AIGC-based Character Behavior Prediction Method, System and Storage Medium

Abstract

The present application relates to the field of data processing technology, and discloses a method, system and storage medium for predicting character behavior based on AIGC. The method comprises: performing spatiotemporal fractal analysis on character behavior data to obtain a character behavior pattern spectrum and perform structural extraction to obtain a character behavior topological structure; performing multi-scale entropy analysis on environmental information to obtain a scene complexity index; performing co-evolutionary calculation on the character behavior topological structure and the scene complexity index to obtain a behavior-environment interaction map and perform behavioral path analysis and screening to obtain a target behavior path set; performing causal emergence analysis on the target behavior path set to obtain a character behavior prediction result. By performing spatiotemporal fractal analysis on character behavior data, the present application can capture the complexity and self-similarity of behavior patterns at different spatiotemporal scales, thereby more comprehensively understanding the intrinsic structure and dynamic characteristics of character behavior, and improving the efficiency and accuracy of character behavior prediction based on AIGC.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method, system and storage medium for predicting character behavior based on AIGC. Background Art

[0002] Existing character behavior prediction technologies mainly include rule-based methods, statistical learning methods, and deep learning methods. Rule-based methods simulate the character decision-making process through predefined if-then statements, and can produce explainable prediction results in specific scenarios. Statistical learning methods, such as hidden Markov models and conditional random fields, learn behavior patterns by analyzing historical data and perform well in processing time series data. Deep learning methods, especially recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have made significant progress in capturing complex time series dependencies. These methods have been widely used in game AI, social media user behavior prediction, intelligent customer service and other fields.

[0003] However, these existing technologies have some obvious limitations. Rule-based methods lack flexibility, have difficulty coping with complex and changing situations, and require a lot of manual maintenance. Although statistical learning methods can learn patterns from data, they often have difficulty capturing the deep motivations and complex psychological states behind character behaviors. Deep learning methods perform well in processing large-scale data and complex pattern recognition, but their decision-making processes often lack explainability and it is difficult to provide a reasonable explanation for the prediction results. In addition, existing technologies generally have problems such as insufficient adaptability to dynamic changes in the environment, difficulty in processing multimodal data, and deficiencies in handling long-term dependencies and causal reasoning. In particular, in complex scenarios where multi-dimensional factors such as the character's personality traits, social background, emotional state, and long-term goals need to be considered, the performance of existing technologies is often unsatisfactory. Summary of the invention

[0004] The present application provides a method, system and storage medium for predicting character behavior based on AIGC, which are used to improve the efficiency and accuracy of character behavior prediction based on AIGC.

[0005] In the first aspect, the present application provides a character behavior prediction method based on AIGC, and the character behavior prediction method based on AIGC includes: performing spatiotemporal fractal analysis on pre-collected character behavior data to obtain a character behavior pattern spectrum; performing structural extraction on the character behavior pattern spectrum to obtain a character behavior topological structure; performing multi-scale entropy analysis on environmental information to obtain a scene complexity index; performing co-evolutionary calculation on the character behavior topological structure and the scene complexity index to obtain a behavior-environment interaction map; performing behavior path analysis and screening on the behavior-environment interaction map through a simulated annealing algorithm to obtain a target behavior path set; performing causal emergence analysis on the target behavior path set to obtain a character behavior prediction result.

[0006] In combination with the first aspect, in a first implementation method of the first aspect of the present application, the pre-collected character behavior data is subjected to spatiotemporal fractal analysis to obtain a character behavior pattern spectrum, including: semantically annotating the character behavior data to obtain a semantically enhanced behavior sequence, and decoupling the semantically enhanced behavior sequence in spatiotemporal dimensions to obtain a time series and a spatial trajectory; performing multi-scale decomposition on the time series to obtain a time scale feature set, and performing fractal dimension calculation on the spatial trajectory to obtain a spatial complexity index; fusing the time scale feature set and the spatial complexity index to obtain a spatiotemporal feature matrix; performing self-similarity analysis on the spatiotemporal feature matrix to obtain a fractal feature vector; clustering the fractal feature vector to obtain a behavior pattern cluster center; performing time series correlation analysis on the behavior pattern cluster center to obtain a behavior transition probability matrix; performing feature decomposition on the behavior transition probability matrix to obtain main behavior pattern eigenvalues; performing spectrum transformation on the main behavior pattern eigenvalues ​​to obtain a character behavior pattern spectrum.

[0007] In combination with the first aspect, in a second implementation of the first aspect of the present application, the structural extraction of the role behavior pattern spectrum to obtain the role behavior topological structure includes: performing generative semantic segmentation on the role behavior pattern spectrum to obtain behavior pattern units, and performing context-aware encoding on the behavior pattern units to obtain a coded behavior sequence; performing temporal association analysis on the coded behavior sequence to obtain a behavior transition sequence, and performing frequency statistics on the behavior transition sequence to obtain a frequency matrix; performing adaptive threshold filtering on the frequency matrix to obtain a main behavior transition path, and performing graph structuring on the main behavior transition path to obtain an initial behavior graph; performing centrality calculation on the initial behavior graph to obtain a key behavior node set, and performing connectivity analysis on the key behavior node set to obtain a behavior subgraph set; performing hierarchical clustering on the behavior subgraph set to obtain a multi-level behavior structure, and performing topological relationship extraction on the multi-level behavior structure to obtain a role behavior topological structure.

[0008] In combination with the first aspect, in a third implementation method of the first aspect of the present application, the multi-scale entropy analysis of the environmental information to obtain the scene complexity index includes: performing generative context encoding on the environmental information to obtain a semantically enhanced environmental representation, and performing spatiotemporal decomposition on the semantically enhanced environmental representation to obtain a multi-scale environmental feature sequence; performing conditional sample entropy calculation on the multi-scale environmental feature sequence to obtain an initial entropy value matrix, and performing adaptive threshold segmentation on the initial entropy value matrix to obtain discretized entropy features; performing role behavior correlation analysis on the discretized entropy features to obtain a behavior-environment coupling index, and performing recursive pattern extraction on the behavior-environment coupling index to obtain an environmental complexity pattern; performing multi-dimensional feature fusion on the environmental complexity pattern to obtain a target complexity vector; and performing time series dynamic modeling on the target complexity vector to obtain the scene complexity index.

[0009] In combination with the first aspect, in a fourth implementation method of the first aspect of the present application, the co-evolutionary calculation of the character behavior topological structure and the scene complexity index is performed to obtain a behavior-environment interaction graph, including: generative semantic embedding of the character behavior topological structure to obtain a behavior semantic vector, and context-aware mapping of the scene complexity index to obtain an environment semantic vector; performing a tensor product operation on the behavior semantic vector and the environment semantic vector to obtain an initial interaction tensor; performing sparsity constrained decomposition on the initial interaction tensor to obtain key interaction factors, and adaptively weighting the key interaction factors to obtain weighted interaction features; performing dynamic time warping on the weighted interaction features to obtain an aligned interaction sequence, and performing conditional entropy calculation on the aligned interaction sequence to obtain a mutual information matrix; performing spectral clustering on the mutual information matrix to obtain behavior-environment clusters, and performing topological structure reconstruction on the behavior-environment clusters to obtain an initial interaction graph; performing evolutionary dynamics modeling on the initial interaction graph to obtain a behavior-environment interaction graph.

[0010] In combination with the first aspect, in a fifth implementation of the first aspect of the present application, the behavior path analysis and screening of the behavior-environment interaction graph by a simulated annealing algorithm to obtain a target behavior path set includes: performing node importance calculation on the behavior-environment interaction graph to obtain a target behavior node set, and performing generative path construction on the target behavior node set to obtain an initial behavior path set; normalizing the path length of the initial behavior path set to obtain standardized path representation data; calculating the energy value of the standardized path representation data by a preset path energy function to obtain an initial path energy value; performing annealing temperature initialization processing on the initial path energy value to obtain a starting temperature parameter; iteratively optimizing the standardized path representation data by the simulated annealing algorithm to obtain an optimized behavior path, and performing environmental constraint index analysis on the optimized behavior path to obtain a feasibility index; performing adaptive threshold screening on the feasibility index to obtain a candidate behavior path subset, and performing difference analysis on the candidate behavior path subset to obtain a path difference matrix; performing cluster sampling based on the path difference matrix to obtain a target behavior path set.

