transfer demand satisfaction rate have significantly improved, indicating that empowerment differential rewards have played a practical role in the strategy learning process, and the human resource management solution no longer just focuses on "filling in" people for the current scheduling, but takes into account medium- and long-term capability building and individual employee development, so that both management and business departments can perceive the improvement in strategy quality during use.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A human resource management and intelligent decision-making method based on reinforcement learning, characterized in that, The method comprises the following steps: Collecting organization data, post data and employee data, constructing an organization hierarchy diagram, generating a node state vector and a constraint vector for each node; The organization hierarchy diagram and the node state vector form a reinforcement learning state; Based on FeUdal Networks, a reinforcement learning network is constructed, which includes a management network, an execution network, a mapping unit and an empowerment comparison unit. The management network generates an unconstrained high-level target vector based on the reinforcement learning state in a long time scale, which encodes empowerment information and compliance information at the same time; The mapping unit performs constraint manifold mapping based on the unconstrained high-level target vector and the constraint vector to obtain a constraint-consistent high-level target vector; The execution network determines human resource management actions based on the node state vector and the constraint-consistent high-level target vector in a short time scale. The empowerment comparison unit calculates the empowerment difference reward based on the human resource management action and the hypothetical action; A reward structure is constructed based on business performance, empowerment difference reward and constraint violation, and the parameters of the management network, execution network, mapping unit and empowerment comparison unit are adjusted jointly according to the reward structure; The trained reinforcement learning network is applied in the human resource management system to generate constraint-consistent high-level target vectors and human resource management actions based on the organization hierarchy diagram and the node state vector, forming a human resource management scheme. 2.The human resource management and intelligent decision-making method based on reinforcement learning according to claim 1, wherein, The construction of the organization hierarchy diagram specifically comprises: Reading organization data, post data and employee data from a human resource management system and a business system, registering each piece of organization data as an organization node, registering each piece of post data as a post node, and registering each piece of employee data as an employee node; According to the superior organization identifier field, management relationship edges are established between organization nodes. According to the organization identifier field, post relationship edges are established between organization nodes and post nodes. According to the post identifier field, employee relationship edges are established between post nodes and employee nodes. The organization nodes, post nodes, employee nodes, management relationship edges, post relationship edges and employee relationship edges are combined to form an organization hierarchy diagram; For each organization node, post node and employee node, fields related to business size, work content, personnel capacity and historical working hours are selected from the corresponding data records, arranged in a predetermined order and subjected to numerical processing to generate a node state vector; For each organization node, post node and employee node, fields related to working hour limit, budget limit, establishment limit and labor rules are selected from the corresponding data records and arranged in a predetermined order to generate a constraint vector; In the organization hierarchy diagram, the node state vector and the constraint vector are bound to each organization node, post node and employee node to obtain an organization hierarchy diagram containing node state vectors and constraint vectors. 3.The method of claim 1, wherein, The generation of the reinforcement learning state specifically comprises: In the organization hierarchy diagram containing node state vectors and constraint vectors, each organization node, post node and employee node is assigned a node number, and the node number is associated with the node type, node state vector and constraint vector; According to the management relationship edges, post relationship edges and employee relationship edges in the organization hierarchy diagram, the node numbers of the two nodes at the ends of each relationship edge are recorded according to the node numbers to generate a graph structure record; The node state vectors of each node are arranged in a uniform field order, a numerical conversion is performed on non-numerical fields in the node state vectors, and a standardization processing is performed on numerical fields in the node state vectors to obtain standardized node state vectors; The standardized node state vectors are spliced with constraint vectors in a feature dimension to form node feature vectors, all node feature vectors are arranged in a node number order to form a feature matrix, and the graph structure record and the feature matrix are combined to form a reinforcement learning state.

4. The method of claim 1, wherein the method is based on reinforcement learning. The reinforcement learning network is constructed specifically as follows: Management time scales and execution time scales are set on a time axis, the time axis is divided into management time step sequences and execution time step sequences, and each management time step covers a preset number of execution time steps; A management network is constructed, the management network includes a state coding layer and a target generation layer based on a FeUdal Networks framework, the state coding layer generates a management hidden representation according to the reinforcement learning state in each management time step, and the target generation layer generates an unconstrained high-level target vector according to the management hidden representation, the unconstrained high-level target vector is divided into an empowerment information part and a compliance information part in a vector dimension; An execution network is constructed, the execution network includes a local state processing layer and a target condition processing layer based on the FeUdal Networks framework, the local state processing layer generates an execution hidden representation according to the node state vector in each execution time step, and the target condition processing layer generates an execution feature vector according to the execution hidden representation and the unconstrained high-level target vector, the unconstrained high-level target vector remains unchanged in execution time steps covering the same management time step; A mapping unit and an empowerment comparison unit are set in the reinforcement learning network, the mapping unit is connected with the unconstrained high-level target vector output by the management network, the empowerment comparison unit is connected with the execution feature vector output by the execution network, and the management network, the execution network, the mapping unit and the empowerment comparison unit form a reinforcement learning network based on the FeUdal Networks.

