Team collaboration training method and system based on scene reproduction and multi-agent collaboration

Abstract

This invention provides a team collaboration training method and system based on scenario replication and multi-agent collaboration, involving the technical field of team collaboration. The method involves constructing interactive scenarios by establishing a team collaboration knowledge base, using deep reinforcement learning to train initial agent behavior patterns, building a multi-agent collaborative learning network enhanced by knowledge distillation, using counterfactual reasoning to analyze the impact of collaborative behavior, dynamically tracking the evolution of capabilities, and outputting collaborative optimization solutions. This invention can effectively improve team collaboration effectiveness, accurately identify collaborative shortcomings, and achieve continuous optimization of team capabilities.

Description

Technical Field

[0001] The present invention relates to team collaboration technology, and in particular to a team collaboration training method and system based on scene reproduction and multi-agent collaboration. Background Art

[0002] Teamwork is a crucial core competency in modern organizations, especially in complex environments. Effective teamwork can significantly improve work efficiency and the quality of task completion. Traditional teamwork training relies primarily on case studies, role-playing, and field exercises. While these methods can enhance team members' collaborative awareness to a certain extent, they often lack specificity and precision. With the rapid development of artificial intelligence (AI), teamwork training methods based on multi-agent systems are gaining popularity. These methods can simulate real-world task scenarios, providing a richer and more precise training experience.

[0003] In the field of team collaboration training, existing technologies often employ static scenario designs and simplified interaction models, which fail to truly reflect the challenges of team collaboration in complex environments. Furthermore, traditional training methods often suffer from incomplete effectiveness evaluation systems, disconnection between training scenarios and actual work, and a lack of personalized adaptation during the training process, which effectively limit the improvement of team collaboration capabilities.

[0004] Existing team collaboration training systems lack effective use of real historical data. Training scenarios are often preset standardized environments that cannot accurately reproduce the specific challenging situations encountered by the team in actual work. This leads to a large gap between training content and actual needs, and training results are difficult to translate into actual work capabilities.

[0005] Traditional collaborative training methods lack precise quantitative evaluation mechanisms and are unable to objectively measure the synergy between team members and their impact on task outcomes. Training feedback usually relies on subjective judgment, making it difficult to identify key behaviors and capability shortcomings in the collaborative process, resulting in a lack of accuracy and specificity in training adjustments.

[0006] Existing team collaboration training systems generally lack adaptive learning capabilities and are unable to dynamically adjust training content and difficulty based on the performance and ability levels of team members. They are also unable to capture the evolution of team capabilities, resulting in inefficient training and difficulty in meeting the needs of different stages of team development and the cultivation of differentiated collaboration capabilities. Summary of the Invention

[0007] The embodiments of the present invention provide a team collaboration training method and system based on scene reproduction and multi-agent collaboration, which can solve the problems in the existing technology.

[0008] A first aspect of an embodiment of the present invention provides a team collaboration training method based on scene reproduction and multi-agent collaboration, comprising:

[0009] Establishing a team collaboration knowledge base based on pre-acquired team collaboration historical data, constructing a multi-agent interaction scenario based on the team collaboration knowledge base, mapping virtual environment parameters, agent role information, and task goal information in the team collaboration knowledge base to the multi-agent interaction scenario, and generating a scenario reproduction training environment;

[0010] Generate an intelligent agent in the scenario reproduction training environment, and train the initial behavior pattern of the intelligent agent through a deep reinforcement learning model based on the historical behavior data in the team collaboration knowledge base;

[0011] Constructing a knowledge distillation-enhanced multi-agent collaborative learning network, and feeding back collaborative behavior data generated by team members interacting with the agent in the scenario reproduction training environment to the multi-agent collaborative learning network in real time;

[0012] Based on the collaborative behavior data, the impact of team collaborative behavior on task results is analyzed using counterfactual reasoning methods, and a causal relationship diagram between collaborative behavior and task performance is constructed. The causal relationship diagram dynamically quantifies the strength of the synergy between team members and uses a temporal attention mechanism to track the evolution of capabilities.

[0013] According to the capability evolution trajectory, a team collaboration evaluation report is generated; based on the team collaboration evaluation report, a team collaboration optimization plan is output, the parameter configuration of the scenario reproduction training environment is adjusted in a targeted manner, and the training strategy of the multi-agent collaborative learning network is updated.

[0014] Mapping the virtual environment parameters, agent role information, and task goal information in the team collaboration history data to the multi-agent interaction scene to generate a scene reproduction training environment includes:

[0015] Constructing a feature vector space mapping model, and extracting features of virtual environment parameters, agent role information, and task target information based on the feature vector space mapping model to generate a target scene feature vector, wherein the target scene feature vector includes an environment feature component, a role feature component, and a task feature component;

[0016] constructing a multi-objective optimization function based on the target scene feature vector, calculating feature mapping losses for the environment feature component, the role feature component, and the task feature component according to the multi-objective optimization function, and optimizing the target scene feature vector based on a weighted result of the feature mapping losses;

[0017] The optimized target scene feature vector is input into the generator of the generative adversarial network, and the generator reconstructs the target scene feature vector through a multi-layer transposed convolutional network to generate an initial scene structure including environment layout parameters, character attribute parameters, and task configuration parameters;

[0018] Building a scene difficulty adaptive adjustment module based on the initial scene structure, the scene difficulty adaptive adjustment module calculates a scene difficulty coefficient according to a preset performance indicator, and dynamically adjusts the initial scene structure based on the scene difficulty coefficient;

[0019] The adjusted initial scene structure is instantiated as a training environment, and an intelligent agent interaction interface and a task evaluation module are configured in the training environment. The task evaluation module dynamically adjusts the evaluation criteria based on the scene difficulty coefficient to achieve adaptive generation of a scene reproduction training environment.

[0020] Constructing a multi-objective optimization function based on the target scene feature vector, and calculating feature mapping losses for the environment feature component, the role feature component, and the task feature component according to the multi-objective optimization function respectively include:

[0021] Constructing a feature component representation model, encoding the environment feature component, the role feature component, and the task feature component into feature vectors based on the feature component representation model, and generating an initial feature vector set;

[0022] Constructing a multi-objective mapping loss function based on the initial feature vector set, wherein the multi-objective mapping loss function performs weighted calculation on the environment feature mapping loss, the role feature mapping loss, and the task feature mapping loss to generate a base mapping loss value;

[0023] Establishing a feature causal graph structure using the initial feature vector set, wherein the nodes of the feature causal graph structure are feature vectors in the initial feature vector set, the edges of the feature causal graph structure represent causal relationships between feature vectors, and calculating the causal strength between feature nodes through causal intervention operations;

[0024] The causal strength is integrated with the base mapping loss value to construct a causal enhancement overall loss function; parameter gradients are calculated based on the overall loss function, and parameters of the feature component representation model are updated based on the parameter gradients;

[0025] Based on the updated feature component representation model, corresponding weight coefficients are calculated for the environment feature mapping loss, the role feature mapping loss, and the task feature mapping loss to obtain the feature mapping loss of each feature component.

[0026] Generating an agent in the scenario reproduction training environment and training the agent's initial behavior pattern through a deep reinforcement learning model based on historical behavior data in the team collaboration knowledge base includes:

[0027] Generate an agent in the scenario reproduction training environment, extract historical behavior data from a team collaboration knowledge base, and build a deep reinforcement learning model based on the historical behavior data, wherein the deep reinforcement learning model includes a behavior state space, an action space, and a reward function;

[0028] The agent is iteratively trained using the deep reinforcement learning model, the policy network parameters of the agent are updated according to the feedback information of the reward function, and the initial behavior pattern of the agent is trained.

[0029] Based on the collaborative behavior data, the impact of team collaborative behavior on task results is analyzed through counterfactual reasoning methods. A causal relationship diagram between collaborative behavior and task performance is constructed. The causal relationship diagram dynamically quantifies the intensity of the synergy between team members. The ability evolution trajectory is tracked in combination with the temporal attention mechanism, including:

[0030] Constructing a collaborative behavior vector and a task result vector based on the collaborative behavior data, calculating the causal effect of each behavioral feature in the collaborative behavior vector on each result indicator in the task result vector using a causal interference model, wherein the causal effect represents the intensity of the impact of the collaborative behavior on the task result, and constructing a causal relationship diagram between the team collaborative behavior and the task performance based on the causal effect;

[0031] Calculating the synergy strength among team members based on the causal relationship graph, combining the timing weight and the attention weight to weight the causal effect, and obtaining the dynamic synergy strength;

[0032] Constructing a team capability phase space based on the dynamic synergy effect strength, calculating a capability index based on the team capability phase space, wherein the capability index is used to characterize the stability of the team capability evolution, and inputting the capability index into a multi-scale temporal attention mechanism;

[0033] The multi-scale temporal attention mechanism calculates the dynamic association weights of temporal features using the query matrix, the key matrix, and the value matrix, and constructs a Jacobian matrix based on the dynamic association weights, wherein the Jacobian matrix describes the local linear features of the capability phase space;

[0034] The capability index and the Jacobian matrix are combined to calculate the bifurcation discriminant, and the transition probability of the team capability is predicted based on the bifurcation discriminant, where the transition probability indicates the turning point of the capability evolution; the dynamic association weight is combined with the transition probability to track and predict the evolution trajectory of the team capability.

