Game strategy evaluation and calibration method and device based on style perception

Abstract

The invention provides a game strategy evaluation calibration method and device based on style perception, and is applied to the technical field of artificial intelligence, the method comprises the following steps: obtaining evaluation feature vectors of a plurality of game strategies, the evaluation feature vectors comprising a winning rate and an ELO score extracted from a game evaluation system; unsupervised clustering is carried out based on the evaluation feature vectors, a plurality of style clusters are obtained, and each style cluster corresponds to a behavior mode of one game strategy; style conditional evaluation recalculation is carried out on each game strategy in each style cluster to obtain calibrated evaluation indexes, and the calibrated evaluation indexes comprise the percentile of the comprehensive score in the style cluster to which the comprehensive score belongs, the relative position of the ELO score in the style cluster to which the ELO score belongs and the relative position of the win rate in the style cluster to which the ELO score belongs; and generating a structured evaluation report of the target game strategy based on the calibrated evaluation index. According to the invention, the accuracy and the refinement degree of game strategy evaluation are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a game strategy evaluation calibration method and device based on style perception. BACKGROUND

[0002] In the field of multi-agent game decision-making, strategy evaluation is a key link of strategy training and optimization. Existing methods mainly include qualitative evaluation based on expert experience, quantitative statistics based on a single index, and ranking generation based on algorithm models.

[0003] However, in a complex game environment with diverse strategy styles and non-transitive significant, the existing evaluation system is difficult to accurately depict the real ability boundary and relative advantage of a specific style strategy due to the use of a global unified standard.

[0004] Therefore, the game strategy evaluation method in the related art has the technical problem of low accuracy. SUMMARY

[0005] The present application provides a game strategy evaluation calibration method and device based on style perception, which solves the defect of low accuracy of the game strategy evaluation method in the prior art and improves the accuracy and refinement of game strategy evaluation.

[0006] The present application provides a game strategy evaluation calibration method based on style perception, comprising the following steps. Obtain evaluation feature vectors of a plurality of game strategies, wherein the evaluation feature vectors include win rate and ELO score extracted from a game evaluation system; perform unsupervised clustering based on the evaluation feature vectors to obtain a plurality of style clusters, wherein each style cluster corresponds to a behavior pattern of a game strategy; perform style conditional evaluation recalibration on each game strategy in each style cluster to obtain calibrated evaluation indicators, wherein the calibrated evaluation indicators include the percentile of the comprehensive score in the corresponding style cluster, the relative position of the ELO score in the corresponding style cluster, and the relative position of the win rate in the corresponding style cluster; generate a structured evaluation report of a target game strategy based on the calibrated evaluation indicators, wherein the structured evaluation report includes style cluster identification and relative level information in the corresponding style cluster.

[0007] According to a style-aware game strategy evaluation and calibration method provided by the present invention, the step of obtaining evaluation feature vectors of multiple game strategies includes: extracting the original evaluation index of each game strategy from a preset game evaluation system, wherein the original evaluation index includes ability value, win rate and ELO score; obtaining environmental parameters of each game strategy, wherein the environmental parameters include map complexity, venue type and game format; fusing the original evaluation index with the environmental parameters to obtain an initial feature vector of each game strategy; and standardizing the initial feature vector to obtain the evaluation feature vector of each game strategy.

[0008] According to the present invention, a style-aware game strategy evaluation and calibration method is provided, wherein unsupervised clustering is performed based on the evaluation feature vector to obtain multiple style clusters, including: clustering the evaluation feature vector using the K-Means clustering algorithm and determining the optimal number of clusters using the elbow rule; dividing the multiple game strategies into multiple style clusters according to the optimal number of clusters, and assigning a style cluster identifier to each game strategy in the multiple game strategies.

[0009] According to a style-aware game strategy evaluation calibration method provided by the present invention, the method involves recalculating the style-conditional evaluation of each game strategy within each style cluster to obtain a calibrated evaluation index. The method includes: for each style cluster, performing the following processes: determining the overall score distribution, ELO score distribution, and win rate distribution of all game strategies within the style cluster; determining the percentile of the overall score for each game strategy within the style cluster based on the overall score distribution; determining the relative position of the ELO score for each game strategy within the style cluster based on the ELO score distribution; determining the relative position of the win rate for each game strategy within the style cluster based on the win rate distribution; and determining the calibrated evaluation index for each game strategy based on the percentile of the overall score, the relative position of the ELO score, and the relative position of the win rate.