[0011] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, the target behavior path set is subjected to causal emergence analysis to obtain a role behavior prediction result, including: performing sequence enhancement on the target behavior path set to obtain an enhanced behavior chain; performing spatiotemporal coupling analysis on the enhanced behavior chain to obtain a behavior interdependence grid; performing topological relationship extraction on the behavior interdependence grid to obtain an original causal topology, and performing temporal coordination verification on the original causal topology to obtain an optimized causal structure; performing emergent feature extraction on the optimized causal structure to obtain multi-level causal association data, and performing interpretable transformation on the multi-level causal association to obtain target causal criterion data; performing scene adaptability analysis on the target causal criterion data to obtain environmental perception causal data, and performing behavioral random simulation analysis on the environmental perception causal data to obtain a behavior possibility spectrum; performing cross-modal integration on the behavior possibility spectrum to obtain a target behavior descriptor, and performing temporal feature extraction and probability distribution calculation based on the target behavior descriptor to obtain role behavior prediction data.

[0012] In a second aspect, the present application provides a role behavior prediction system based on AIGC, and the role behavior prediction system based on AIGC includes:

[0013] The spatiotemporal fractal module is used to perform spatiotemporal fractal analysis on the pre-collected character behavior data to obtain the character behavior pattern spectrum;

[0014] A structure extraction module, used for performing structure extraction on the role behavior pattern spectrum to obtain a role behavior topological structure;

[0015] The first analysis module is used to perform multi-scale entropy analysis on environmental information to obtain scene complexity indicators;

[0016] An index calculation module, used for performing co-evolutionary calculation on the character behavior topological structure and the scene complexity index to obtain a behavior-environment interaction map;

[0017] A path screening module, used to analyze and screen the behavior path of the behavior-environment interaction graph through a simulated annealing algorithm to obtain a target behavior path set;

[0018] The second analysis module is used to perform causal emergence analysis on the target behavior path set to obtain role behavior prediction results.

[0019] A third aspect of the present application provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned AIGC-based character behavior prediction method.

[0020] In the technical solution provided by the present application, by performing spatiotemporal fractal analysis on the character behavior data, the method can capture the complexity and self-similarity of the behavior pattern at different spatiotemporal scales, so as to more comprehensively understand the intrinsic structure and dynamic characteristics of the character behavior. This multi-scale analysis method helps to identify subtle behavior patterns that may be overlooked by conventional statistical methods, and provides a richer information basis for subsequent behavior prediction. The character behavior pattern spectrum is structurally extracted to obtain the character behavior topological structure. This step not only reveals the correlation between different behaviors, but also constructs a hierarchical behavior model, so that the system can better understand and predict complex behavior sequences. The extraction of this topological structure helps to identify key behavior nodes and behavior conversion paths. By performing multi-scale entropy analysis on environmental information, the scene complexity index is obtained, which can accurately quantify the complexity and uncertainty of the environment, which is crucial for understanding the impact of the environment on the character behavior. This quantification of environmental complexity can dynamically adjust the behavior prediction strategy according to different environmental conditions, greatly improving the environmental adaptability of the prediction. The present application obtains the behavior-environment interaction map through co-evolutionary calculation, effectively simulating the dynamic interaction between character behavior and environment, so that the prediction model can take into account the immediate and long-term impact of environmental changes on character behavior. The introduction of this dynamic interaction model significantly improves the timeliness and accuracy of predictions, especially in rapidly changing environments. At the same time, the use of simulated annealing algorithm to analyze and screen the behavior path of the behavior-environment interaction graph not only optimizes the selection of behavior paths, but also improves the computational efficiency of predictions, and improves the efficiency and accuracy of AIGC-based role behavior predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0022] Figure 1 This is a schematic diagram of an embodiment of a method for predicting a role behavior based on AIGC in an embodiment of the present application;

[0023] Figure 2 This is a schematic diagram of an embodiment of a character behavior prediction system based on AIGC in an embodiment of the present application. DETAILED DESCRIPTION

[0024] The embodiments of the present application provide a method, system and storage medium for predicting role behavior based on AIGC. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the role behavior prediction method based on AIGC includes:

[0026] Step S101, performing spatiotemporal fractal analysis on the pre-collected character behavior data to obtain a character behavior pattern spectrum;

[0027] It is understandable that the execution subject of the present application may be a role behavior prediction system based on AIGC, or may be a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0028] Specifically, the key step of this method is to perform spatiotemporal fractal analysis on the pre-collected character behavior data to obtain the character behavior pattern spectrum. The pre-collected character behavior data contains the behavior information of the character at different time and space points. These data may come from various sources, such as character descriptions in literary works, character lines and behavior records in movie and TV scripts, and character behavior cases in psychological research. Spatiotemporal fractal analysis is a complex system analysis method used to study the self-similarity and complexity of data in time and space dimensions. The character behavior data is preprocessed, including data cleaning, standardization and structuring. Then, the data is decomposed into multiple scales using wavelet transform to capture the behavior characteristics at different time scales. Then, the fractal dimension at each scale is calculated, which reflects the complexity of the behavior pattern. At the same time, the spatial autocorrelation analysis method, such as Moran's I index, is used to evaluate the distribution characteristics of the behavior in space. Then, a spatiotemporal fractal spectrum is constructed, which reflects the complexity and self-similarity of the character behavior at different spatiotemporal scales. By analyzing the spatiotemporal fractal spectrum, the main patterns and characteristics of the character behavior are identified to form a character behavior pattern spectrum. This pattern spectrum includes the main characteristics of character behavior, such as the periodicity, suddenness, and spatial clustering of behavior.

[0029] For example, through spatiotemporal fractal analysis, it is found that the character's behavior shows obvious periodicity on a daily scale (the fractal dimension is close to 1), and shows high complexity on a monthly scale (the fractal dimension is close to 1.8). Spatially, the character's behavior is highly clustered in an urban environment (Moran's I index is 0.75), but is more dispersed in a rural environment (Moran's I index is 0.2).

[0030] Step S102, extracting the structure of the role behavior pattern spectrum to obtain the role behavior topological structure;

[0031] Specifically, extracting the structure of the role behavior pattern spectrum and obtaining the role behavior topology is a key step, which aims to extract the intrinsic connection and structural relationship between behaviors from the role behavior pattern spectrum. First, the role behavior pattern spectrum contains the complexity and self-similarity information of the role behavior at different spatiotemporal scales. The structure extraction process first normalizes the pattern spectrum to ensure that data at different scales are comparable. Then, the spectral clustering algorithm is used to perform cluster analysis on the normalized data to identify behavior pattern clusters with similar characteristics.

[0032] Feature extraction is performed on each behavior pattern cluster, including calculating indicators such as the centrality, connectivity, and frequency of the pattern. Based on these features, an association network between behavior patterns is constructed, in which nodes represent behavior patterns and edges represent the strength of association between patterns. In order to extract the skeleton structure of the network, the minimum spanning tree algorithm is applied to retain the most important behavior pattern connections. At the same time, the community detection algorithm is used to identify densely connected substructures in the network, which represent sets of interrelated behavior patterns. By topologically sorting the behavior pattern network, a hierarchical relationship between behavior patterns is established, thereby obtaining the role behavior topology. This topology is a directed acyclic graph that reflects the dependency and influence relationship between behavior patterns. For example, the role behavior pattern spectrum may include multiple behavior patterns such as "work", "social", and "leisure". Through structural extraction, it is found that there is a strong correlation between the "work" mode and the "social" mode (correlation strength 0.8), while the correlation with the "leisure" mode is weaker (correlation strength 0.3). Community detection finds that "work" and "social" belong to the same behavior community, while "leisure" belongs to another community. Topological sorting shows that the "work" mode usually occurs before the "social" mode, while the "leisure" mode is the last.

[0033] Step S103: Perform multi-scale entropy analysis on the environmental information to obtain a scene complexity index;

[0034] Specifically, it is an important step to perform multi-scale entropy analysis on environmental information to obtain scene complexity indicators, which aims to quantify the complexity and uncertainty of the environment. First, environmental information includes various physical and social attributes of the scene in which the character is located, such as spatial layout, population density, frequency of social activities, etc. Multi-scale entropy analysis is a method for evaluating signal complexity, which captures the multi-level complexity of the system by calculating information entropy at different time or spatial scales. The environmental information is preprocessed, including data normalization and denoising. Then, the environmental information sequence is divided into subsequences of different lengths using a sliding window technique, each length corresponding to an analysis scale. Sample entropy is calculated for each subsequence at each scale, and the sample entropy reflects the irregularity and complexity of the sequence. Then, by calculating the average value of sample entropy at different scales, the multi-scale entropy value is obtained, which comprehensively reflects the complexity of the environment at multiple scales.