[0035] Calculating a bifurcation discriminant by combining the capability index and the Jacobian matrix, and predicting the transition probability of team capability based on the bifurcation discriminant includes:

[0036] Combining the capability index with the eigenvalue of the Jacobian matrix to construct a bifurcation parameter vector, wherein the bifurcation parameter vector includes the capability index and the maximum real part of the eigenvalue;

[0037] A bifurcation discriminant is calculated based on the bifurcation parameter vector through the determinant of the Jacobian matrix and the exponential smoothing term, and the bifurcation discriminant is used to detect the mutation characteristics of the team capability status;

[0038] Calculating a critical value of the bifurcation discriminant within an observation time window, the critical value representing a threshold condition for a transition in the team capability state, and inputting a difference between the bifurcation discriminant and the critical value into a nonlinear mapping function;

[0039] The nonlinear mapping function is constructed based on the sigmoid function to calculate the transition probability of the team's ability, and the transition probability represents the possibility of a sudden change in the team's ability;

[0040] A confidence interval of the transition probability is constructed, and the upper and lower bounds of the confidence interval are calculated by an uncertainty function. The uncertainty function is dynamically adjusted as the bifurcation discriminant changes, thereby achieving an accurate prediction of the team capability transition.

[0041] Generating a team collaboration evaluation report based on the capability evolution trajectory; outputting a team collaboration optimization plan based on the team collaboration evaluation report, adjusting the parameter configuration of the scenario reproduction training environment in a targeted manner, and updating the training strategy of the multi-agent collaborative learning network include:

[0042] The capability evolution trajectory includes multidimensional capability indicators. Based on these multidimensional capability indicators, the individual capability growth rate and team collaboration efficiency are calculated. The individual capability growth rate represents the speed of capability improvement, and the team collaboration efficiency represents the collaboration effect. The individual capability growth rate and the team collaboration efficiency are weighted and combined to obtain a comprehensive team collaboration score, which reflects the overall performance of the team.

[0043] A multidimensional evaluation system is constructed based on the comprehensive team collaboration score, the multidimensional evaluation system including individual dimensions, team dimensions, and task dimensions, and a team collaboration evaluation report is generated; an optimization objective function is constructed based on the team collaboration evaluation report, the optimization objective function combines performance loss and optimization cost, and generates a team collaboration optimization plan based on resource constraints, time constraints, and effect constraints;

[0044] The team collaboration optimization plan is converted into the parameter configuration of the scenario reproduction training environment, where the parameter configuration includes a training difficulty parameter and a scenario complexity parameter. The scenario reproduction training environment is dynamically adjusted, and the training strategy of the multi-agent collaborative learning network is updated based on the parameter configuration.

[0045] A second aspect of an embodiment of the present invention provides a team collaboration training system based on scene reproduction and multi-agent collaboration, including:

[0046] The first unit is configured to establish a team collaboration knowledge base based on pre-acquired team collaboration historical data, construct a multi-agent interaction scenario based on the team collaboration knowledge base, map virtual environment parameters, agent role information, and task objective information in the team collaboration knowledge base to the multi-agent interaction scenario, and generate a scenario reproduction training environment;

[0047] The second unit is configured to generate an intelligent agent in the scenario reproduction training environment, and train an initial behavior pattern of the intelligent agent through a deep reinforcement learning model based on the historical behavior data in the team collaboration knowledge base;

[0048] The third unit is used to build a multi-agent collaborative learning network enhanced by knowledge distillation, and to feed back the collaborative behavior data generated by the team members interacting with the agents in the scenario reproduction training environment to the multi-agent collaborative learning network in real time;

[0049] The fourth unit is used to analyze the impact of team collaboration behavior on task outcomes using counterfactual reasoning based on the collaboration behavior data, construct a causal relationship diagram between collaboration behavior and task performance, dynamically quantify the intensity of synergy between team members, and track the evolution of capabilities in combination with a temporal attention mechanism;

[0050] The fifth unit is used to generate a team collaboration evaluation report based on the capability evolution trajectory; based on the team collaboration evaluation report, output a team collaboration optimization plan, adjust the parameter configuration of the scenario reproduction training environment in a targeted manner, and update the training strategy of the multi-agent collaborative learning network.

[0051] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0052] processor;

[0053] a memory for storing processor-executable instructions;

[0054] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0055] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0056] The beneficial effects of this application are as follows:

[0057] Through the training method of scene reproduction and multi-agent collaboration, the present invention can accurately simulate real team collaboration scenarios, provide team members with an immersive collaborative skills training experience, greatly improve the authenticity and pertinence of team collaboration training, thereby effectively shortening the team running-in period and improving collaboration efficiency.

[0058] This invention adopts a technical architecture that combines deep reinforcement learning with knowledge distillation, which realizes the continuous optimization of intelligent agent behavior patterns and adaptive adjustment of environmental parameters. It can dynamically adjust the training difficulty and collaboration requirements according to the actual performance of team members, ensuring that the training content is always in the "zone of proximal development" of team members, significantly improving the personalization level of training effects and learning efficiency.

[0059] The counterfactual reasoning analysis framework and temporal attention mechanism constructed by the present invention can accurately quantify the contribution of different collaborative behaviors to task results and the intensity of synergy among team members, forming an objective and detailed team collaboration evaluation report and optimization plan, providing data-driven decision support for team building and collaboration capacity improvement, and effectively solving the problems of vague evaluation standards and unclear optimization directions in traditional team training. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the process of a team collaboration training method based on scene reproduction and multi-agent collaboration according to an embodiment of the present invention;

[0061] Figure 2 A histogram comparing task completion rates of the scenario difficulty adaptation method according to an embodiment of the present invention;

[0062] Figure 3 This is a histogram comparing the multi-objective optimization effects of characteristic components according to an embodiment of the present invention;

[0063] Figure 4 This is a logic diagram of team collaboration behavior analysis and capability evolution tracking in an embodiment of the present invention;

[0064] Figure 5 This is a bar chart comparing the probability of capability transition of different teams according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0066] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0067] Figure 1 FIG is a flow chart of a team collaboration training method based on scene reproduction and multi-agent collaboration according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0068] Establishing a team collaboration knowledge base based on pre-acquired team collaboration historical data, constructing a multi-agent interaction scenario based on the team collaboration knowledge base, mapping virtual environment parameters, agent role information, and task goal information in the team collaboration knowledge base to the multi-agent interaction scenario, and generating a scenario reproduction training environment;

[0069] Generate an intelligent agent in the scenario reproduction training environment, and train the initial behavior pattern of the intelligent agent through a deep reinforcement learning model based on the historical behavior data in the team collaboration knowledge base;

[0070] Constructing a knowledge distillation-enhanced multi-agent collaborative learning network, and feeding back collaborative behavior data generated by team members interacting with the agent in the scenario reproduction training environment to the multi-agent collaborative learning network in real time;

[0071] Based on the collaborative behavior data, the impact of team collaborative behavior on task results is analyzed using counterfactual reasoning methods, and a causal relationship diagram between collaborative behavior and task performance is constructed. The causal relationship diagram dynamically quantifies the strength of the synergy between team members and uses a temporal attention mechanism to track the evolution of capabilities.

[0072] According to the capability evolution trajectory, a team collaboration evaluation report is generated; based on the team collaboration evaluation report, a team collaboration optimization plan is output, the parameter configuration of the scenario reproduction training environment is adjusted in a targeted manner, and the training strategy of the multi-agent collaborative learning network is updated.