[0010] According to the present invention, a style-aware game strategy evaluation calibration method is provided, wherein generating a structured evaluation report of the target game strategy based on the calibrated evaluation index includes: inputting the evaluation feature vector of the target game strategy into a pre-trained unsupervised clustering model to obtain a predicted style cluster of the target game strategy; determining the comprehensive score distribution, ELO score distribution, and win rate distribution of all game strategies corresponding to the predicted style cluster; comparing and analyzing the evaluation feature vector of the target game strategy with the comprehensive score distribution, the ELO score distribution, and the win rate distribution to obtain the calibrated evaluation index of the target game strategy; and generating a structured evaluation report of the target game strategy based on the calibrated evaluation index of the target game strategy.

[0011] According to the present invention, a style-aware game strategy evaluation and calibration method further includes: obtaining calibrated evaluation indices for all game strategies and their corresponding style cluster identifiers; performing dimensionality reduction processing on the calibrated evaluation indices for all game strategies to obtain low-dimensional spatial coordinates; visualizing and rendering the low-dimensional spatial coordinates to generate a game strategy capability spatial distribution map, wherein game strategies with different style cluster identifiers in the game strategy capability distribution map are distinguished by different colors or shapes; and performing capability boundary analysis based on the strategy capability spatial distribution map to obtain the clustering region and boundary features of each style cluster.

[0012] This invention also provides a style-aware game strategy evaluation and calibration device, comprising the following modules: an acquisition module for acquiring evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include win rate and ELO score extracted from the game evaluation system; a clustering module for performing unsupervised clustering based on the evaluation feature vectors to obtain multiple style clusters, wherein each style cluster corresponds to a behavioral pattern of a game strategy; a re-evaluation module for recalculating the style-conditional evaluation of each game strategy within each style cluster to obtain calibrated evaluation indicators, wherein the calibrated evaluation indicators include the percentile of the comprehensive score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster; and a generation module for generating a structured evaluation report of the target game strategy based on the calibrated evaluation indicators, wherein the structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the style-aware game strategy evaluation and calibration method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the style-aware game strategy evaluation and calibration method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the style-aware game strategy evaluation and calibration method as described above.

[0016] The present invention provides a style-aware game strategy evaluation and calibration method and apparatus. First, it acquires evaluation feature vectors (including win rate and ELO score) for multiple game strategies, providing a standardized data foundation for subsequent analysis and ensuring the comprehensiveness and consistency of the evaluation criteria. Next, it performs unsupervised clustering based on these feature vectors to automatically identify different style clusters, thereby objectively distinguishing the behavioral patterns of game strategies and avoiding subjective classification bias. Then, it recalculates the style-conditional evaluation of game strategies within each style cluster, obtaining calibrated evaluation indicators (such as the relative positions of the overall score percentile, ELO score, and win rate), effectively eliminating the interference of style differences on the evaluation, achieving fair comparison under the same behavioral pattern, and improving the accuracy and refinement of the evaluation. Finally, it generates a structured evaluation report for the target game strategy based on the calibration indicators, including style cluster identifiers and relative level information within the style cluster, making the evaluation results clearer and more interpretable, facilitating users' quick understanding of the performance positioning and style characteristics of the game strategy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced one by one below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the style-aware game strategy evaluation and calibration method provided by the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the use of the adversarial game strategy evaluation method and system based on unsupervised clustering provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the adversarial game strategy evaluation method and system construction based on unsupervised clustering provided by the present invention.

[0021] Figure 4 This is a distribution diagram of the strategy capability space analysis results provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the module of the style-aware game strategy evaluation and calibration device provided by the present invention.

[0023] Figure 6 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] Game strategy evaluation and calibration methods in related technologies mainly include the following three categories: Expert qualitative evaluation: This method relies on manual observation for assessment. It is highly reliable, but highly subjective and difficult to scale up.

[0026] Single-indicator statistics (such as win rate): These rely on data statistics to uncover patterns, but they are one-sided and cannot reflect comprehensive capabilities.

[0027] Global strategy ranking model: It can handle large-scale games, but the output is a global scalar score, which implicitly assumes that "all strategies are comparable".