[0035] In order to further enhance the robustness of the analysis, the concept of fuzzy entropy is introduced. Fuzzy entropy uses fuzzy set theory to deal with uncertainty in data and can better handle noise and ambiguity in environmental information. At the same time, the permutation entropy analysis method is used to capture the dynamic relationship between environmental variables. Finally, the results of multi-scale entropy, fuzzy entropy and permutation entropy are weighted and fused to obtain a comprehensive scene complexity index.

[0036] For example, environmental data including population density, building density, and social activity frequency were collected. Multi-scale entropy analysis was performed on these data. At a small scale (such as hourly), the sample entropy value was 1.2, indicating that the environment changed greatly in a short period of time; at a large scale (such as weekly), the sample entropy value dropped to 0.8, indicating that the environment was relatively stable in the long run. The result of fuzzy entropy analysis was 0.9, reflecting a certain degree of uncertainty in the environment. Permutation entropy analysis showed that there was a strong dynamic correlation between population density and social activity frequency, with a permutation entropy value of 2.1. These results were weighted and fused (weights were 0.4, 0.3, and 0.3, respectively), and the final scene complexity index was 1.41. This index comprehensively reflects the complexity of the environment at different scales and dimensions, and provides important environmental background information for subsequent role behavior prediction.

[0037] Step S104, performing co-evolutionary calculation on the role behavior topological structure and the scene complexity index to obtain a behavior-environment interaction map;

[0038] Specifically, co-evolutionary computation of the character behavior topology and scene complexity index to obtain the behavior-environment interaction map is a key step, aiming to simulate the dynamic interaction between character behavior and the environment. First, the character behavior topology is a network that describes the relationship between behavior patterns, while the scene complexity index quantifies the complexity of the environment. The core idea of ​​co-evolutionary computation is to reveal the deep interaction between behavior and environment by simulating the mutual adaptation process of behavior and environment.

[0039] First, the character behavior topology is converted into an adjacency matrix, where the matrix elements represent the strength of the association between behaviors. The scene complexity index is used as part of the environment fitness function. Next, a genetic algorithm is used to evolve the behavior topology, where each individual represents a possible behavior strategy. In each generation of evolution, the adaptability of each behavior strategy is evaluated according to the environment fitness function, and strategies with high fitness have a greater probability of being retained and propagated. At the same time, the environment will also be adjusted according to the current optimal behavior strategy, which is achieved by updating the scene complexity index.

[0040] In order to capture the dynamic characteristics of behavior-environment interaction, a time series analysis method is introduced. A dynamic time warping algorithm is used to align behavior sequences and environmental changes on different time scales, thereby identifying time-dependent interaction patterns. In addition, a multi-agent system is used to simulate the interaction of multiple roles in the same environment. Each agent represents a role. They affect the environment through their behavior, and environmental changes in turn affect the decisions of each agent.

[0041] After multiple rounds of iterative calculations, a stable behavior-environment interaction graph is finally obtained. This graph is a dynamic network, where nodes represent behavior patterns and environmental factors, and edges represent the intensity and direction of the interaction between them. The graph also includes a time dimension, reflecting the evolution of the interaction relationship over time.

[0042] For example, the initial behavior topology contains three main nodes: "work", "social" and "leisure", and the scene complexity index is 1.41. After 100 rounds of coevolutionary calculations, it was found that the "work" behavior has enhanced adaptability in a complex environment (fitness increased from 0.6 to 0.8), while the "leisure" behavior is more frequent in a less complex environment (the frequency increases by 30% when the complexity is less than 1.0). At the same time, the complexity of the environment also fluctuates with the changes in the character's behavior. When the character frequently "socializes", the complexity increases slightly (an average increase of 0.1). The final behavior-environment interaction map shows that the correlation strength between "work" and "social" behaviors increases with the increase in environmental complexity (the correlation coefficient increases from 0.3 to 0.7), while "leisure" behavior is negatively correlated with environmental complexity (correlation coefficient -0.5).

[0043] Step S105, analyzing and screening the behavior path of the behavior-environment interaction graph by using a simulated annealing algorithm to obtain a target behavior path set;

[0044] Specifically, the simulated annealing algorithm is used to analyze and screen the behavior paths of the behavior-environment interaction graph to obtain the target behavior path set, which is a key step to find the optimal behavior path from the complex interaction graph. The simulated annealing algorithm is a heuristic optimization method that simulates the change of atomic energy state during metal cooling and is used to solve combinatorial optimization problems. In this method, the behavior path is regarded as an optimization problem, and the goal is to find the most suitable behavior sequence under a given environment.

[0045] The behavior-environment interaction graph is transformed into a weighted directed graph, where nodes represent behavior states, edges represent behavior transitions, and the weights of edges reflect the fitness of transitions. The initial temperature is set to a high value to allow a wide search space. Then, an initial behavior path is randomly generated as the current solution. In each iteration, the current path is slightly modified, such as exchanging the positions of two behavior nodes or inserting new behavior nodes, to generate a candidate solution. The fitness of the candidate solution is calculated. If it is better than the current solution, it is accepted; if it is worse than the current solution, it is accepted with a certain probability, and this probability decreases as the temperature decreases. The temperature is gradually reduced according to a predetermined cooling strategy, such as exponential decay. In order to improve the search efficiency, an adaptive neighborhood search strategy is introduced. The size of the search neighborhood is dynamically adjusted according to the quality of the current solution. A large neighborhood is used in the early stage of the search to increase diversity, and a small neighborhood is used in the later stage to increase accuracy. At the same time, the taboo search technique is used to maintain a short-term memory list to avoid repeated visits to the recently explored solutions, so as to escape from the local optimum.

[0046] During the screening process, a multi-objective evaluation function is used to comprehensively consider the fitness, diversity, and complexity of the behavior path. Fitness is calculated by summing the fitness of each behavior node on the path under the current environment. Diversity is measured by calculating the number of different behavior types in the path. Complexity considers the path length and the difficulty of behavior conversion. These goals are combined into a comprehensive score through weighted summation.

[0047] Finally, the algorithm outputs a target behavior path set, which includes multiple high-quality candidate paths. These paths achieve a good balance between fitness, diversity, and complexity, providing a reliable basis for subsequent behavior prediction.

[0048] For example, when analyzing the training behavior of a professional athlete, the behavior-environment interaction map includes nodes such as "physical training", "technical training", "tactical learning", and "recovery". The initial temperature is set to 100 and the cooling rate is 0.95. After 1,000 iterations, the temperature dropped to 5.18. In this process, the algorithm explored about 5,000 candidate paths. The target behavior path set finally screened out contains 10 high-quality paths. One of the optimal paths is: physical training (fitness 0.8) > technical training (0.9) > recovery (0.7) > tactical learning (0.85) > technical training (0.95), with a total fitness of 4.2. While maintaining high fitness, this path also reflects the diversity and scientific nature of training, which is in line with the training rules of professional athletes.

[0049] Step S106: Perform causal emergence analysis on the target behavior path set to obtain role behavior prediction results.

[0050] Specifically, the last key step of this method is to perform causal emergence analysis on the target behavior path set and obtain the role behavior prediction results, which aims to extract deep causal relationships from the optimized behavior paths and make behavior predictions based on these relationships. Causal emergence analysis is a complex system analysis method used to identify nonlinear and non-obvious causal relationships in the system. In this method, it is used to discover potential key factors in the behavior sequence that may affect future behaviors. Each path in the target behavior path set is temporally decomposed to convert the continuous behavior sequence into discrete time steps. Then, the conditional random field model is used to capture the local dependencies between behaviors. This model takes into account the contextual information of the behavior and can better describe complex sequence relationships. Next, the Granger causality test is applied to identify the temporal dependencies between behaviors, which helps to determine which previous behaviors have a significant impact on subsequent behaviors.

[0051] In order to deal with nonlinear relationships, transfer entropy analysis based on information theory is introduced. Transfer entropy calculates the information contribution of one behavior to another behavior, which can capture more subtle causal relationships. At the same time, a dynamic Bayesian network is used to simulate the dynamic probabilistic dependencies between behaviors. This network evolves over time, reflecting the dynamic changes of causal relationships. After identifying the key causal relationships, a deep reinforcement learning method is used to predict future behaviors. Specifically, the Double Deep Q Network (Double DQN) algorithm is used to use the identified causal relationships as state features and the behavior selection as the action space. Through interaction and feedback with the environment, the model learns the optimal behavior strategy.

[0052] Finally, in order to enhance the robustness of prediction, an ensemble learning method is used. The results of multiple prediction models are weighted and fused, such as decision trees, support vector machines, and neural networks. Each model is trained based on different causal relationship features. The fused result is the final character behavior prediction result.