5. The method of claim 1, wherein the method is based on reinforcement learning. The generation of the constraint consistent high-level target vector specifically includes: According to the work time limit field, the budget limit field, the compilation limit field and the labor rule field in the constraint vector, a constraint feature vector is arranged in a preset order, a threshold parameter and a weight parameter are set for each field in the constraint feature vector; The unconstrained high-level target vector is divided in a vector dimension to obtain the empowerment information part and the compliance information part, and a one-to-one correspondence is established between each dimension of the compliance information part and the work time limit field, the budget limit field, the compilation limit field and the labor rule field in the constraint feature vector according to a preset mapping relationship; In the mapping unit, constraint manifold mapping is performed according to the constraint vector and the constraint feature vector, the values along the compliance information part dimensions are adjusted in a constraint direction, and the values of the compliance information part are limited in a feasible interval determined by the work time limit field, the budget limit field, the compilation limit field and the labor rule field. The compliance information part of the mapped constraint manifold is recombined with the empowerment information part to generate a constraint-consistent high-level target vector, and the dimension of the constraint-consistent high-level target vector is kept consistent with the dimension of the un-constrained high-level target vector.

6. The method of claim 1, wherein the method is based on reinforcement learning. The calculation of the empowerment difference reward specifically includes: At each execution time step, for a post node and an employee node in the organizational hierarchy diagram, a node state vector related to the post node and the employee node is selected from a set of node state vectors, the node state vector is spliced with a constraint-consistent high-level target vector used in a current management time step in a feature dimension to form an execution input vector, and the execution input vector is input into a local state processing layer and a target condition processing layer in the execution network to generate an execution hidden representation; An action mapping is calculated from the execution hidden representation to obtain a human resource management action score vector, a human resource management action is determined on the human resource management action score vector according to a preset action selection rule, and the human resource management action is recorded at the execution time step; In the empowerment comparison unit, the node state vector, the constraint-consistent high-level target vector, and the human resource management action are input, a set of hypothetical actions is constructed according to a preset action combination rule, the human resource management action is combined with the hypothetical actions one by one to form comparison samples, an empowerment index value is calculated according to each comparison sample, and an empowerment difference reward is generated according to the difference between the empowerment index value corresponding to the human resource management action and the empowerment index value corresponding to the hypothetical action.

7. The method of claim 1, wherein the method is based on reinforcement learning. The construction and processing of the reward structure specifically include: A business performance value is calculated for each management time step according to a business index recorded by the human resource management system, the business performance value is linearly scaled and mapped to an external reward interval to form an external reward of the management time step; The empowerment difference rewards output by the empowerment comparison unit at each execution time step are weighted and summed according to a preset weight coefficient within the execution time steps covered by the management time step to obtain an intrinsic reward within the management time step, while the empowerment difference rewards of each execution time step are retained as intrinsic rewards of the execution time steps; A constraint violation penalty is defined according to the degree of deviation between the compliance information part of the constraint-consistent high-level target vector obtained by the mapping unit and the constraint vector, and the working hour limit violation, the budget limit violation, and the staffing limit violation triggered by the human resource management action in the execution network, and the constraint violation penalty is calculated at the management time step and the execution time step; The external reward, the intrinsic reward, and the constraint violation penalty are linearly combined according to a preset coefficient to obtain a total reward value of each management time step and each execution time step, and the management network parameters, the execution network parameters, the mapping unit parameters, and the empowerment comparison unit parameters are synchronously adjusted according to the total reward value using a reinforcement learning strategy update rule. 8.The method of claim 1, wherein, The generation of the human resource management scheme specifically includes: The trained management network, the execution network, the mapping unit, and the empowerment comparison unit are loaded in the human resource management system, the management network, the execution network, the mapping unit, and the empowerment comparison unit are combined into a reinforcement learning network, and a calling interface with organizational data, post data, and employee data is established; According to the current organization data, the post data and the employee data, the organization hierarchy diagram, the node state vector and the constraint vector are updated, and the organization hierarchy diagram, the node state vector and the constraint vector are combined to form a reinforcement learning state of an operation stage; In each management time step, the reinforcement learning state is input into a management network to generate an unconstrained high-level target vector, constraint manifold mapping is performed through a mapping unit to generate a constraint-consistent high-level target vector, and the node state vector and the constraint-consistent high-level target vector are input into an execution network in an execution time step covering the management time step to generate a human resource management action; According to the constraint-consistent high-level target vector and the human resource management action, a post configuration arrangement, a scheduling arrangement and a training arrangement are generated, the post configuration arrangement, the scheduling arrangement and the training arrangement are combined into a human resource management scheme, and the human resource management scheme and an execution result of the human resource management scheme are recorded in a human resource management system.

Images