[0028] Although existing methods have some effectiveness in their respective applicable scenarios, they share a fundamental flaw: in game environments with highly heterogeneous strategy styles, they force all strategies to be compared in the same evaluation space.

[0029] In other words, existing technologies lack style perception capabilities and cannot perform fair and accurate relative evaluations within subspaces with consistent behavioral patterns.

[0030] This invention aims to address a key technical problem in existing game strategy evaluation systems: in adversarial environments with highly heterogeneous strategy styles, the use of a globally uniform evaluation standard leads to distorted evaluation results for strategies of specific styles, manifesting as blurred capability boundaries and overall scores deviating from their true relative levels among similar strategies. Specifically, existing evaluation methods (such as win rate statistics and ELO methods) place all strategies in the same evaluation space for ranking or scoring.

[0031] Specifically, existing evaluation systems include: adversarial win rate, Bayesian ELO score, strategic capability boundary, and comprehensive score. However, in scenarios with diverse strategic styles (such as the coexistence of fast-paced and operational strategies), the criteria for judging the merits of different styles of strategies differ fundamentally. When a globally unified standard is used for evaluation, it can lead to: some strategies performing well among their peers being systematically underestimated by the global scoring system, resulting in distorted comprehensive scores; heatmaps based on win rate or resource allocation becoming blurred due to the mixed styles, failing to accurately reflect the true critical points of specific style strategies; and evaluation results failing to support targeted strategy optimization.

[0032] To address the aforementioned technical problems, this invention proposes a style-aware evaluation calibration method. This method, as an enhancement to existing game evaluation systems, does not introduce a new scoring model. Instead, it utilizes unsupervised clustering to group strategies and recalculates the original evaluation indicators using style-conditionalization.

[0033] Figure 1 This is a flowchart illustrating the style-aware game strategy evaluation and calibration method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps.

[0034] Step 101: Obtain the evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include the win rate and ELO score extracted from the game evaluation system.

[0035] In this embodiment of the invention, raw data constituting the evaluation feature vector is collected from an existing game evaluation system. This mainly includes: Win Rate (W): the output of the adversarial win rate module, reflecting the overall winning probability of the game strategy. ELO Score (R): the output of the Bayesian ELO module, used to measure the relative strength level of the game strategy. Ability Value (C): input data from the comprehensive distribution module, serving as a basic value for evaluating the overall ability of the strategy. Environmental Parameters: environmental context information that significantly affects strategy performance is obtained from the adversarial experiment configuration, such as map complexity (M), venue type (T), game format (F), and initial resource quantity.

[0036] The heterogeneous data collected above are integrated to construct a multi-dimensional initial feature vector for each game strategy. This vector centrally represents the objective performance of the strategy and its environment.

[0037] The initial feature vectors are standardized and preprocessed to obtain the evaluation feature vectors. For example, this process uses the StandardScaler tool in a machine learning library (such as Scikit-learn) to perform Z-score normalization on all feature dimensions, that is, subtracting the mean and dividing by the standard deviation, so that the mean of each feature data after processing is 0 and the variance is 1.

[0038] This solves the problem of large differences in the dimensions and numerical ranges of different evaluation metrics (such as win rate and ELO score), and avoids the situation where certain features dominate the model results due to excessively large values ​​in subsequent clustering analysis, thus ensuring that the unsupervised clustering algorithm can fairly consider all feature dimensions.

[0039] Step 102: Perform unsupervised clustering based on the evaluation feature vectors to obtain multiple style clusters, where each style cluster corresponds to a behavioral pattern of a game strategy.

[0040] In this embodiment of the invention, an unsupervised clustering algorithm is applied to the standardized evaluation feature vector to automatically divide the game strategy into multiple style clusters.

[0041] For example, the K-Means algorithm (implemented through the KMeans class in Scikit-learn) can be used, with parameters such as n_clusters=5 (determining the optimal number of clusters using the elbow rule), init='k-means++', and n_init=10 to improve clustering stability and efficiency.

[0042] The clustering process requires no manual intervention. Each style cluster corresponds to a data-driven behavioral pattern (such as fast-paced or operational), and cluster identifiers (such as 0, 1, 2, ...) serve as identifiers. This effectively identifies the heterogeneity of strategy styles and decomposes the global evaluation space into homogeneous subsets.

[0043] Step 103: Perform style-conditional evaluation recalculation for each game strategy within each style cluster to obtain the calibrated evaluation index.