[0053] For example, when analyzing the game behavior of a professional chess player, the target behavior path set contains multiple high-quality game strategy sequences. Through causal emergence analysis, some non-intuitive causal relationships were discovered. For example, there is a significant Granger causal relationship (p value < 0.01) between conservative strategies (such as a robust layout) in the early stage of the game and offensive moves in the middle game. Transfer entropy analysis shows that the opponent's specific response (such as trying to crack the layout) has a high information contribution to the player's behavior choices in the next 20 steps (transfer entropy value 0.45). The dynamic Bayesian network reveals the probability of a chess player's strategy switching in different situations, such as the probability of switching from defense to counterattack presents a U-shaped curve as the game progresses.

[0054] In the embodiment of the present application, by performing spatiotemporal fractal analysis on the character behavior data, the method can capture the complexity and self-similarity of the behavior pattern at different spatiotemporal scales, so as to more comprehensively understand the intrinsic structure and dynamic characteristics of the character behavior. This multi-scale analysis method helps to identify subtle behavior patterns that may be overlooked by conventional statistical methods, and provides a richer information basis for subsequent behavior prediction. The character behavior pattern spectrum is structurally extracted to obtain the character behavior topological structure. This step not only reveals the correlation between different behaviors, but also constructs a hierarchical behavior model, so that the system can better understand and predict complex behavior sequences. The extraction of this topological structure helps to identify key behavior nodes and behavior conversion paths. By performing multi-scale entropy analysis on environmental information, the scene complexity index is obtained, which can accurately quantify the complexity and uncertainty of the environment, which is crucial for understanding the impact of the environment on the character behavior. The quantification of this environmental complexity can dynamically adjust the behavior prediction strategy according to different environmental conditions, greatly improving the environmental adaptability of the prediction. The present application obtains the behavior-environment interaction map through co-evolutionary calculation, effectively simulating the dynamic interaction between the character behavior and the environment, so that the prediction model can take into account the immediate and long-term effects of environmental changes on the character behavior. The introduction of this dynamic interaction model significantly improves the timeliness and accuracy of predictions, especially in rapidly changing environments. At the same time, the use of simulated annealing algorithm to analyze and screen the behavior path of the behavior-environment interaction graph not only optimizes the selection of behavior paths, but also improves the computational efficiency of predictions, and improves the efficiency and accuracy of AIGC-based role behavior predictions.

[0055] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0056] (1) Semantically annotate the character behavior data to obtain a semantically enhanced behavior sequence, and decouple the semantically enhanced behavior sequence in terms of time and space to obtain a time series and spatial trajectory;

[0057] (2) Perform multi-scale decomposition on the time series to obtain the time scale feature set, and calculate the fractal dimension of the spatial trajectory to obtain the spatial complexity index;

[0058] (3) Fusing the time scale feature set and the spatial complexity index to obtain the spatiotemporal feature matrix;

[0059] (4) Perform self-similarity analysis on the spatiotemporal feature matrix to obtain the fractal feature vector;

[0060] (5) Cluster the fractal feature vectors to obtain the behavior pattern cluster center;

[0061] (6) Perform temporal correlation analysis on the behavior pattern cluster centers to obtain the behavior conversion probability matrix;

[0062] (7) Perform eigendecomposition on the behavior transition probability matrix to obtain the eigenvalues ​​of the main behavior patterns;

[0063] (8) Perform spectrum transformation on the characteristic values ​​of the main behavior patterns to obtain the character behavior pattern spectrum.

[0064] Specifically, semantically annotating character behavior data and decoupling spatiotemporal dimensions is a complex process that aims to extract meaningful behavior patterns from raw data. First, natural language processing techniques are used to semantically annotate behavior data. This involves using a pre-trained language model to understand the contextual meaning of behavior descriptions and assigning semantic labels to each behavior. In this way, raw behavior data is transformed into enhanced behavior sequences that carry rich semantic information.

[0065] The time and space dimensions of semantically enhanced behavior sequences are decoupled. The time dimension is represented by extracting the timestamp and duration of the behavior, while the space dimension is represented by the location coordinates of the behavior. Multi-scale decomposition of time series is achieved using wavelet transform, which can decompose the time series into subsequences of different frequencies, each of which represents a behavior feature on a time scale. The formula of wavelet transform can be expressed as:

[0066] ;

[0067] in, is the wavelet coefficient, a is the scale parameter, b is the translation parameter, is the original signal, is a wavelet function.

[0068] The box counting method is used to calculate the fractal dimension of the spatial trajectory, and the calculation formula is:

[0069] ;

[0070] Where D is the fractal dimension, The size required to cover the entire collection is The number of boxes.

[0071] The fusion of the time scale feature set and the spatial complexity index is achieved by constructing a spatiotemporal feature matrix. The self-similarity analysis of the spatiotemporal feature matrix uses the recursive graph technology to calculate the similarity between different time points and obtain the self-similarity matrix. The calculation formula of the self-similarity matrix is:

[0072] ;

[0073] in, are the elements of the recursive matrix, is the Heaviside function, is the threshold value, and is the state vector.

[0074] The clustering of fractal feature vectors is implemented using the K-means algorithm. The time series correlation analysis of the behavior pattern cluster centers is performed, and the Markov chain model is used to calculate the transition probability between behavior patterns. The eigendecomposition of the behavior transition probability matrix uses singular value decomposition to extract the main eigenvectors and eigenvalues ​​of the matrix. Finally, the eigenvalues ​​of the main behavior patterns are spectrally transformed, and the eigenvalues ​​are converted from the time domain to the frequency domain using Fourier transform.

[0075] In the AIGC-based character behavior prediction method, it is assumed that the behavior data of a virtual game character is analyzed. The original data may contain descriptions such as "the character fights", "the character buys equipment", "the character completes the task", etc. Through semantic annotation, these descriptions are converted into standardized behavior types and intensity levels. For example, "the character fights" may be annotated as "behavior type: fighting, intensity: high".

[0076] After time-space decoupling, the character's daily behavior time series and spatial coordinate sequence in the game world were obtained. Multi-scale decomposition shows that there is a cycle of combat-rest on the hourly scale, and a law of "main quest-side quest-free exploration" on the daily scale. The fractal dimension of the spatial trajectory is calculated to be 1.8, indicating that the character's activities in the game world have a high spatial complexity. The self-similarity analysis of the spatiotemporal feature matrix shows that the character's weekly behavior patterns have a similarity of 65%. Cluster analysis obtained 5 main behavioral pattern cluster centers, corresponding to combat, quests, social activities, economic activities and exploration. Time series association analysis shows that the probability of transition from combat to economic activities (such as purchasing equipment and repairing equipment) is 0.7, while the probability of transition from quests to exploration is 0.5.

[0077] The first three main eigenvalues ​​obtained after feature decomposition are 2.4, 1.9, and 1.3, respectively, and the corresponding eigenvectors reflect the three most important combinations of behavioral patterns. The behavioral pattern spectrum after spectrum transformation has significant peaks at 24 hours and 7 days, reflecting the existence of daily cycles and weekly cycles. For example, AIGC can generate a daily behavior schedule that conforms to the character's habits based on these patterns: combat activities in the morning, economic affairs at noon, tasks in the afternoon, and social and exploration activities in the evening. This data-driven character behavior generation method greatly improves the authenticity and coherence of the generated content.

[0078] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0079] (1) Generative semantic segmentation is performed on the character behavior pattern spectrum to obtain behavior pattern units, and context-aware encoding is performed on the behavior pattern units to obtain encoded behavior sequences;

[0080] (2) Perform temporal correlation analysis on the coded behavior sequence to obtain the behavior conversion sequence, and perform frequency statistics on the behavior conversion sequence to obtain a frequency matrix;

[0081] (3) Adaptively filter the frequency matrix by threshold to obtain the main behavior transition path, and then perform graph-structured processing on the main behavior transition path to obtain the initial behavior graph;

[0082] (4) Calculate the centrality of the initial behavior graph to obtain the key behavior node set, and perform connectivity analysis on the key behavior node set to obtain the behavior subgraph set;

[0083] (5) Perform hierarchical clustering on the behavior subgraph set to obtain a multi-level behavior structure, and extract topological relationships from the multi-level behavior structure to obtain the role behavior topological structure.

[0084] Specifically, in the AIGC-based character behavior prediction method, generative semantic segmentation of the character behavior pattern spectrum is a key step, which aims to decompose continuous behavior patterns into discrete, meaningful behavior units. This process uses generative AI models, such as the transformer architecture, to understand and segment complex behavior sequences. The segmented behavior pattern units are then context-aware encoded, a step that takes into account the contextual information in which the behavior occurs and enhances the richness of the behavior representation.