[0073] In an optional embodiment, mapping the virtual environment parameters, agent role information, and task objective information in the team collaboration history data to the multi-agent interaction scene to generate a scene reproduction training environment includes:

[0074] Constructing a feature vector space mapping model, and extracting features of virtual environment parameters, agent role information, and task target information based on the feature vector space mapping model to generate a target scene feature vector, wherein the target scene feature vector includes an environment feature component, a role feature component, and a task feature component;

[0075] constructing a multi-objective optimization function based on the target scene feature vector, calculating feature mapping losses for the environment feature component, the role feature component, and the task feature component according to the multi-objective optimization function, and optimizing the target scene feature vector based on a weighted result of the feature mapping losses;

[0076] The optimized target scene feature vector is input into the generator of the generative adversarial network, and the generator reconstructs the target scene feature vector through a multi-layer transposed convolutional network to generate an initial scene structure including environment layout parameters, character attribute parameters, and task configuration parameters;

[0077] Building a scene difficulty adaptive adjustment module based on the initial scene structure, the scene difficulty adaptive adjustment module calculates a scene difficulty coefficient according to a preset performance indicator, and dynamically adjusts the initial scene structure based on the scene difficulty coefficient;

[0078] The adjusted initial scene structure is instantiated as a training environment, and an intelligent agent interaction interface and a task evaluation module are configured in the training environment. The task evaluation module dynamically adjusts the evaluation criteria based on the scene difficulty coefficient to achieve adaptive generation of a scene reproduction training environment.

[0079] A feature vector space mapping model was constructed, employing a bidirectional long short-term memory (LSTM) network structure consisting of an environment encoding layer, a role encoding layer, and a task encoding layer. The environment encoding layer uses a 1024-dimensional hidden state vector to process virtual environment parameters, such as terrain information, climate conditions, and obstacle distribution. The role encoding layer uses a 512-dimensional embedding vector to represent the agent's role attributes, including capabilities, behavior patterns, and decision preferences. The task encoding layer uses a 768-dimensional task representation vector to process task objective information, including task type, completion conditions, and time constraints.

[0080] The model integrates three layers of encoding information through the attention mechanism to generate a 2304-dimensional target scene feature vector, of which the environmental feature component accounts for 1024 dimensions, the role feature component accounts for 512 dimensions, and the task feature component accounts for 768 dimensions.

[0081] Based on the generated target scene feature vector, the system constructs a multi-objective optimization function to optimize the feature mapping. This optimization function simultaneously considers three objectives: environmental consistency, character behavior rationality, and task challenge. The environmental consistency objective is evaluated by calculating the structural similarity between the generated environment and historical environments, using convolution operations to extract environmental topological features. The character behavior rationality objective quantifies the degree of deviation between the generated character behavior patterns and the historical agent's behavior trajectories. The task challenge objective evaluates the match between the task difficulty and historical tasks based on completion rate and average time.

[0082] The feature map losses generated by the calculations for each target were combined using adaptive weights: a weight of 0.4 for the environment feature loss, a weight of 0.35 for the character feature loss, and a weight of 0.25 for the task feature loss. The optimization process used gradient descent for 5000 iterations, with a learning rate set to 0.001 and decayed by 10% every 500 iterations, until the combined loss of the target scene feature vector fell below a preset threshold of 0.05.

[0083] The optimized target scene feature vector is fed into the generator of a generative adversarial network for feature reconstruction. This generator employs a five-layer transposed convolutional network architecture. The first layer receives a 2304-dimensional target scene feature vector and performs feature expansion using 128 5×5 convolution kernels. The second layer uses 64 4×4 transposed convolution kernels for upsampling. The third layer applies 32 3×3 transposed convolution kernels to generate the basic scene outline. The fourth layer uses 16 3×3 transposed convolution kernels to refine the scene details. The fifth layer uses 8 2×2 transposed convolution kernels to generate the final scene layout.

[0084] Each layer of transposed convolution is followed by batch normalization and a LeakyReLU activation function with an activation parameter of 0.2. The generator output includes a 128×128×3 matrix of environment layout parameters, representing terrain height, material type, and obstacle density in three-dimensional space; a 64×64×6 matrix of character attribute parameters, representing the agent's position, orientation, speed, energy value, communication range, and skill cooldown; and a 32×32×4 matrix of task configuration parameters, representing the task trigger point, target area, resource distribution, and time node.

[0085] Based on the generated initial scenario structure, an adaptive scenario difficulty adjustment module was implemented. This module uses reinforcement learning to dynamically adjust scenario parameters to accommodate varying user skill levels. The module first defines the scenario difficulty coefficient, with an initial value of 0.5 (medium difficulty) and a range of 0.1-0.9. The system then calculates adaptive difficulty based on three pre-defined performance metrics: task completion rate (target value 85%), average decision time (target value 1.5 seconds), and collaborative efficiency index (target value 0.7).

[0086] When the user's performance exceeds the target, the difficulty coefficient increases by 0.05 per round; otherwise, it decreases by 0.05. Scene adjustments follow specific rules: When the difficulty coefficient increases, the obstacle density in the environment layout parameters increases by 20%, and the terrain complexity increases by 15%; the energy value in the character attribute parameters decreases by 10%, and the skill cooldown time increases by 25%; the time constraint in the task configuration parameters decreases by 15%, and the target area decreases by 30%. Conversely, these parameters are relaxed accordingly.

[0087] The adjusted scene structure is instantiated into an interactive training environment. The instantiation process uses a 3D engine to construct the physical scene and configures an agent interaction interface consisting of four main interfaces: a state observation interface that returns a 128-dimensional vector of the current environment state; an action execution interface that accepts 12 discrete actions and 3 continuous action parameters; a communication interaction interface that supports 256-bit message transmission between agents; and a reward feedback interface that calculates reward signals in the range of -10 to +10 in real time.

[0088] The task assessment module adjusts the evaluation criteria based on the scenario's difficulty coefficient. When the difficulty coefficient is 0.3, completing the basic objectives qualifies as a pass. When the difficulty coefficient reaches 0.7, a pass is achieved only if all objectives are completed within the time limit and collaboration efficiency reaches 90%. During the evaluation process, team performance data is recorded every 100 milliseconds for subsequent dynamic difficulty adjustments and training effectiveness analysis.

[0089] Figure 2 This is a bar chart comparing task completion rates of the scenario difficulty adaptation method according to an embodiment of the present invention:

[0090] This figure compares the performance of three solutions (this technique, a fixed difficulty solution, and a random difficulty method) in four scenarios of varying difficulty. In simple scenarios, all three solutions achieved high performance, reaching 95.3%, 93.5%, and 91.8%, respectively, with relatively small differences. In medium scenarios, performance began to diverge significantly, with this technique maintaining a high performance of 92.1%, while the fixed difficulty solution dropped to 84.2%, and the random difficulty method further decreased to 79.5%. In complex scenarios, the performance gap widened further, with this technique maintaining a strong performance of 87.4%, while the fixed difficulty solution dropped to 68.7%, and the random difficulty method only 63.2%. In extreme scenarios, the performance difference was most significant, with this technique maintaining a strong performance of 82.9%, despite a slight decline, while the fixed difficulty solution and the random difficulty method dropped significantly to 53.1% and 48.6%, respectively. The data clearly demonstrates that as the scenario difficulty increases, this technique demonstrates significant performance advantages and greater robustness, particularly in high-difficulty scenarios, far surpassing the other two solutions in adaptability.

[0091] In an optional embodiment, constructing a multi-objective optimization function based on the target scene feature vector, and calculating the feature mapping loss for the environment feature component, the role feature component, and the task feature component according to the multi-objective optimization function respectively includes:

[0092] Constructing a feature component representation model, encoding the environment feature component, the role feature component, and the task feature component into feature vectors based on the feature component representation model, and generating an initial feature vector set;

[0093] Constructing a multi-objective mapping loss function based on the initial feature vector set, wherein the multi-objective mapping loss function performs weighted calculation on the environment feature mapping loss, the role feature mapping loss, and the task feature mapping loss to generate a base mapping loss value;

[0094] Establishing a feature causal graph structure using the initial feature vector set, wherein the nodes of the feature causal graph structure are feature vectors in the initial feature vector set, the edges of the feature causal graph structure represent causal relationships between feature vectors, and calculating the causal strength between feature nodes through causal intervention operations;

[0095] The causal strength is integrated with the base mapping loss value to construct a causal enhancement overall loss function; parameter gradients are calculated based on the overall loss function, and parameters of the feature component representation model are updated based on the parameter gradients;

[0096] Based on the updated feature component representation model, corresponding weight coefficients are calculated for the environment feature mapping loss, the role feature mapping loss, and the task feature mapping loss to obtain the feature mapping loss of each feature component.

[0097] A feature component representation model is constructed, which consists of three encoder modules: an environmental feature encoder, a role feature encoder, and a task feature encoder. The environmental feature encoder receives input environmental features, such as weather conditions, lighting conditions, and terrain information, and converts them into a 64-dimensional environmental feature vector through a multi-layer neural network.