[0044] The calibrated evaluation metrics include the percentile of the composite score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster.

[0045] In this embodiment of the invention, after clustering is completed, the game strategies are grouped according to the cluster identifier of the style cluster, and style conditional recalculation is performed on the strategies within each style cluster.

[0046] For example, the percentile of the overall score S of each game strategy within its style cluster is calculated as the "relative level within the style". For example, if a strategy is at the 88th percentile within the cluster, it means that it is better than 88% of the strategies in the same style.

[0047] The distribution of ELO scores (R) and win rates (W) within a cluster is statistically analyzed, and the relative position of each strategy in these metrics (such as percentile or comparison with the cluster mean) is calculated. This recalculation method calibrates the evaluation feature vector, eliminates distortion caused by style differences, and makes the evaluation results more focused on the true performance of similar strategies.

[0048] Step 104: Based on the calibrated evaluation indicators, generate a structured evaluation report of the target game strategy. The structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

[0049] In this embodiment of the invention, a structured evaluation report is generated for the target game strategy (such as a newly added game strategy).

[0050] For example, after extracting the evaluation feature vector of the target game strategy, a pre-trained clustering model is used to predict its style cluster identifier (e.g., cluster 3). Then, the recalculation results of the corresponding cluster are called to generate a structured evaluation report in JSON format. The structured evaluation report includes the style cluster identifier, the percentile of the comprehensive score S within the cluster, the ELO score R, and the relative position of the win rate W within the cluster (including mean and percentile). In addition, visualization output is supported, such as drawing a policy capability space distribution map through dimensionality reduction, which intuitively shows the separation of style clusters and capability boundaries.

[0051] This invention first obtains evaluation feature vectors (including win rate and ELO score) for multiple game strategies, providing a standardized data foundation for subsequent analysis and ensuring the comprehensiveness and consistency of the evaluation criteria. Next, unsupervised clustering is performed based on these feature vectors to automatically identify different style clusters, thereby objectively distinguishing the behavioral patterns of game strategies and avoiding subjective classification bias. Then, style-conditional evaluation recalculation is performed on game strategies within each style cluster to obtain calibrated evaluation indicators (such as the relative positions of the overall score percentile, ELO score, and win rate), effectively eliminating the interference of style differences on the evaluation, achieving fair comparison under the same behavioral pattern, and improving the accuracy and refinement of the evaluation. Finally, a structured evaluation report of the target game strategy is generated based on the calibrated indicators, including style cluster identifiers and relative level information within the style cluster, making the evaluation results clearer and more interpretable, facilitating users' quick understanding of the performance positioning and style characteristics of the game strategy.

[0052] According to the style-aware game strategy evaluation calibration method provided by the present invention, the evaluation feature vectors of multiple game strategies are obtained, including: The original evaluation indicators for each game strategy are extracted from multiple game strategies in a pre-defined game evaluation system. The original evaluation indicators include ability value, win rate and ELO score. Obtain the environmental parameters for each game strategy, including map complexity, venue type, and game format; The original evaluation indicators are fused with environmental parameters to obtain the initial feature vector for each game strategy; The initial feature vectors are standardized to obtain the evaluation feature vectors for each game strategy.

[0053] In this embodiment of the invention, the original evaluation indicators for each game strategy are first collected from an existing game evaluation system. These original evaluation indicators are the core outputs of the pre-defined game evaluation system and mainly include: the capability value (C) obtained from the comprehensive distribution module, the win rate (W) obtained from the adversarial win rate module, and the ELO score (R) obtained from the Bayesian ELO module. These indicators quantify the adversarial capability of the game strategy from different dimensions.

[0054] Simultaneously, the system acquires the environmental parameters of each game strategy during the evaluation process. These environmental parameters define the context conditions for strategy execution, specifically including map complexity (M), site type (T), and game format (F). These parameters are derived from the configuration information of the adversarial experiment.

[0055] After data collection, the original evaluation indicators are combined with environmental parameters to construct a multi-dimensional initial feature vector for each game strategy. This initial feature vector correlates the inherent performance indicators of the game strategy (i.e., the original evaluation indicators) with the environmental information (i.e., the environmental parameters), forming a comprehensive feature representation.

[0056] To ensure the comparability of initial feature vectors of different dimensions and to eliminate the influence of dimensions, the initial feature vectors need to be standardized.