[0085] Context-aware encoding can be expressed by the following formula:

[0086] ;

[0087] in, is the code of the ith behavior unit, is the behavioral unit itself, is context information, is the encoding function.

[0088] The encoded behavior sequence is subjected to temporal correlation analysis to generate a behavior transition sequence. This step captures the temporal dependency between behaviors and is crucial for identifying the behavior pattern of the character. The frequency matrix F is obtained by performing frequency statistics on the behavior transition sequence, whose elements represents the frequency of transition from behavior i to behavior j.

[0089] The frequency matrix is ​​filtered by the adaptive threshold to screen out the main behavior conversion path. The adaptive threshold T can be calculated by the following formula:

[0090] ;

[0091] Among them, is the average value of the frequency matrix, is the standard deviation, is an adjustable parameter.

[0092] The main behavior conversion path is processed through graph structuring to form an initial behavior graph. The nodes of this graph represent behaviors, and the edges represent conversion relationships. Centrality calculation is performed on the initial behavior graph to identify key behavior nodes. Commonly used centrality metrics include degree centrality, betweenness centrality, and eigenvector centrality.

[0093] The set of key behavior nodes undergoes connectivity analysis to obtain a set of behavior subgraphs. This step aims to discover closely connected behavior groups. Hierarchical clustering is performed on the set of behavior subgraphs to generate a multi-level behavior structure. Hierarchical clustering can be represented by the following formula:

[0094] ;

[0095] Among them, is the cluster and the distance between, is the distance between individual behaviors x and y.

[0096] Finally, topological relationship extraction is performed on the multi-level behavior structure to obtain a role behavior topological structure. This structure reflects the hierarchical and dependency relationships between role behaviors, providing an important basis for subsequent behavior prediction.

[0097] For example, in an open-world game based on AIGC, the behavior data of a virtual character is analyzed. The initial role behavior pattern spectrum contains complex behavior sequences such as "exploration", "combat", "trading", "dialogue", etc. After generative semantic segmentation, finer-grained behavior pattern units are obtained, such as "hiking exploration", "riding exploration", "melee attack", "ranged attack", "buying items", "selling items", "friendly dialogue", "threatening dialogue", etc. These behavior units are encoded through context awareness, considering the time, location, and previous behaviors when the behavior occurs. For example, the encoding of "buying items" not only includes the behavior itself but also information such as whether it is in a town or the wild, whether it is day or night, and whether there was a battle before.

[0098] Temporal correlation analysis shows that "combat" is often followed by "buying items" or "dialogue", and "exploration" is often followed by "combat" or "dialogue". In the frequency matrix obtained by frequency statistics, the conversion frequency from "exploration" to "combat" is 0.3, and the frequency from "combat" to "buying items" is 0.5.

[0099] After filtering with an adaptive threshold (assuming the calculated threshold is 0.2), the main behavior transition paths are retained, such as "explore-fight-buy items", "dialogue-trade-explore", etc. These paths constitute the initial behavior graph. The centrality calculation shows that "explore" and "dialogue" are the most critical behavior nodes, and they have the most connections with other behaviors. Connectivity analysis further reveals several closely connected behavior subgraphs, such as "explore-fight-buy items" and "dialogue-trade-explore". Hierarchical clustering organizes these behaviors into a multi-layer structure, with the top layer possibly being "interaction" and "independent activity", and the next layer being subdivided into categories such as "social", "economy", and "adventure". The final character behavior topology clearly shows the relationship and hierarchy between various behaviors.

[0100] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0101] (1) Generate context encoding of environmental information to obtain semantically enhanced environmental representation, and then perform spatiotemporal decomposition of the semantically enhanced environmental representation to obtain a multi-scale environmental feature sequence;

[0102] (2) Calculate the conditional sample entropy of the multi-scale environmental feature sequence to obtain the initial entropy value matrix, and perform adaptive threshold segmentation on the initial entropy value matrix to obtain discretized entropy features;

[0103] (3) Perform role-behavior correlation analysis on the discretized entropy features to obtain the behavior-environment coupling index, and perform recursive pattern extraction on the behavior-environment coupling index to obtain the environment complexity model;

[0104] (4) Perform multi-dimensional feature fusion on the environmental complexity pattern to obtain the target complexity vector;

[0105] (5) Perform temporal dynamic modeling on the target complexity vector to obtain the scene complexity index.

[0106] Specifically, in the AIGC-based character behavior prediction method, generative context encoding of environmental information is a key step, aiming to capture the semantic features of the environment. This process uses generative AI models such as GPT or BERT to convert raw environmental data into vector representations with rich semantic information. Generative context encoding can be expressed by the following formula:

[0107] ;

[0108] in, is the semantic vector representation of the environment, is the original environment information, is context information, is a generative encoding function.

[0109] The semantic enhanced environment representation is decomposed in time and space to obtain a multi-scale environment feature sequence. This step uses wavelet transform or Fourier transform to decompose the environment information at different time and space scales. Next, the conditional sample entropy is calculated for the multi-scale environment feature sequence to obtain the initial entropy value matrix. The conditional sample entropy calculation formula is as follows:

[0110] ;

[0111] in, is the conditional sample entropy, Represents environmental characteristics, represents conditions (such as time or spatial location), yes and The joint probability distribution of yes The marginal probability distribution of .

[0112] The initial entropy value matrix is ​​segmented by an adaptive threshold to obtain discretized entropy features. The adaptive threshold can be determined by the OTSU method or an entropy-based method. The discretized entropy features are subjected to role-behavior correlation analysis to obtain the behavior-environment coupling index. This step uses mutual information or Pearson correlation coefficient to quantify the degree of association between behavior and environmental features. Recursive pattern extraction is performed on the behavior-environment coupling index to obtain the environmental complexity pattern. Recursive pattern extraction can be implemented using the dynamic time warping (DTW) algorithm or recurrent neural network (RNN). The environmental complexity pattern reflects the regularity and complexity of environmental changes.

[0113] Perform multi-dimensional feature fusion on the environmental complexity pattern to obtain a comprehensive complexity vector. Feature fusion can be achieved through weighted summation or principal component analysis (PCA). Finally, perform time series dynamic modeling on the comprehensive complexity vector to obtain the scene complexity index. Time series dynamic modeling can use the ARIMA model or the long short-term memory network (LSTM) to capture the temporal variation of complexity.

[0114] For example, in an open world game based on AIGC, environmental information includes factors such as terrain, weather, and NPC distribution. Through generative context encoding, the original environmental data is converted into a 300-dimensional semantic vector. For example, "mountainous terrain, rainy weather, and dense hostile NPCs" may be encoded as a specific vector representation.

[0115] After the spatiotemporal decomposition, we obtained sequences reflecting the characteristics of different scales, such as hourly weather changes, daily NPC activity patterns, weekly terrain changes, etc. The conditional sample entropy is calculated for these sequences to obtain a 10x10 initial entropy value matrix, in which each element represents the environmental complexity at a specific spatiotemporal scale.

[0116] The entropy matrix is ​​converted into discretized entropy features through adaptive threshold segmentation (assuming the threshold is 0.5). The correlation analysis of character behaviors shows that the coupling index between "combat" behavior and the environment is 0.8, while the coupling index of "exploration" behavior is 0.6, indicating that the environment has a greater impact on combat behavior. Recursive pattern extraction found a periodic pattern of environmental complexity, such as an environmental complexity peak every 7 game days. After multi-dimensional feature fusion, a 50-dimensional comprehensive complexity vector is obtained, and each dimension represents a complexity feature of the environment. Finally, the comprehensive complexity vector is modeled through the LSTM network to obtain a scene complexity index that changes dynamically over time. This index shows different trends in different stages of the game. For example, during the peak period of the main task, the complexity index may increase significantly, reflecting that the environment has become more challenging. In scenes with high complexity, AIGC may generate more strategic and cautious behaviors, while in scenes with low complexity, more exploration and social behaviors are generated.

[0117] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0118] (1) Generate semantic embedding of the character behavior topology structure to obtain the behavior semantic vector, and perform context-aware mapping of the scene complexity index to obtain the environment semantic vector;

[0119] (2) Perform a tensor product operation on the behavior semantic vector and the environment semantic vector to obtain the initial interaction tensor;

[0120] (3) Perform sparsity constraint decomposition on the initial interaction tensor to obtain key interaction factors, and adaptively assign weights to the key interaction factors to obtain weighted interaction features;

[0121] (4) Dynamically time warping the weighted interaction features to obtain an aligned interaction sequence, and then conditional entropy calculation is performed on the aligned interaction sequence to obtain a mutual information matrix;

[0122] (5) Perform spectral clustering on the mutual information matrix to obtain behavior-environment clusters, and reconstruct the topological structure of the behavior-environment clusters to obtain the initial interaction graph;

[0123] (6) Perform evolutionary dynamics modeling on the initial interaction graph to obtain the behavior-environment interaction graph.