[0098] The role feature encoder receives input role features, such as role type, behavior patterns, and interaction preferences, and converts them into a 32-dimensional role feature vector using a convolutional neural network and a fully connected layer. The task feature encoder receives input task features, such as task difficulty, completion time limit, and resource requirements, and converts them into a 48-dimensional task feature vector using a recurrent neural network. These three feature vectors together constitute the initial feature vector set.

[0099] For example, for a specific scenario, environmental features include "indoor space, ample lighting, low noise level," role features include "adult user, medium technical proficiency, focused attention," and task features include "medium difficulty, 30 minutes to complete, medium accuracy required." These features are converted into corresponding feature vectors through their respective encoders.

[0100] Based on the initial set of feature vectors, a multi-objective mapping loss function is constructed. This function performs a weighted calculation on the three feature mapping losses to generate a base mapping loss value. The environment feature mapping loss is calculated as the Euclidean distance between the environment feature vector and the target scene feature vector in the environment dimension; the character feature mapping loss is calculated as the negative of the cosine similarity between the character feature vector and the target scene feature vector in the character dimension; and the task feature mapping loss is calculated as the cross entropy between the task feature vector and the target scene feature vector in the task dimension. These three losses are combined using weight coefficients, with initial weights set to 0.35, 0.35, and 0.3, respectively, to obtain the base mapping loss value.

[0101] A feature causal graph structure is established using the initial set of feature vectors. In this graph structure, nodes are feature vectors from the initial set of feature vectors, and edges represent causal relationships between feature vectors. Causal relationships are determined by analyzing the mutual information and conditional independence between feature vectors. Specifically, a sliding window method is used to collect historical samples of feature vectors, the mutual information matrix between samples is calculated, and the PC algorithm is used to determine the direction and existence of edges. For a simple scenario, the possible causal graphs may include causal links such as "environmental features → role features", "environmental features → task features", and "role features → task features".

[0102] Causal intervention operations are used to calculate the causal strength between feature nodes. Causal intervention operations are performed by changing the feature value of a source node and observing the magnitude of the change in the target node. For example, if the environmental feature "sufficient light" is changed to "insufficient light," the change in the character's feature vector can be observed. If the change is significant, it indicates that the environmental feature has a strong causal influence on the character's feature. Through multiple intervention experiments, a causal strength matrix is ​​calculated, where each element represents the causal strength from one feature to another.

[0103] The causal strength is integrated with the base mapping loss to construct a causally enhanced overall loss function. This integration is achieved by adding the causal strength as a regularizer to the base mapping loss. If the causal strength between two features is high, a corresponding penalty term is added to the loss function, allowing the model to learn the causal relationship between the features. For example, if the causal strength of the environment feature to the task feature is 0.7 (strong), then an environment-task feature mapping loss term with a weight of 0.7 is added to the overall loss function.

[0104] Parameter gradients are calculated based on the overall loss function, and gradient descent is used to update the parameters of the feature component representation model. In each training batch, a set of environment-role-task feature samples is collected. The overall loss is calculated via forward propagation, and then the gradients are calculated and the model parameters are updated via backpropagation. The learning rate is set to 0.001, and the Adam optimizer is used for parameter updates. The training batch size is 64, and the number of training epochs is 100.

[0105] Based on the updated feature component representation model, the corresponding weight coefficients for the three feature mapping losses are calculated. The weight coefficients are calculated based on the contribution of each feature to the overall loss, which is determined through sensitivity analysis. Each feature is slightly perturbed and the magnitude of the change in the overall loss is observed. The larger the magnitude of the change, the greater the contribution of the feature to the overall loss, and the larger the corresponding weight coefficient. After multiple iterative calculations, the final weight coefficient for the environment feature mapping loss is 0.4, the weight coefficient for the role feature mapping loss is 0.35, and the weight coefficient for the task feature mapping loss is 0.25. These weight coefficients reflect the importance of different feature components in the current scene.

[0106] Through the above method, a technical solution is realized to construct a multi-objective optimization function based on the target scene feature vector, and calculate the feature mapping loss for the environment, role and task feature components respectively according to the optimization function, providing an effective solution for the target representation of scene adaptation.

[0107] Figure 3 This is a histogram comparing the multi-objective optimization effects of characteristic components according to an embodiment of the present invention:

[0108] The figure compares the performance of three different models (the basic feature representation model, the multi-objective mapping model, and the causal enhancement mapping model) in five different scenarios. In complex environments, the three models achieved performance levels of 73.5%, 81.3%, and 88.7%, respectively, with the causal enhancement mapping model performing the best. In multi-role collaboration, the performance levels were 68.2%, 77.6%, and 85.9%, respectively, maintaining similar performance gaps. In dynamic task scenarios, the three models achieved performance levels of 71.8%, 79.5%, and 87.2%, respectively, demonstrating the adaptability of the causal enhancement mapping model in dynamic environments. In cross-domain fusion scenarios, the performance levels were 65.4%, 74.8%, and 82.6%, respectively. Although overall performance declined, the relative advantages between the models remained significant. In high-stress scenarios, the performance dropped to 62.7%, 72.3%, and 80.4%, respectively, demonstrating the impact of stressful environments on model performance. Across all test scenarios, the causal enhancement mapping model consistently maintained a significant performance advantage, averaging about 15-20 percentage points higher than the basic feature representation model and about 8-10 percentage points higher than the multi-target mapping model, demonstrating the model's robustness and generalization capabilities in complex and changing environments.

[0109] In an optional embodiment, generating an agent in the scenario reproduction training environment and training the initial behavior pattern of the agent using a deep reinforcement learning model based on historical behavior data in the team collaboration knowledge base includes:

[0110] Generate an agent in the scenario reproduction training environment, extract historical behavior data from a team collaboration knowledge base, and build a deep reinforcement learning model based on the historical behavior data, wherein the deep reinforcement learning model includes a behavior state space, an action space, and a reward function;

[0111] The agent is iteratively trained using the deep reinforcement learning model, the policy network parameters of the agent are updated according to the feedback information of the reward function, and the initial behavior pattern of the agent is trained.

[0112] A scenario-reproduction training environment is constructed based on actual application scenarios. This environment includes a multi-agent interaction area, environmental state information, and a task objective setting module. N agents are generated in the environment, each with perception, decision-making, and execution capabilities. The agent's perception capability includes observing the environmental state and obtaining information such as the current position, the distribution of surrounding objects, and the positions of other agents. The decision-making capability is composed of a deep neural network, which receives perception information and outputs action instructions. The execution capability converts decisions into actual actions in the environment. For example, in a warehousing and logistics scenario, 10 transport robots can be generated as agents. Each agent is equipped with a position sensor, lidar, and a communication module, and can perceive obstacles, other agents, and target goods in the environment.

[0113] The system accesses a pre-built team collaboration knowledge base, which contains historical behavioral data on various agents during task execution, including location trajectories, action selections, and task assignments. In a warehousing and logistics scenario, the knowledge base contains behavioral records of 5,000 collaborative tasks completed by handling robots over the past three months, including 500,000 trajectory data points. Each data point records information such as the agent ID, timestamp, location coordinates, current task, battery status, and load.

[0114] The system extracts representative historical behavior data from the knowledge base, including both routine task execution data and special case handling data. The system uses a data cleaning algorithm to remove outliers and redundant data, retaining valid training samples. In this example, the system selects 4,000 task data instances as valid samples, eliminating abnormal behavior records and duplicate records caused by equipment failures.

[0115] A deep reinforcement learning model is constructed based on filtered historical behavior data. This model defines three core elements: a behavior state space, an action space, and a reward function. The behavior state space is the collection of environmental information perceivable by the agent, including the agent's own state (position, orientation, velocity, etc.), the state of the environment (obstacle distribution, target location, etc.), and the state of other agents (position, task status, etc.). In this example, the state space has 100 dimensions and includes the robot's current two-dimensional coordinates (x, y), orientation angle, linear and angular velocity, battery percentage, obstacle distances in eight directions within a 10-meter radius, the current weight of the cargo being carried, the coordinates of the target shelf, and the positions and states of five other nearby robots.

[0116] An action space defines all the behaviors an agent can perform. It can be categorized into two types: discrete and continuous. Discrete action spaces are suitable for scenarios with a limited number of behavioral choices, while continuous action spaces are suitable for scenarios requiring precise control. In this example, the robot's action space includes eight discrete movement directions (forward, backward, left, right, and four diagonal directions) and three manipulation actions (grab, drop, and wait), for a total of 11 basic actions.