[0057] For example, the StandardScaler from Python's Scikit-learn library can be used to perform Z-score normalization on the initial feature vector, making its mean 0 and standard deviation 1, thus obtaining a standardized evaluation feature vector that can be used for model computation.

[0058] Through the embodiments of the present invention, the original evaluation indicators and environmental parameters are integrated to construct a standardized evaluation feature vector, which effectively improves the accuracy and adaptability of game strategy evaluation and achieves precise calibration under style perception.

[0059] According to the present invention, a style-aware game strategy evaluation and calibration method is provided, which performs unsupervised clustering based on evaluation feature vectors to obtain multiple style clusters, including: The K-Means clustering algorithm is used to cluster the evaluation feature vectors, and the optimal number of clusters is determined by the elbow rule. Based on the optimal number of clusters, multiple game strategies are divided into multiple style clusters, and a style cluster identifier is assigned to each game strategy in the multiple game strategies.

[0060] In this embodiment of the invention, the standardized evaluation feature vectors (sets) are grouped using an unsupervised clustering algorithm (such as K-Means) to automatically identify style clusters of several game strategies. Each style cluster represents a data-driven behavioral pattern, without the need for manually defined semantic labels or preset rules. The clustering results are identified by cluster identifiers (such as 0, 1, 2, ...), which serve as the basis for subsequent grouping and recalculation.

[0061] First, it is necessary to determine the optimal number of clusters for unsupervised clustering. This invention employs the K-Means clustering algorithm and utilizes the elbow rule to determine the optimal number of clusters K. The elbow rule makes a judgment by calculating the sum of squares within each cluster corresponding to different K values ​​and observing the inflection point of its change curve (i.e., the "elbow"). In practice, as the K value gradually increases from 1, the sum of squares within each cluster decreases significantly; when the K value increases to the actual number of clusters, the decrease tends to level off, forming an inflection point similar to an elbow. The K value corresponding to this point is the optimal number of clusters.

[0062] For example, after determining the optimal number of clusters K=5, the cluster analysis is then performed. The KMeans class from Python's Scikit-learn machine learning library is used. Specific parameters are set as follows: n_clusters=5, specifying that the data will be divided into 5 style clusters; init='k-means++', employing a smart initial centroid selection method to accelerate convergence and improve result stability; n_init=10, indicating that the algorithm will run 10 times with different initial centroids and select the best result as the final model. The standardized evaluation feature vector dataset is then input into this configured K-Means model for fitting.

[0063] After model training, each game strategy is assigned to the group of its nearest cluster center based on the position of its evaluation feature vector in the feature space. Ultimately, multiple game strategies are automatically divided into multiple style clusters (five in this implementation). Each game strategy is assigned a unique style cluster identifier, such as the numbers 0, 1, 2, 3, and 4. Each style cluster represents a common game strategy behavior pattern discovered through data-driven analysis.

[0064] Through the embodiments of the present invention, the optimal number of clusters is automatically determined by K-Means clustering and the elbow rule, realizing unsupervised style classification of game strategies, accurately assigning style cluster labels to each strategy, and significantly improving the objectivity and personalized calibration capability of game strategy evaluation.

[0065] According to the style-aware game strategy evaluation calibration method provided by the present invention, style-conditional evaluation recalculation is performed on each game strategy within each style cluster to obtain calibrated evaluation indices, including: For each style family, perform the following procedure: Determine the overall score distribution, ELO score distribution, and win rate distribution of all game strategies within the style cluster; Based on the overall score distribution, determine the percentile of the overall score for each game strategy within the style cluster; Based on the ELO score distribution, determine the relative position of the ELO score of each game strategy within the style cluster; Based on the win rate distribution, determine the relative position of the win rate of each game strategy within the style cluster; Based on the percentile of the overall score, the relative position of the ELO score, and the relative position of the win rate, a calibrated evaluation index is determined for each game strategy.

[0066] In this embodiment of the invention, the percentile of the overall score S of the game strategy within its cluster is calculated as the "relative level within the style"; the ELO and win rate distributions within the style cluster are statistically analyzed, and the relative position of the game strategy within them is calculated.

[0067] Based on the obtained style cluster identifiers (e.g., 0, 1, 2, 3, 4), all game strategies are divided into corresponding subsets. For each style cluster, an independent statistical analysis process is performed to determine the baseline distribution of each evaluation index within that style cluster.