[0124] Specifically, in the AIGC-based character behavior prediction method, generative semantic embedding of the character behavior topology is a key step, which aims to transform the complex behavior structure into a computable vector representation. This process uses generative AI models, such as Transformer or BERT, to encode the behavior topology into a high-dimensional semantic vector. At the same time, the scene complexity index is context-aware mapped to obtain the environment semantic vector. This step takes into account the dynamic characteristics and contextual information of the environment.

[0125] Generative semantic embedding can be expressed by the following formula:

[0126] ;

[0127] in, is the behavior semantic vector, is the behavioral topology, is a generative embedding function, are model parameters.

[0128] The tensor product operation is performed on the behavior semantic vector and the environment semantic vector to obtain the initial interaction tensor. The tensor product operation can capture the complex interaction between behavior and environment. The initial interaction tensor is decomposed with sparsity constraints to obtain key interaction factors. This step uses techniques such as CANDECOMP / PARAFAC decomposition to extract the most important interaction patterns. The key interaction factors are adaptively weighted to obtain weighted interaction features. Adaptive weight assignment takes into account the importance of different interaction factors and can be achieved through the attention mechanism. Dynamic time warping (DTW) is performed on the weighted interaction features to obtain aligned interaction sequences. The DTW algorithm can handle the nonlinear alignment problem of time series, so that interaction sequences of different lengths can be compared.

[0129] The conditional entropy calculation is performed on the aligned interaction sequence to obtain the mutual information matrix. The conditional entropy calculation formula is as follows:

[0130] ;

[0131] in, is the mutual information, A and B represent the behavior and environment characteristics respectively, is the joint probability distribution, p(a) and p(b) are the marginal probability distributions.

[0132] The mutual information matrix is ​​spectrally clustered to obtain the behavior-environment clusters. Spectral clustering can cluster based on the spectral characteristics of the data and is suitable for processing complex nonlinear relationships. The behavior-environment clusters are reconstructed through topological structure to obtain the initial interaction graph. Finally, the initial interaction graph is modeled using evolutionary dynamics to obtain the behavior-environment interaction graph. The evolutionary dynamics model can describe the dynamic changes of the interaction relationship over time.

[0133] For example, in an AIGC-based multiplayer online role-playing game, the character behavior topology contains nodes such as "fight", "trade", and "exploration". Through generative semantic embedding, each behavior node is converted into a 256-dimensional vector. For example, the "fight" behavior may be encoded as [0.1, 0.3, ..., 0.2]. At the same time, scene complexity indicators (such as terrain complexity and NPC density) are also mapped into 256-dimensional environmental semantic vectors. The tensor product operation produces an initial 256x256 interaction tensor. After sparsity constrained decomposition, 10 key interaction factors are extracted. Adaptive weight assignment dynamically adjusts the importance of these factors according to the current game state. For example, in areas with frequent battles, the weight of factors related to battles may increase to 0.8, while in peaceful areas it may be only 0.2. Dynamic time warping aligns interaction sequences of different lengths to a uniform time step of 100. Conditional entropy calculation obtains a 10x10 mutual information matrix, in which each element represents the mutual information of a pair of behavior-environment features. Spectral clustering classifies these features into five main clusters, representing five typical behavior-environment interaction patterns.

[0134] The final behavior-environment interaction graph is a dynamically evolving network, with nodes representing behavior and environment features and edges representing the strength of interaction between them. This graph is constantly updated as the game progresses. For example, at the beginning of the game, the connection between the "exploration" and "resource collection" nodes may be strong (weight 0.9), while as the game progresses, the connection between the "combat" and "upgrade" nodes may become more important (weight increases from 0.3 to 0.8).

[0135] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0136] (1) Calculate the node importance of the behavior-environment interaction graph to obtain the target behavior node set, and construct a generative path for the target behavior node set to obtain the initial behavior path set;

[0137] (2) Normalize the path length of the initial behavior path set to obtain standardized path representation data;

[0138] (3) Calculate the energy value of the standardized path representation data using a preset path energy function to obtain the initial path energy value;

[0139] (4) Perform annealing temperature initialization processing on the initial path energy value to obtain the starting temperature parameter;

[0140] (5) Iteratively optimize the standardized path representation data through the simulated annealing algorithm to obtain the optimized behavior path, and analyze the environmental constraint indicators of the optimized behavior path to obtain the feasibility index;

[0141] (6) Adaptively screening the feasibility indicators with threshold values ​​to obtain a subset of candidate behavior paths, and then performing difference analysis on the subset of candidate behavior paths to obtain a path difference matrix;

[0142] (7) Perform cluster sampling based on the path difference matrix to obtain the target behavior path set.

[0143] Specifically, in the AIGC-based character behavior prediction method, node importance calculation of the behavior-environment interaction graph is a key step, aiming to identify the key nodes that have the greatest impact on the character behavior. This process uses centrality metrics in graph theory, such as eigenvector centrality or PageRank algorithm, to evaluate the importance of each node. The node importance calculation can be expressed by the following formula:

[0144] ;

[0145] in, is the importance score of node i, is the adjacency matrix element, is the importance of node j connected to node i.

[0146] After obtaining the target behavior node set, a generative path is constructed for it. This step uses AIGC technology, such as recurrent neural network (RNN) or Transformer model, to generate possible behavior paths. The generated initial behavior path set is then normalized to obtain standardized path representation data. The normalization process can be expressed as:

[0147] ;

[0148] in, is the normalized path length, is the original path length, and are the shortest and longest paths in the path set, respectively.

[0149] The energy value of the standardized path representation data is calculated by the preset path energy function to obtain the initial path energy value. The path energy function reflects the quality or fitness of the path, which may take into account factors such as path length and environmental adaptability. The initial path energy value is initialized with annealing temperature to obtain the starting temperature parameter, which is a key step in the simulated annealing algorithm.

[0150] The simulated annealing algorithm continuously adjusts the path to find the global optimal solution through an iterative optimization process. In each iteration, the algorithm randomly modifies the current path, such as swapping the positions of two nodes, and then calculates the energy value of the new path. If the new path is better, it is accepted; if it is worse, it is accepted with a certain probability, and this probability decreases as the temperature decreases. The temperature reduction formula can be expressed as:

[0151] ;

[0152] in, is the temperature at the kth iteration, is the cooling coefficient (0 < < 1).

[0153] The optimized behavior path is analyzed for environmental constraint indicators to obtain feasibility indicators. This step evaluates the feasibility of the path in the actual environment. The feasibility indicators are screened by adaptive thresholds to obtain a subset of candidate behavior paths, and then a difference analysis is performed on this subset to obtain a path difference matrix. Finally, cluster sampling is performed based on the path difference matrix to obtain the target behavior path set.

[0154] For example, in an open world role-playing game based on AIGC, the behavior-environment interaction graph contains nodes such as "exploration", "combat", "trading", and "dialogue". Through node importance calculation, it is found that the "exploration" node has the highest importance score of 0.85, followed by "combat" (0.72) and "trading" (0.63). Based on these important nodes, the generative path construction algorithm generates 100 initial behavior paths, such as "exploration-combat-trading-dialogue".

[0155] Path length normalization normalizes these paths to the range of 0-1. For example, if the shortest path length is 2 and the longest is 10, the normalized value of a path with a length of 6 is 0.5. The path energy function takes into account the length, diversity, and environmental adaptability of the path and calculates the energy value of each path. The initial temperature is set to 100 and the cooling coefficient is 0.95. The simulated annealing algorithm is iterated 1000 times, during which the temperature drops from 100 to about 5.18. The environmental constraint analysis of the optimized behavior path is performed, and the feasibility index is obtained by considering factors such as terrain difficulty and enemy distribution. The adaptive threshold is set to 0.7, and the paths with feasibility index greater than 0.7 are selected as candidate subsets. The candidate path subsets are analyzed for differences and the path difference matrix is ​​constructed. For example, the difference between the two paths "Explore-Fight-Trade" and "Explore-Dialogue-Trade" is 0.33 (only 1 / 3 of the nodes are different). Finally, the paths are clustered by the K-means clustering algorithm, and the representative path of each cluster is selected, and finally 10 representative and diverse target behavior paths are obtained.