[0117] Reward function design is crucial for reinforcement learning, used to evaluate the behavior of intelligent agents and guide the learning process. The reward function comprehensively considers factors such as task completion, resource utilization efficiency, collaborative effectiveness, and safety. In this example, the reward function is set as follows: successfully delivering cargo to the target location receives a reward of +100; each time step consumes a reward of -0.1 to encourage rapid task completion; colliding with other robots receives a reward of -50; colliding with static obstacles receives a reward of -30; maintaining an appropriate distance (0.5-1.5 meters) from other robots receives a reward of +0.5 to promote collaboration; and choosing the shortest path receives a reward of +0.2 per step.

[0118] Based on the above definitions, a deep reinforcement learning network architecture was constructed, consisting of a policy network and a value network. The policy network takes the current state as input and outputs an action probability distribution; the value network takes the current state as input and outputs an estimate of the state's value. The policy network uses a four-layer fully connected neural network structure. The input layer has 100 neurons, the same as the state space dimension. The two hidden layers contain 128 and 64 neurons, respectively. The output layer has 11 neurons, the same as the action space dimension. ReLU activation function and a softmax output layer are used. The value network uses a three-layer fully connected neural network structure. The input layer also has 100 neurons, the hidden layer contains 64 neurons, and the output layer has one neuron representing the state value.

[0119] The agent is iteratively trained using the constructed deep reinforcement learning model. The training process uses a batch training approach, with each batch containing 256 state-action-reward-new-state transition samples. In this example, the system sets a total of 5,000 training rounds, each consisting of 1,000 steps of interaction with the environment. The agent performs actions in the environment, and the environment provides feedback on the new state and reward. The agent records these transitions in an experience pool. When the experience pool reaches 10,000 data points, the system begins randomly sampling batches to update the network parameters.

[0120] During the parameter update process, the system calculates policy loss and value loss. Policy loss measures the difference between the current policy and the optimal policy, while value loss measures the difference between the predicted state value and the actual value. The optimization algorithm uses the Adam optimizer, with a learning rate of 0.0003, a momentum parameter of 0.9, and a learning rate decay of 10% every 50 rounds. In this example, the loss function values ​​of the policy network and value network were 4.2 and 8.5, respectively, at the beginning of training. After 5000 rounds of training, they dropped to 0.7 and 1.3, indicating that the network parameters are gradually converging.

[0121] During training, the system evaluates the performance of the current strategy every 100 rounds, using metrics including average task completion time, success rate, number of collisions, and collaboration efficiency. In a sample scenario, the initial strategy achieved an average task completion time of 120 seconds, a success rate of 65%, and an average number of collisions per task of 3.2. After 5,000 rounds of training, the average task completion time dropped to 72 seconds, the success rate increased to 93%, and the average number of collisions decreased to 0.5.

[0122] The trained agents ultimately possessed initial behavioral patterns and were capable of making collaborative decisions based on sensory information. In a real-world warehouse environment, 10 agents were able to collaboratively complete a multi-target cargo handling task, autonomously avoiding obstacles and optimizing routes. This resulted in a 62% improvement in task efficiency compared to manual operations and a 25% improvement compared to traditional algorithms. These initial behavioral patterns will serve as the foundation for subsequent personalized training to further adapt to the specific needs of different scenarios.

[0123] In an optional embodiment, the impact of team collaboration on task results is analyzed using a counterfactual reasoning method based on the collaboration behavior data, and a causal relationship diagram between collaboration behavior and task performance is constructed. The causal relationship diagram dynamically quantifies the intensity of synergy between team members. The ability evolution trajectory is tracked in conjunction with a temporal attention mechanism, including:

[0124] Constructing a collaborative behavior vector and a task result vector based on the collaborative behavior data, calculating the causal effect of each behavioral feature in the collaborative behavior vector on each result indicator in the task result vector using a causal interference model, wherein the causal effect represents the intensity of the impact of the collaborative behavior on the task result, and constructing a causal relationship diagram between the team collaborative behavior and the task performance based on the causal effect;

[0125] Calculating the synergy strength among team members based on the causal relationship graph, combining the timing weight and the attention weight to weight the causal effect, and obtaining the dynamic synergy strength;

[0126] Constructing a team capability phase space based on the dynamic synergy effect strength, calculating a capability index based on the team capability phase space, wherein the capability index is used to characterize the stability of the team capability evolution, and inputting the capability index into a multi-scale temporal attention mechanism;

[0127] The multi-scale temporal attention mechanism calculates the dynamic association weights of temporal features using the query matrix, the key matrix, and the value matrix, and constructs a Jacobian matrix based on the dynamic association weights, wherein the Jacobian matrix describes the local linear features of the capability phase space;

[0128] The capability index and the Jacobian matrix are combined to calculate the bifurcation discriminant, and the transition probability of the team capability is predicted based on the bifurcation discriminant, where the transition probability indicates the turning point of the capability evolution; the dynamic association weight is combined with the transition probability to track and predict the evolution trajectory of the team capability.

[0129] like Figure 4 As shown, the method includes:

[0130] Collect data on the collaborative behaviors of team members during task execution. This data can include metrics such as communication frequency, response time, task allocation, and decision-making participation. Data collection can be achieved through collaboration platform logging, sensor monitoring, or behavioral annotation during specific tasks. For example, during the development of a five-person team, record data points such as the number of code submissions, number of review comments, and contribution to problem solving for each member.

[0131] The collected collaborative behavior data is converted into a structured collaborative behavior vector. Simultaneously, evaluation indicators for task completion quality, efficiency, and innovation are constructed into a task outcome vector. For example, a collaborative behavior vector might include dimensions like [communication frequency: 0.8, resource sharing: 0.6, decision-making participation: 0.7], while a task outcome vector might include dimensions like [completion quality: 0.85, time efficiency: 0.75, innovation index: 0.6]. The system uses causal intervention models to calculate the impact of each behavioral characteristic on the outcome indicators. In practice, a counterfactual intervention is performed on a specific collaborative behavior characteristic, such as changing "communication frequency" from 0.8 to 0.4. The magnitude of change in each indicator in the task outcome vector is observed, and this process is repeated to construct a causal effect matrix. Calculations show that the causal effect of communication frequency on completion quality is 0.35, and on time efficiency is 0.28, with higher values ​​indicating more significant impacts.

[0132] Based on the causal effect matrix, a causal relationship diagram is constructed between team collaboration behavior and task performance. Nodes in the diagram represent collaborative behavior characteristics or task outcome indicators, and edge weights represent the magnitude of the causal effect. In practical applications, this causal relationship diagram can be visualized to show, for example, that the impact of "resource sharing" on "completion quality" is 0.42, and its impact on "innovation index" is 0.56, helping managers intuitively understand the strength of the correlation between behavior and performance.

[0133] Calculate the strength of collaboration between team members. First, identify the collaborative interaction patterns between members, such as the frequency of collaboration between members A and B, their complementarity, and their synergy. Combine the importance weights of the time series data with the attention weights of specific behaviors to dynamically weight the aforementioned causal effects. For example, in the early stages of a project, the weight of communication and collaboration might be 0.65, while in the later stages of the project, the weight of execution might increase to 0.72. The calculation results show that over time, the synergy between team members A and B in problem solving has increased from 0.56 to 0.78, indicating that their collaborative model is becoming increasingly mature.

[0134] Based on the strength of dynamic synergy effects, a team capability phase space is constructed. This space describes the changes in the team's state across multiple dimensions of capability indicators. Within this phase space, a capability index is calculated to characterize the stability of the team's capability evolution. In specific implementation, key capability dimensions such as problem-solving, innovation, and adaptability are selected to construct the phase space, tracking the dynamic changes in the team's state along these dimensions. Calculations show that one team's capability index is 0.12, indicating relatively stable capability evolution; while another team's index reaches 0.45, indicating significant capability fluctuations and potential need for management intervention.

[0135] The calculated capability index is fed into a multi-scale temporal attention mechanism. This mechanism computes the dynamic association weights of temporal features using a query matrix, a key matrix, and a value matrix. In practice, the query matrix represents the current capability state of interest, the key matrix stores historical capability state features, and the value matrix contains the corresponding capability performance data. Through the interaction of these three matrices, the association weights of capability features at different time points are derived. For example, the impact of recent collaboration patterns on current innovation capabilities is weighted at 0.63, while the impact of early training effectiveness on current problem-solving capabilities is weighted at 0.24.