[0068] For the style cluster being processed, calculate the statistical distribution of the original evaluation metrics for all game strategies within it: Overall score (S) distribution: Collect the overall score S of each game strategy within the cluster, calculated by the existing game evaluation system, to form a dataset.

[0069] ELO score (R) distribution: Collect the Bayesian ELO score R for each game strategy within the cluster to form a dataset.

[0070] Win rate (W) distribution: Collect the win rate W of each game strategy within the cluster to form a dataset.

[0071] These distributions describe the overall situation of each evaluation indicator within the current style cluster, providing a reference for subsequent calculations of relative levels.

[0072] After obtaining the above distribution, the system performs the following calculations for each game strategy within this style cluster to obtain the calibrated evaluation index: Calculate the percentile of the overall score (S): Based on the overall score distribution of the cluster, calculate the percentile of the overall score S of the current game strategy within the cluster. For example, if a strategy's S value ranks in the top 12% within the cluster, then its percentile is 88%.

[0073] Determine the relative position of the ELO score (R): Similarly, based on the intra-cluster ELO score distribution, determine the relative position of the current strategy's ELO score R within it. For example, its percentile can be calculated, or its difference from the intra-cluster mean can be recorded to illustrate the strategy's ranking among its peers.

[0074] Determine the relative position of the win rate (W): Based on the win rate distribution within the cluster, determine the relative position of the current strategy's win rate W within it, calculated in the same way as the ELO score.

[0075] The evaluation index after the strategy is calibrated consists of three pieces of information: the percentile of the overall score calculated for a single game strategy, the relative position of the ELO score, and the relative position of the win rate.

[0076] The calibrated evaluation metrics are no longer global scores that are compared across all types of strategies, but rather fair evaluation results focused on the style cluster to which the strategy belongs, eliminating the interference of style differences. For example, a complete calibrated metric might be recorded as: {"Overall S percentile": 88, "ELO R percentile": 82, "Win Percentile": 75}.

[0077] Through the embodiments of the present invention, evaluation indicators are calculated by relative distribution within style clusters, eliminating global evaluation bias, accurately reflecting the true performance of the strategy in a specific style environment, and greatly improving the relevance and reliability of the evaluation.

[0078] According to the present invention, a style-aware game strategy evaluation calibration method generates a structured evaluation report of the target game strategy based on the calibrated evaluation indicators, including: The evaluation feature vector of the objective game strategy is input into a pre-trained unsupervised clustering model to obtain the predicted style cluster of the objective game strategy. Determine the overall score distribution, ELO score distribution, and win rate distribution of all game strategies corresponding to the prediction style cluster; By comparing and analyzing the evaluation feature vector of the objective game strategy with the comprehensive score distribution, ELO score distribution and win rate distribution, the calibrated evaluation index of the objective game strategy is obtained. Based on the calibrated evaluation indicators of the objective game strategy, a structured evaluation report of the objective game strategy is generated.

[0079] In this embodiment of the invention, the first step is to determine the style classification of the target game strategy. The evaluation feature vector of the target game strategy (including its ability value C, win rate W, ELO score R, and related environmental parameters) is input into an unsupervised clustering model (such as the K-Means model) that has been trained using historical game strategy data. This pre-trained model quickly predicts and outputs the style cluster identifier to which the target game strategy belongs based on the position of the feature vector in the feature space; for example, the prediction result is "cluster 3". This step achieves rapid and automated style classification of new strategies.

[0080] After identifying the predictive style cluster of the target game strategy, the evaluation data of all historical game strategies corresponding to the cluster identifier (e.g., "Cluster 3") of that predictive style cluster are retrieved. This data constitutes the baseline distribution within that style cluster, primarily including: the overall score (S) distribution, ELO score (R) distribution, and win rate (W) distribution of all game strategies belonging to that cluster. Obtaining the statistical information of these distributions (such as mean, standard deviation, quantiles, etc.) establishes a reference system for subsequent precise comparisons.

[0081] The original evaluation metrics of the target game strategy (i.e., its specific values ​​of S, R, and W) are precisely compared with the baseline distribution of its style cluster retrieved in the previous step. The relative position of each metric of the target game strategy within the cluster distribution is calculated, usually presented in percentile form.