[0156] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0157] (1) Perform sequence enhancement on the target behavior path set to obtain an enhanced behavior chain;

[0158] (2) Perform spatiotemporal coupling analysis on the reinforcement behavior chain to obtain a behavior interdependence grid;

[0159] (3) Extract the topological relationship of the behavior interdependence grid to obtain the original causal topology, and verify the temporal coordination of the original causal topology to obtain the optimized causal structure;

[0160] (4) Extract emergent features from the optimized causal structure to obtain multi-level causal association data, and perform interpretable transformation on the multi-level causal association to obtain target causal criterion data;

[0161] (5) Performing scenario adaptability analysis on the target causal criterion data to obtain environmental perception causal data, and performing behavioral random simulation analysis on the environmental perception causal data to obtain a behavioral possibility spectrum;

[0162] (6) The behavior possibility spectrum is cross-modally integrated to obtain the target behavior descriptor, and temporal feature extraction and probability distribution calculation are performed based on the target behavior descriptor to obtain the role behavior prediction data.

[0163] Specifically, in the AIGC-based character behavior prediction method, sequence enhancement of the target behavior path set is a key step, which aims to enrich the semantic information and contextual associations of the behavior sequence. This process uses generative AI models such as GPT or BERT to expand and refine the original behavior sequence to obtain an enhanced behavior chain. Sequence enhancement can be expressed by the following formula:

[0164] ;

[0165] in, is the reinforced behavior sequence, is the original behavior sequence, is the context information, and F is the sequence reinforcement function. The reinforcement behavior chain is subjected to spatiotemporal coupling analysis to obtain the behavior interdependence grid. This step uses spatiotemporal data mining techniques, such as spatiotemporal autocorrelation analysis or spatiotemporal covariance function, to capture the spatiotemporal dependencies between behaviors. The behavior interdependence grid reflects the degree of mutual influence between different behaviors in the temporal and spatial dimensions.

[0166] The topological relationship of the behavior interdependence grid is extracted to obtain the original causal topology. This process uses causal discovery algorithms, such as the PC algorithm or the FCI algorithm, to infer the causal relationship between behaviors. The original causal topology is verified by temporal coordination to ensure the consistency of the causal relationship in the time series and obtain the optimized causal structure. Temporal coordination verification can be achieved using Granger causality test or dynamic Bayesian network. The optimized causal structure is extracted through emergent features to obtain multi-level causal association data. Emergent feature extraction aims to discover high-order patterns and structures in complex systems, and complex network analysis or multi-scale entropy analysis methods can be used. The multi-level causal association data is subjected to interpretability transformation to obtain the target causal criterion data. The purpose of interpretability transformation is to transform complex causal relationships into rules or descriptions that can be understood by humans. The target causal criterion data is subjected to scenario adaptability analysis to obtain environmental perception causal data. This step considers the impact of different scenarios on causal relationships and is achieved using scenario modeling or conditional probability reasoning. The environmental perception causal data is subjected to behavioral random simulation analysis to obtain the behavioral possibility spectrum. Behavioral random simulation uses Monte Carlo method or Markov chain simulation to generate a large number of possible behavior sequences.

[0167] The behavior possibility spectrum is cross-modally integrated to obtain the target behavior descriptor. Cross-modal integration fuses behavioral features of different dimensions (such as time, space, semantics, etc.) into a unified representation. Finally, based on the target behavior descriptor, time series feature extraction and probability distribution calculation are performed to obtain the role behavior prediction data. Time series feature extraction can use the autoregressive integrated moving average (ARIMA) model or the long short-term memory (LSTM) network, while probability distribution calculation can be achieved through kernel density estimation or Gaussian mixture model.

[0168] For example, in a massively multiplayer online role-playing game based on AIGC, the original target behavior path set contains a simple sequence such as "exploration-battle-trade". After sequence enhancement, this sequence may be expanded to "carefully explore unknown areas-encounter and fight with hostile NPCs-trade equipment with merchants in nearby towns". The spatiotemporal coupling analysis shows that the "exploration" behavior is highly correlated with the "battle" behavior in space (correlation coefficient 0.8), while "battle" and "trade" show a certain lag relationship in time (average time interval 30 minutes).

[0169] Causal topology extraction found that "exploration" is the direct cause of "combat" (confidence 0.9), and "combat" leads to "trading" (confidence 0.7). Emergent feature extraction identified a high-order pattern: frequent "exploration-combat-trading" cycles lead to rapid character level increase. This pattern was converted into an interpretable rule: "High-risk exploration behavior accelerates character growth."

[0170] The scenario adaptability analysis shows that in high-difficulty areas, the transition probability from "exploration" to "combat" increases by 50%, while in safe areas, the duration of "trading" behavior is extended by an average of 20 minutes. The behavioral random simulation generates 1,000 possible behavior sequences, forming a behavioral probability spectrum. After cross-modal integration, a 300-dimensional target behavior descriptor is obtained, which contains temporal, spatial, and semantic features.

[0171] The final character behavior prediction shows that in the next hour, the character has a 70% chance of continuing to explore high-difficulty areas, a 20% chance of returning to safe areas for trading, and a 10% chance of teaming up with other players. This prediction is based on time series features (such as the behavior pattern in the past 3 hours) and the probability distribution of the current state (such as character health value, backpack capacity, etc.).

[0172] The above describes the role behavior prediction method based on AIGC in the embodiment of the present application. The following describes the role behavior prediction system based on AIGC in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the role behavior prediction system based on AIGC includes:

[0173] The spatiotemporal fractal module 201 is used to perform spatiotemporal fractal analysis on the pre-collected character behavior data to obtain a character behavior pattern spectrum;

[0174] A structure extraction module 202 is used to extract the structure of the role behavior pattern spectrum to obtain a role behavior topological structure;

[0175] The first analysis module 203 is used to perform multi-scale entropy analysis on the environmental information to obtain a scene complexity index;

[0176] An index calculation module 204 is used to perform co-evolutionary calculation on the character behavior topological structure and the scene complexity index to obtain a behavior-environment interaction map;

[0177] A path screening module 205 is used to analyze and screen the behavior paths of the behavior-environment interaction graph by using a simulated annealing algorithm to obtain a target behavior path set;

[0178] The second analysis module 206 is used to perform causal emergence analysis on the target behavior path set to obtain role behavior prediction results.

[0179] Through the synergy of the above components, by performing spatiotemporal fractal analysis on the character behavior data, this method can capture the complexity and self-similarity of the behavior pattern at different spatiotemporal scales, so as to more comprehensively understand the intrinsic structure and dynamic characteristics of the character behavior. This multi-scale analysis method helps to identify subtle behavior patterns that may be overlooked by conventional statistical methods, and provides a richer information basis for subsequent behavior prediction. The character behavior pattern spectrum is structurally extracted to obtain the character behavior topology. This step not only reveals the correlation between different behaviors, but also constructs a hierarchical behavior model, so that the system can better understand and predict complex behavior sequences. The extraction of this topological structure helps to identify key behavior nodes and behavior conversion paths. By performing multi-scale entropy analysis on environmental information, the scene complexity index is obtained, which can accurately quantify the complexity and uncertainty of the environment, which is crucial for understanding the impact of the environment on character behavior. This quantification of environmental complexity can dynamically adjust the behavior prediction strategy according to different environmental conditions, greatly improving the environmental adaptability of the prediction. The present application obtains the behavior-environment interaction map through co-evolutionary calculation, effectively simulating the dynamic interaction between character behavior and environment, so that the prediction model can take into account the immediate and long-term impact of environmental changes on character behavior. The introduction of this dynamic interaction model significantly improves the timeliness and accuracy of predictions, especially in rapidly changing environments. At the same time, the use of simulated annealing algorithm to analyze and screen the behavior path of the behavior-environment interaction graph not only optimizes the selection of behavior paths, but also improves the computational efficiency of predictions, and improves the efficiency and accuracy of AIGC-based role behavior predictions.

[0180] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the AIGC-based character behavior prediction method.