[0136] Based on dynamic correlation weights, a Jacobian matrix is ​​constructed to describe the mutual influence between dimensions in the capability phase space. The matrix elements represent the immediate impact of changes in one capability dimension on another. Practical applications show that teamwork ability has an impact coefficient of 0.58 on problem-solving efficiency, while information sharing level has an impact coefficient of 0.71 on innovation capability.

[0137] The capability index and Jacobian matrix are combined to calculate the bifurcation discriminant, which is used to detect conditions that indicate a potential qualitative change in capability status. When the discriminant exceeds a preset threshold (e.g., 0.85), it indicates a potential shift in team capability. The discriminant is used to calculate the probability of a team capability transition. For example, after a specific team training session, the system calculates a 0.72 probability of improving team collaboration in the next phase, indicating a high likelihood of positive results from the training.

[0138] By combining dynamic correlation weights with transition probabilities, a predictive model for team capability evolution is constructed. This model can predict the trajectory of team capability development over a period of time and identify potential turning points. For example, a forecast indicates that a team's innovation capability will steadily improve by approximately 15% over the next quarter, but a 30% chance of a capability transition may occur when a cross-departmental project is launched. These predictions provide team managers with precise guidance on the timing and direction of intervention.

[0139] In an optional embodiment, the capability index and the Jacobian matrix are combined to calculate a bifurcation discriminant, and predicting the transition probability of team capability based on the bifurcation discriminant includes:

[0140] Combining the capability index with the eigenvalue of the Jacobian matrix to construct a bifurcation parameter vector, wherein the bifurcation parameter vector includes the capability index and the maximum real part of the eigenvalue;

[0141] A bifurcation discriminant is calculated based on the bifurcation parameter vector through the determinant of the Jacobian matrix and the exponential smoothing term, and the bifurcation discriminant is used to detect the mutation characteristics of the team capability status;

[0142] Calculating a critical value of the bifurcation discriminant within an observation time window, the critical value representing a threshold condition for a transition in the team capability state, and inputting a difference between the bifurcation discriminant and the critical value into a nonlinear mapping function;

[0143] The nonlinear mapping function is constructed based on the sigmoid function to calculate the transition probability of the team's ability, and the transition probability represents the possibility of a sudden change in the team's ability;

[0144] A confidence interval of the transition probability is constructed, and the upper and lower bounds of the confidence interval are calculated by an uncertainty function. The uncertainty function is dynamically adjusted as the bifurcation discriminant changes, thereby achieving an accurate prediction of the team capability transition.

[0145] Time-series data from team members' collaboration processes is captured, and various team capability indicators are represented using system state vectors. For a five-member R&D team, 30 consecutive days of collaboration data were collected, including metrics such as communication frequency, task completion rate, knowledge contribution, problem-solving time, and number of innovative proposals. Using adaptive sliding window technology, the system calculates the capability index and Jacobian matrix for each time point.

[0146] The capability index is calculated using a phase space reconstruction method, with an embedding dimension of 4 and a time delay of 2. The nearest neighbor tracking algorithm is used to analyze the dynamic time series of teamwork. For example, the capability index for the task completion rate indicator is calculated to be 0.127, indicating a certain degree of instability.

[0147] The Jacobian matrix calculation is based on a dynamic model of team capabilities, which includes the coupled relationships between five key variables. Numerical differentiation methods are used to evaluate the impact of state variables on the system's evolution at each time point. For example, in the observations on day 15, the calculated Jacobian matrix quantitatively describes the interaction effects between team members. The coefficient of influence of knowledge contribution on problem-solving time is -0.42, indicating that knowledge sharing can significantly reduce problem-solving time.

[0148] To construct the bifurcation parameter vector, the system combines the capability index with the eigenvalues ​​of the Jacobian matrix. Specifically, the five eigenvalues ​​of the Jacobian matrix are extracted, and the eigenvalue with the largest real part is selected and combined with the capability index to form a two-dimensional parameter vector. For example, in the data for day 20, the eigenvalue with the largest real part is 0.086, which, combined with the capability index of 0.127, forms the bifurcation parameter vector [0.127, 0.086].

[0149] The calculation of the bifurcation discriminant incorporates the determinant of the Jacobian matrix and an exponential smoothing term. The system first calculates the determinant of the Jacobian matrix, obtaining a value of -0.0032. Then, a smoothing factor of 0.85 is introduced, resulting in an exponential smoothing term of 0.0241. The weighted combination of these two terms yields the bifurcation discriminant of 0.0209, which is used to determine whether the team's capabilities are approaching the bifurcation point. The discriminant is dynamically updated within a 10-day observation window, forming a time series of the discriminant.

[0150] The critical value was determined through statistical analysis of historical data. The system analyzed team capability changes over the past six months, identifying 20 capability transitions and extracting the discriminant values ​​before each transition. Percentile analysis determined the critical value to be 0.025, indicating that when the discriminant exceeds this value, there is a high probability of a team capability transition.

[0151] The transition probability was calculated using a nonlinear mapping function based on the sigmoid function, with a steepness parameter of 15 and an offset parameter of 0.5. Specifically, when the bifurcation discriminant was 0.0209 and the critical value was 0.025, the difference between the two was -0.0041. After inputting the nonlinear mapping function, the resulting transition probability was 0.437, indicating a 43.7% probability that the team's capabilities would undergo a significant change in the near future.

[0152] Confidence intervals for the transition probability are also constructed and dynamically adjusted using an uncertainty function. This uncertainty function accounts for data noise, model parameters, and external interference factors, providing narrower confidence intervals when the discriminant is close to the critical value and wider intervals when it is far from the critical value. For example, when the transition probability is 0.437, the upper and lower bounds of the 95% confidence interval are 0.512 and 0.362, respectively.

[0153] The system continuously monitors changes in indicators and triggers an early warning mechanism when the probability of a transition exceeds 0.6 for three consecutive days. For example, mid-project, the system detected a drop in communication frequency and fluctuations in task completion rates, leading to a transition probability of 0.78. This successfully predicted the subsequent decline in team collaboration efficiency.

[0154] The advantage of this method is its ability to proactively detect potential sudden changes in team capabilities, providing managers with a window of opportunity for intervention. Experimental validation demonstrated that across 30 test cases, this method achieved 84% prediction accuracy, with an average lead time of 3.7 days. This represents a 26% improvement in accuracy and a 1.5-day improvement in warning time compared to traditional statistical methods.

[0155] Figure 5 This is a bar chart comparing the probability of capability transition of different teams according to an embodiment of the present invention:

[0156] This figure compares the performance of three different methods (basic classification, Lyapunov enhancement, and dynamic confidence interval) across four test scenarios. In the standard scenario test, the three methods achieved performance of 67.8%, 78.4%, and 83.6%, respectively, with the dynamic confidence interval method performing the best. In the high-stress environment test, the performance decreased to 59.3%, 72.6%, and 79.2%, respectively, while maintaining a relatively strong performance. In the team collaboration test, the three methods achieved high performance levels of 74.2%, 85.1%, and 88.7%, respectively, demonstrating their adaptability in collaborative scenarios. In the long-term project test, the performance reached 62.5%, 76.9%, and 84.3%, respectively, demonstrating the more stable performance of the dynamic confidence interval method in sustained tasks. Overall, the dynamic confidence interval method maintained a significant performance advantage across all test scenarios, averaging approximately 20 percentage points ahead of the basic classification method and 5-8 percentage points ahead of the Lyapunov enhancement method, demonstrating its robustness and adaptability across diverse application scenarios.

[0157] In an optional embodiment, generating a team collaboration evaluation report based on the capability evolution trajectory; outputting a team collaboration optimization plan based on the team collaboration evaluation report, adjusting the parameter configuration of the scenario reproduction training environment in a targeted manner, and updating the training strategy of the multi-agent collaborative learning network include:

[0158] The capability evolution trajectory includes multidimensional capability indicators. Based on these multidimensional capability indicators, the individual capability growth rate and team collaboration efficiency are calculated. The individual capability growth rate represents the speed of capability improvement, and the team collaboration efficiency represents the collaboration effect. The individual capability growth rate and the team collaboration efficiency are weighted and combined to obtain a comprehensive team collaboration score, which reflects the overall performance of the team.

[0159] A multidimensional evaluation system is constructed based on the comprehensive team collaboration score, the multidimensional evaluation system including individual dimensions, team dimensions, and task dimensions, and a team collaboration evaluation report is generated; an optimization objective function is constructed based on the team collaboration evaluation report, the optimization objective function combines performance loss and optimization cost, and generates a team collaboration optimization plan based on resource constraints, time constraints, and effect constraints;

[0160] The team collaboration optimization plan is converted into the parameter configuration of the scenario reproduction training environment, where the parameter configuration includes a training difficulty parameter and a scenario complexity parameter. The scenario reproduction training environment is dynamically adjusted, and the training strategy of the multi-agent collaborative learning network is updated based on the parameter configuration.