[0082] For example, the calculated overall score S of the target game strategy is at the 88th percentile in "Cluster 3," meaning its overall score is better than 88% of the strategies in that cluster. Similarly, its ELO score R is calculated to be at the 82nd percentile, and its win rate W is at the 75th percentile. These percentile values ​​together constitute the calibrated evaluation metrics for the target game strategy.

[0083] Finally, based on the calibrated evaluation metrics of the target game strategy, a structured evaluation report of the target game strategy is automatically generated. This structured evaluation report is organized using a machine-readable and user-friendly structured format (such as JSON) to ensure the standardization and parsability of information. The report not only includes the basic identifier of the target game strategy and its predicted style cluster affiliation, but more importantly, it provides a detailed presentation of the evaluation results of its various metrics after calibration, i.e., its relative level among similar style strategies.

[0084] Through the embodiments of the present invention, by mapping the target strategy to the predicted style cluster and performing intra-cluster comparative analysis, an accurate and interpretable structured evaluation report is generated, which effectively improves the pertinence and practical value of the evaluation results.

[0085] According to the style-aware game strategy evaluation and calibration method provided by the present invention, the method further includes: Obtain the calibrated evaluation metrics for all game strategies and their corresponding style cluster identifiers; Dimensionality reduction is performed based on the calibrated evaluation metrics of all game strategies to obtain low-dimensional space coordinates. Visualize and render low-dimensional spatial coordinates to generate a spatial distribution map of game strategy capabilities. In the game strategy capability distribution map, game strategies identified by different style clusters are distinguished by different colors or shapes. Capability boundary analysis is performed based on the spatial distribution map of strategy capabilities to obtain the clustering regions and boundary characteristics of each style cluster.

[0086] In this embodiment of the invention, when evaluating the target game strategy, its evaluation feature vector (C, W, R, environmental parameters) is extracted; the style cluster identifier to which it belongs is predicted through a pre-trained clustering model; the recalculation result of the corresponding style cluster is called to generate a structured evaluation report containing the following: style cluster identifier; percentile of the comprehensive score S in the same category; relative position of ELO and win rate in the same category. Global heatmaps are supported for easy observation of the situation of each category set.

[0087] For example, collect data on all game strategies that have been evaluated and calibrated, including the calibrated evaluation metrics for each game strategy (such as percentile of overall score, relative position of ELO score, etc.) and their corresponding style cluster identifiers.

[0088] The collected calibrated evaluation metrics are subjected to dimensionality reduction. Since the original feature vectors and calibrated evaluation metrics have high dimensionality, they cannot be directly visualized in two or three dimensions. In practice, dimensionality reduction algorithms such as principal component analysis are used to map the calibrated evaluation metrics of each game strategy to a two-dimensional plane coordinate system, thus calculating the low-dimensional spatial coordinates of each game strategy in the visualization space.

[0089] After obtaining the low-dimensional coordinates of all game strategies, the visualization and rendering stage begins. Using visualization libraries in Python such as Matplotlib, Seaborn, or Plotly, each game strategy is plotted on a two-dimensional plane based on its coordinates. The key is that, during rendering, each game strategy is assigned a different color or shape marker based on its style cluster identifier. For example, strategies belonging to "style cluster 1" are represented by red dots, while strategies belonging to "style cluster 2" are represented by blue triangles. In this way, a clear power space distribution map of the game strategies is generated.

[0090] To verify the effectiveness of the present invention, a comparative experiment was conducted in an intelligent game simulation environment. Based on 1,000 AI strategy game data, three experts familiar with the environment were invited to evaluate 100 representative strategies, and the results were used as the evaluation benchmark.

[0091] Table 1 below shows the comparison of results.

[0092] Table 1. Comprehensive comparison of the effectiveness of this invention with various existing evaluation methods: Evaluation method Method of the present invention In-style ranking result rationality expert scoring Style recognition accuracy Whether to support quantitative analysis Evaluation dimension and processing capacity Expert evaluation Expert qualitative evaluation / / No Extremely limited dimension, only limited samples can be processed Win rate Single index statistics 68% Not supported Yes, support visualization 1 dimension, large-scale processing Elo Global strategy ranking 75% Not supported Yes, support visualization 1 dimension, large-scale processing The invention / 93% 95% Yes, support visualization Several dimensions can be achieved, large-scale processing Experimental results show that the present invention is significantly superior to existing methods in terms of evaluation rationality and style recognition capabilities, while maintaining large-scale automated processing capabilities.