[0181] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0182] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A role behavior prediction method based on AIGC, characterized in that: The AIGC-based role behavior prediction method includes: Perform spatiotemporal fractal analysis on the pre-collected character behavior data to obtain the character behavior pattern spectrum; The character behavior pattern spectrum is subjected to structural extraction to obtain a character behavior topological structure, including: Generative semantic segmentation is performed on the character behavior pattern spectrum to obtain behavior pattern units, and context-aware encoding is performed on the behavior pattern units to obtain a coded behavior sequence; Temporal association analysis is performed on the coded behavior sequence to obtain a behavior transition sequence, and frequency statistics are performed on the behavior transition sequence to obtain a frequency matrix; Adaptive threshold filtering is performed on the frequency matrix to obtain a main behavior transition path, and graph-structured processing is performed on the main behavior transition path to obtain an initial behavior graph; Centrality calculation is performed on the initial behavior graph to obtain a key behavior node set, and connectivity analysis is performed on the key behavior node set to obtain a behavior subgraph set; Hierarchical clustering is performed on the behavior subgraph set to obtain a multi-level behavior structure, and topological relationship extraction is performed on the multi-level behavior structure to obtain a character behavior topological structure; Perform multi-scale entropy analysis on environmental information to obtain scene complexity indicators; A co-evolutionary calculation is performed on the character behavior topological structure and the scene complexity index to obtain a behavior-environment interaction graph, including: generative semantic embedding of the character behavior topological structure to obtain a behavior semantic vector, and context-aware mapping of the scene complexity index to obtain an environment semantic vector; a tensor product operation is performed on the behavior semantic vector and the environment semantic vector to obtain an initial interaction tensor; a sparsity constrained decomposition is performed on the initial interaction tensor to obtain key interaction factors, and adaptive weight assignment is performed on the key interaction factors to obtain weighted interaction features; dynamic time warping is performed on the weighted interaction features to obtain an aligned interaction sequence, and conditional entropy calculation is performed on the aligned interaction sequence to obtain a mutual information matrix; spectral clustering is performed on the mutual information matrix to obtain behavior-environment clusters, and topological structure reconstruction is performed on the behavior-environment clusters to obtain an initial interaction graph; evolutionary dynamics modeling is performed on the initial interaction graph to obtain a behavior-environment interaction graph; Performing behavior path analysis and screening on the behavior-environment interaction graph by using a simulated annealing algorithm to obtain a target behavior path set; A causal emergence analysis is performed on the target behavior path set to obtain a role behavior prediction result.

2. The AIGC-based character behavior prediction method according to claim 1, characterized in that: The step of performing spatiotemporal fractal analysis on the pre-collected character behavior data to obtain a character behavior pattern spectrum includes: Performing semantic annotation on the character behavior data to obtain a semantically enhanced behavior sequence, and performing temporal and spatial dimension decoupling on the semantically enhanced behavior sequence to obtain a time series and a spatial trajectory; Performing multi-scale decomposition on the time series to obtain a time scale feature set, and performing fractal dimension calculation on the spatial trajectory to obtain a spatial complexity index; The time scale feature set and the space complexity index are integrated to obtain a spatiotemporal feature matrix; Performing self-similarity analysis on the spatiotemporal feature matrix to obtain a fractal feature vector; Performing clustering processing on the fractal feature vector to obtain a behavior pattern cluster center; Performing time series correlation analysis on the behavior pattern cluster centers to obtain a behavior conversion probability matrix; Performing eigendecomposition on the behavior conversion probability matrix to obtain eigenvalues ​​of main behavior patterns; Performing spectrum transformation on the main behavior pattern characteristic values ​​to obtain the role behavior pattern spectrum.

3. The AIGC-based character behavior prediction method according to claim 1, characterized in that: The multi-scale entropy analysis of the environmental information is performed to obtain scene complexity indicators, including: Performing generative context encoding on the environmental information to obtain a semantically enhanced environmental representation, and performing spatiotemporal decomposition on the semantically enhanced environmental representation to obtain a multi-scale environmental feature sequence; Performing conditional sample entropy calculation on the multi-scale environmental feature sequence to obtain an initial entropy value matrix, and performing adaptive threshold segmentation on the initial entropy value matrix to obtain discretized entropy features; Performing role behavior correlation analysis on the discretized entropy features to obtain a behavior-environment coupling index, and performing recursive pattern extraction on the behavior-environment coupling index to obtain an environment complexity pattern; Performing multi-dimensional feature fusion on the environmental complexity pattern to obtain a target complexity vector; The target complexity vector is subjected to time series dynamic modeling to obtain the scene complexity index.

4. The AIGC-based character behavior prediction method according to claim 1, characterized in that: The behavior path analysis and screening of the behavior-environment interaction map by using a simulated annealing algorithm to obtain a target behavior path set includes: Calculating the importance of nodes on the behavior-environment interaction graph to obtain a target behavior node set, and constructing a generative path on the target behavior node set to obtain an initial behavior path set; Normalizing the path length of the initial behavior path set to obtain standardized path representation data; Calculating the energy value of the standardized path representation data using a preset path energy function to obtain an initial path energy value; Performing annealing temperature initialization processing on the initial path energy value to obtain a starting temperature parameter; Iteratively optimizing the standardized path representation data by using the simulated annealing algorithm to obtain an optimized behavior path, and performing environmental constraint index analysis on the optimized behavior path to obtain a feasibility index; Adaptively screening the feasibility index by threshold value to obtain a subset of candidate behavior paths, and performing difference analysis on the subset of candidate behavior paths to obtain a path difference matrix; Cluster sampling is performed based on the path difference matrix to obtain a target behavior path set.

5. The AIGC-based character behavior prediction method according to claim 4, characterized in that: The causal emergence analysis is performed on the target behavior path set to obtain the role behavior prediction result, including: Performing sequence enhancement on the target behavior path set to obtain an enhanced behavior chain; Performing spatiotemporal coupling analysis on the enhanced behavior chain to obtain a behavior interdependence grid; Extracting the topological relationship of the behavior interdependence grid to obtain the original causal topology, and performing temporal coordination verification on the original causal topology to obtain an optimized causal structure; Extracting emergent features from the optimized causal structure to obtain multi-level causal association data, and performing interpretability conversion on the multi-level causal association to obtain target causal criterion data; Performing a scene adaptability analysis on the target causal criterion data to obtain environmental perception causal data, and performing a behavior random simulation analysis on the environmental perception causal data to obtain a behavior possibility spectrum; The behavior possibility spectrum is cross-modally integrated to obtain a target behavior descriptor, and temporal feature extraction and probability distribution calculation are performed based on the target behavior descriptor to obtain role behavior prediction data.

6. A character behavior prediction system based on AIGC, characterized in that: Used to execute the AIGC-based role behavior prediction method according to any one of claims 1 to 5, the AIGC-based role behavior prediction system comprises: The spatiotemporal fractal module is used to perform spatiotemporal fractal analysis on the pre-collected character behavior data to obtain the character behavior pattern spectrum; A structure extraction module is used to perform structure extraction on the character behavior pattern spectrum to obtain a character behavior topological structure, including: performing generative semantic segmentation on the character behavior pattern spectrum to obtain behavior pattern units, and performing context-aware encoding on the behavior pattern units to obtain a coded behavior sequence; performing temporal association analysis on the coded behavior sequence to obtain a behavior conversion sequence, and performing frequency statistics on the behavior conversion sequence to obtain a frequency matrix; performing adaptive threshold filtering on the frequency matrix to obtain a main behavior conversion path, and performing graph structuring processing on the main behavior conversion path to obtain an initial behavior graph; performing centrality calculation on the initial behavior graph to obtain a key behavior node set, and performing connectivity analysis on the key behavior node set to obtain a behavior subgraph set; performing hierarchical clustering on the behavior subgraph set to obtain a multi-level behavior structure, and performing topological relationship extraction on the multi-level behavior structure to obtain a character behavior topological structure; The first analysis module is used to perform multi-scale entropy analysis on environmental information to obtain scene complexity indicators; An indicator calculation module is used to perform co-evolutionary calculation on the character behavior topological structure and the scene complexity index to obtain a behavior-environment interaction graph, including: performing generative semantic embedding on the character behavior topological structure to obtain a behavior semantic vector, and performing context-aware mapping on the scene complexity index to obtain an environment semantic vector; performing tensor product operation on the behavior semantic vector and the environment semantic vector to obtain an initial interaction tensor; performing sparsity constrained decomposition on the initial interaction tensor to obtain key interaction factors, and performing adaptive weight assignment on the key interaction factors to obtain weighted interaction features; performing dynamic time warping on the weighted interaction features to obtain an aligned interaction sequence, and performing conditional entropy calculation on the aligned interaction sequence to obtain a mutual information matrix; performing spectral clustering on the mutual information matrix to obtain behavior-environment clusters, and performing topological structure reconstruction on the behavior-environment clusters to obtain an initial interaction graph; performing evolutionary dynamics modeling on the initial interaction graph to obtain a behavior-environment interaction graph; A path screening module, used to analyze and screen the behavior path of the behavior-environment interaction graph through a simulated annealing algorithm to obtain a target behavior path set; The second analysis module is used to perform causal emergence analysis on the target behavior path set to obtain role behavior prediction results.

7. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the AIGC-based character behavior prediction method as described in any one of claims 1 to 5 is implemented.

Images