[0161] By analyzing the trajectory of capability evolution, a team collaboration evaluation report is generated, and based on this, a team collaboration optimization plan is output. The parameter configuration of the scenario reproduction training environment is adjusted in a targeted manner, and the training strategy of the multi-agent collaborative learning network is updated.

[0162] The capability evolution trajectory includes multidimensional capability indicators covering five dimensions: communication, decision-making efficiency, resource allocation, problem-solving, and adaptability. Each dimension is scored on a standardized scale of 0-100, generated by collecting performance data during the agent training process. For example, the capability indicators of three agents A, B, and C in a team at the beginning of training were: A(65, 70, 60, 75, 80), B(80, 65, 75, 60, 70), and C(70, 75, 65, 80, 60). After 50 rounds of training, they evolved to: A(80, 85, 75, 85, 90), B(90, 80, 85, 75, 85), and C(85, 90, 80, 90, 75).

[0163] Individual capability growth rates and team collaboration efficiency are calculated based on multidimensional capability indicators. Individual capability growth rates are calculated by dividing the difference between the current capability value and the baseline capability value by the number of training rounds. Taking Agent A as an example, its communication capability growth rate is (80-65) / 50 = 0.3. Team collaboration efficiency is calculated as the weighted average of three indicators: team task completion, collaborative conflict resolution rate, and information sharing efficiency. In practice, the weight for task completion can be set to 0.5, the weight for collaborative conflict resolution rate to 0.3, and the weight for information sharing efficiency to 0.2. Assuming that a team's three indicators at a specific training stage are 85%, 90%, and 80%, respectively, the team collaboration efficiency is 85% × 0.5 + 90% × 0.3 + 80% × 0.2 = 85.5%.

[0164] The weighted combination of individual capability growth rate and team collaboration efficiency yields a comprehensive team collaboration score. The weighting coefficients can be adjusted based on the specific application scenario. A typical configuration is a weight of 0.4 for individual capability growth rate and 0.6 for team collaboration efficiency. For example, if a team's average individual capability growth rate is 0.28 and its team collaboration efficiency is 85.5%, the overall team collaboration score is 0.28 × 0.4 + 0.855 × 0.6 = 0.625. This score reflects the overall performance of the team and ranges from 0 to 1, with closer to 1 indicating better team collaboration.

[0165] A multi-dimensional evaluation system is constructed based on the comprehensive team collaboration score, which includes individual, team, and task dimensions. The individual dimension focuses on the development of each intelligent agent's capabilities, recording the growth curve and bottleneck points of each dimension; the team dimension analyzes the team structure, role allocation, and interaction mode; and the task dimension evaluates the performance differences of the team in tasks of different difficulty levels. The evaluation system associates the indicators of each dimension with the comprehensive team collaboration score through a weight matrix to generate a team collaboration evaluation report. The report includes four parts: a radar chart of the team's capability distribution, an analysis of the collaboration model, an analysis of strengths and weaknesses, and a forecast of development trends. For example, the evaluation results of a team showed that the communication efficiency was high (90 points) but the resource allocation ability was weak (65 points). The collaboration model analysis found that the quality of collaboration of the team decreased by 30% under high-pressure environment.

[0166] An optimization objective function is constructed based on the team collaboration evaluation report. This function combines performance loss and optimization cost, and generates a team collaboration optimization plan based on resource constraints, time constraints, and effect constraints. Performance loss is quantified by the gap between current performance and ideal performance. The optimization cost includes computing resource consumption, training time, and model complexity. The constraints are set as follows: computing resources do not exceed 120% of the budget, training time does not exceed 72 hours, and performance improvement is not less than 15%. The optimization plan includes three aspects: capability improvement strategy, collaboration mechanism optimization, and task adaptation adjustment. Improvement measures are proposed for the deficiencies found in the evaluation report. For example, to address the problem of weak resource allocation capabilities, the proportion of resource competition training scenarios can be increased from 20% to 35%, and a resource sharing reward mechanism can be introduced, with the reward coefficient set at 1.5 times the basic reward.

[0167] The team collaboration optimization plan is converted into the parameter configuration of the scenario reproduction training environment, including the training difficulty parameter and the scenario complexity parameter. The training difficulty parameter controls the challenge of the task, covering three aspects: time pressure, resource constraints, and interference factors; the scenario complexity parameter defines the frequency of environmental changes, information uncertainty, and decision constraints. For example, to meet the needs of improving decision-making efficiency, the time pressure parameter can be adjusted from 0.6 to 0.8, while reducing the information uncertainty parameter from 0.5 to 0.3, in order to train the agent's ability to make more accurate decisions within a limited time. In addition, the training strategy of the multi-agent collaborative learning network is updated based on the parameter configuration, including adjusting the learning rate decay curve, the reward function weight, and the experience replay sampling strategy. For communication ability training, the relevant reward weight can be increased from 0.2 to 0.35, and a communication failure penalty mechanism can be added, with the penalty coefficient set to 0.15.

[0168] By dynamically adjusting the scene-reproducing training environment and updating the training strategy, the collaborative capabilities of the agent team were continuously optimized. In a collaborative navigation task involving five agents, applying this method improved the overall team collaboration score from 0.625 to 0.782, shortened task completion time by 21.5%, and increased resource utilization by 18.3%, significantly validating the effectiveness of this method.

[0169] A second aspect of an embodiment of the present invention provides a team collaboration training system based on scene reproduction and multi-agent collaboration, including:

[0170] The first unit is configured to establish a team collaboration knowledge base based on pre-acquired team collaboration historical data, construct a multi-agent interaction scenario based on the team collaboration knowledge base, map virtual environment parameters, agent role information, and task objective information in the team collaboration knowledge base to the multi-agent interaction scenario, and generate a scenario reproduction training environment;

[0171] The second unit is configured to generate an intelligent agent in the scenario reproduction training environment, and train an initial behavior pattern of the intelligent agent through a deep reinforcement learning model based on the historical behavior data in the team collaboration knowledge base;

[0172] The third unit is used to build a multi-agent collaborative learning network enhanced by knowledge distillation, and to feed back the collaborative behavior data generated by the team members interacting with the agents in the scenario reproduction training environment to the multi-agent collaborative learning network in real time;

[0173] The fourth unit is used to analyze the impact of team collaboration behavior on task outcomes using counterfactual reasoning based on the collaboration behavior data, construct a causal relationship diagram between collaboration behavior and task performance, dynamically quantify the intensity of synergy between team members, and track the evolution of capabilities in combination with a temporal attention mechanism;

[0174] The fifth unit is used to generate a team collaboration evaluation report based on the capability evolution trajectory; based on the team collaboration evaluation report, output a team collaboration optimization plan, adjust the parameter configuration of the scenario reproduction training environment in a targeted manner, and update the training strategy of the multi-agent collaborative learning network.

[0175] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0176] processor;

[0177] a memory for storing processor-executable instructions;

[0178] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0179] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0180] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A team collaboration training method based on scenario reproduction and multi-agent collaboration, characterized by: include: Establishing a team collaboration knowledge base based on pre-acquired team collaboration historical data, constructing a multi-agent interaction scenario based on the team collaboration knowledge base, mapping virtual environment parameters, agent role information, and task goal information in the team collaboration knowledge base to the multi-agent interaction scenario, and generating a scenario reproduction training environment; Generate an intelligent agent in the scenario reproduction training environment, and train the initial behavior pattern of the intelligent agent through a deep reinforcement learning model based on the historical behavior data in the team collaboration knowledge base; Constructing a knowledge distillation-enhanced multi-agent collaborative learning network, and feeding back collaborative behavior data generated by team members interacting with the agent in the scenario reproduction training environment to the multi-agent collaborative learning network in real time; Based on the collaborative behavior data, the impact of team collaborative behavior on task results is analyzed using counterfactual reasoning methods. A causal relationship diagram between collaborative behavior and task performance is constructed. The causal relationship diagram dynamically quantifies the intensity of the synergy between team members and combines the temporal attention mechanism to track the evolution of capabilities, including: Constructing a collaborative behavior vector and a task result vector based on the collaborative behavior data, calculating the causal effect of each behavioral feature in the collaborative behavior vector on each result indicator in the task result vector using a causal interference model, wherein the causal effect represents the intensity of the impact of the collaborative behavior on the task result, and constructing a causal relationship diagram between the team collaborative behavior and the task performance based on the causal effect; Calculating the synergy strength among team members based on the causal relationship graph, combining the timing weight and the attention weight to weight the causal effect, and obtaining the dynamic synergy strength; Constructing a team capability phase space based on the dynamic synergy effect strength, calculating a capability index based on the team capability phase space, wherein the capability index is used to characterize the stability of the team capability evolution, and inputting the capability index into a multi-scale temporal attention mechanism; The multi-scale temporal attention mechanism calculates the dynamic association weights of temporal features using the query matrix, the key matrix, and the value matrix, and constructs a Jacobian matrix based on the dynamic association weights, wherein the Jacobian matrix describes the local linear features of the capability phase space; Combining the capability index with the Jacobian matrix to calculate a bifurcation discriminant, and predicting the transition probability of team capability based on the bifurcation discriminant, wherein the transition probability indicates a turning point in capability evolution; combining the dynamic association weight with the transition probability to track and predict the trajectory of team capability evolution; According to the capability evolution trajectory, a team collaboration evaluation report is generated; based on the team collaboration evaluation report, a team collaboration optimization plan is output, the parameter configuration of the scenario reproduction training environment is adjusted in a targeted manner, and the training strategy of the multi-agent collaborative learning network is updated.

2. The method according to claim 1, characterized in that Mapping the virtual environment parameters, agent role information, and task goal information in the team collaboration history data to the multi-agent interaction scene to generate a scene reproduction training environment includes: Constructing a feature vector space mapping model, and extracting features of virtual environment parameters, agent role information, and task target information based on the feature vector space mapping model to generate a target scene feature vector, wherein the target scene feature vector includes an environment feature component, a role feature component, and a task feature component; constructing a multi-objective optimization function based on the target scene feature vector, calculating feature mapping losses for the environment feature component, the role feature component, and the task feature component according to the multi-objective optimization function, and optimizing the target scene feature vector based on a weighted result of the feature mapping losses; The optimized target scene feature vector is input into the generator of the generative adversarial network, and the generator reconstructs the target scene feature vector through a multi-layer transposed convolutional network to generate an initial scene structure including environment layout parameters, character attribute parameters, and task configuration parameters; Building a scene difficulty adaptive adjustment module based on the initial scene structure, the scene difficulty adaptive adjustment module calculates a scene difficulty coefficient according to a preset performance indicator, and dynamically adjusts the initial scene structure based on the scene difficulty coefficient; The adjusted initial scene structure is instantiated as a training environment, and an intelligent agent interaction interface and a task evaluation module are configured in the training environment. The task evaluation module dynamically adjusts the evaluation criteria based on the scene difficulty coefficient to achieve adaptive generation of a scene reproduction training environment.

3. The method according to claim 2, characterized in that Constructing a multi-objective optimization function based on the target scene feature vector, and calculating feature mapping losses for the environment feature component, the role feature component, and the task feature component according to the multi-objective optimization function respectively include: Constructing a feature component representation model, encoding the environment feature component, the role feature component, and the task feature component into feature vectors based on the feature component representation model, and generating an initial feature vector set; Constructing a multi-objective mapping loss function based on the initial feature vector set, wherein the multi-objective mapping loss function performs weighted calculation on the environment feature mapping loss, the role feature mapping loss, and the task feature mapping loss to generate a base mapping loss value; Establishing a feature causal graph structure using the initial feature vector set, wherein the nodes of the feature causal graph structure are feature vectors in the initial feature vector set, the edges of the feature causal graph structure represent causal relationships between feature vectors, and calculating the causal strength between feature nodes through causal intervention operations; The causal strength is integrated with the base mapping loss value to construct a causal enhancement overall loss function; parameter gradients are calculated based on the overall loss function, and parameters of the feature component representation model are updated based on the parameter gradients; Based on the updated feature component representation model, corresponding weight coefficients are calculated for the environment feature mapping loss, the role feature mapping loss, and the task feature mapping loss to obtain the feature mapping loss of each feature component.

4. The method according to claim 1, wherein Generating an agent in the scenario reproduction training environment and training the agent's initial behavior pattern through a deep reinforcement learning model based on historical behavior data in the team collaboration knowledge base includes: Generate an agent in the scenario reproduction training environment, extract historical behavior data from a team collaboration knowledge base, and build a deep reinforcement learning model based on the historical behavior data, wherein the deep reinforcement learning model includes a behavior state space, an action space, and a reward function; The agent is iteratively trained using the deep reinforcement learning model, the policy network parameters of the agent are updated according to the feedback information of the reward function, and the initial behavior pattern of the agent is trained.

5. The method according to claim 1, wherein Calculating a bifurcation discriminant by combining the capability index and the Jacobian matrix, and predicting the transition probability of team capability based on the bifurcation discriminant includes: Combining the capability index with the eigenvalue of the Jacobian matrix to construct a bifurcation parameter vector, wherein the bifurcation parameter vector includes the capability index and the maximum real part of the eigenvalue; A bifurcation discriminant is calculated based on the bifurcation parameter vector through the determinant of the Jacobian matrix and the exponential smoothing term, and the bifurcation discriminant is used to detect the mutation characteristics of the team capability status; Calculating a critical value of the bifurcation discriminant within an observation time window, the critical value representing a threshold condition for a transition in the team capability state, and inputting a difference between the bifurcation discriminant and the critical value into a nonlinear mapping function; The nonlinear mapping function is constructed based on the sigmoid function to calculate the transition probability of the team's ability, and the transition probability represents the possibility of a sudden change in the team's ability; A confidence interval of the transition probability is constructed, and the upper and lower bounds of the confidence interval are calculated by an uncertainty function. The uncertainty function is dynamically adjusted as the bifurcation discriminant changes, thereby achieving an accurate prediction of the team capability transition.

6. The method according to claim 1, characterized in that Generating a team collaboration evaluation report based on the capability evolution trajectory; outputting a team collaboration optimization plan based on the team collaboration evaluation report, adjusting the parameter configuration of the scenario reproduction training environment in a targeted manner, and updating the training strategy of the multi-agent collaborative learning network include: The capability evolution trajectory includes multidimensional capability indicators. Based on these multidimensional capability indicators, the individual capability growth rate and team collaboration efficiency are calculated. The individual capability growth rate represents the speed of capability improvement, and the team collaboration efficiency represents the collaboration effect. The individual capability growth rate and the team collaboration efficiency are weighted and combined to obtain a comprehensive team collaboration score, which reflects the overall performance of the team. A multidimensional evaluation system is constructed based on the comprehensive team collaboration score, the multidimensional evaluation system including individual dimensions, team dimensions, and task dimensions, and a team collaboration evaluation report is generated; an optimization objective function is constructed based on the team collaboration evaluation report, the optimization objective function combines performance loss and optimization cost, and generates a team collaboration optimization plan based on resource constraints, time constraints, and effect constraints; The team collaboration optimization plan is converted into the parameter configuration of the scenario reproduction training environment, where the parameter configuration includes a training difficulty parameter and a scenario complexity parameter. The scenario reproduction training environment is dynamically adjusted, and the training strategy of the multi-agent collaborative learning network is updated based on the parameter configuration.

7. A team collaboration training system based on scene reproduction and multi-agent collaboration, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is configured to establish a team collaboration knowledge base based on pre-acquired team collaboration historical data, construct a multi-agent interaction scenario based on the team collaboration knowledge base, map virtual environment parameters, agent role information, and task objective information in the team collaboration knowledge base to the multi-agent interaction scenario, and generate a scenario reproduction training environment; The second unit is configured to generate an intelligent agent in the scenario reproduction training environment, and train an initial behavior pattern of the intelligent agent through a deep reinforcement learning model based on the historical behavior data in the team collaboration knowledge base; The third unit is used to build a multi-agent collaborative learning network enhanced by knowledge distillation, and to feed back the collaborative behavior data generated by the team members interacting with the agents in the scenario reproduction training environment to the multi-agent collaborative learning network in real time; The fourth unit is used to analyze the impact of team collaboration behavior on task outcomes using counterfactual reasoning based on the collaboration behavior data, construct a causal relationship diagram between collaboration behavior and task performance, dynamically quantify the intensity of synergy between team members, and track the evolution of capabilities in combination with a temporal attention mechanism; The fifth unit is used to generate a team collaboration evaluation report based on the capability evolution trajectory; based on the team collaboration evaluation report, output a team collaboration optimization plan, adjust the parameter configuration of the scenario reproduction training environment in a targeted manner, and update the training strategy of the multi-agent collaborative learning network.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images