[0093] The following describes an example of the practical application of the style-aware game strategy evaluation and calibration method provided by this invention.

[0094] refer to Figure 2 , Figure 2 This is a schematic diagram of the adversarial game strategy evaluation method and system based on unsupervised clustering provided by the present invention, which includes: data input, GPU processing core and dynamic evaluation output.

[0095] refer to Figure 3 , Figure 3 This is a schematic diagram of the adversarial game strategy evaluation method and system construction based on unsupervised clustering provided by the present invention, which includes: data acquisition, feature construction, pattern discovery, evaluation calibration, and result report generation.

[0096] The system proposed in this invention can be deployed on top of existing game evaluation systems as a calibration and enhancement module. Its system architecture mainly includes the following functional modules: Data acquisition module: Obtains calculated assessment indicators and environmental parameters from existing assessment systems.

[0097] Feature construction module: Constructs feature vectors from existing evaluation metrics (C, W, R) and environmental parameters.

[0098] Pattern discovery module: The core is an unsupervised clustering algorithm, which is responsible for the automatic identification of policy styles.

[0099] Evaluation and calibration module: responsible for recalculating the style conditions of existing evaluation indicators.

[0100] Report generation module: Responsible for generating structured evaluation reports that include relative levels within styles.

[0101] Hardware requirements: A computing platform equipped with an NVIDIA RTX 4090 graphics card and CUDA 12.2 parallel architecture, featuring 24GB of video memory and at least 64GB of system memory. This configuration enables efficient parallel computing during the training phase and supports high throughput during the evaluation phase.

[0102] Software environment: Operating system: Ubuntu 20.04LTS or later.

[0103] Core programming language: Python 3.9.

[0104] Data processing: Pandas, NumPy.

[0105] Machine learning: Scikit-learn (for clustering and data preprocessing).

[0106] The specific implementation method is as follows: Step 1: Specific implementation of feature vector construction.

[0107] Data source: The following data were obtained from existing evaluation systems: Capability value C: Input from the integrated distribution module; Win rate W: Output from the match win rate module; ELO score R: Output from the Bayesian ELO module; Environmental parameters: including map complexity M, venue type T, game format F, etc. (from the adversarial experiment configuration).

[0108] Data preprocessing: StandardScaler from Scikit-learn was used to standardize all feature vectors so that their mean is 0 and their variance is 1.

[0109] Step 2: Specific implementation of unsupervised style clustering.

[0110] Clustering: The K-Means algorithm is used to cluster the preprocessed feature vectors. The optimal number of clusters, K=5, can be determined by methods such as the elbow rule. Using the KMeans class from the Scikit-learn library, with n_clusters=5, init='k-means++', and n_init=10, the entire dataset is fitted to obtain the centroids of 5 clusters and the cluster ID of each sample.

[0111] Step 3: Specific implementation of style conditional evaluation recalculation.

[0112] Data grouping: The entire dataset is divided into 5 subsets based on the cluster ID of each sample.

[0113] Style-conditional recalculation: Perform the following operations for each subset: Reinterpretation of the composite score: The composite score S of the strategy is calculated as the percentile within its cluster, and is used as the "relative level within the style"; ELO vs. Win Rate Comparison: Statistically analyze the distribution of ELO and win rate within a cluster, and calculate the relative position of the strategy within it.

[0114] Step 4: Specific implementation of style adaptive evaluation and structured output.

[0115] Feature extraction: The "new strategy_V1" to be evaluated is tested in a game environment, and the system automatically obtains its ability value C, win rate W, ELO score R, and environmental parameters.

[0116] Style recognition: The feature vector is input into a pre-trained K-Means model, which predicts that it belongs to cluster 3.

[0117] Style conditional assessment: Overall score S: The percentile of S in cluster 3 is 88% (top 12%). ELO and Win Rate Comparison: Displays the relative position of the strategy's ELO score and win rate in cluster 3.

[0118] Report Generation: The system generates an evaluation report in JSON format, as shown below: { "Strategy Identifier": "New Strategy_V1", "Style Cluster ID": "3", "Overall score S percentile": 88, "Comparison within the same style": { "Overall Score S": {"Mean": 0.68, "Percentile": 88} ELO rating (R): {Mean: 1150, Percentile: 82} "Win rate W": {"Mean": 0.62,"Percentile": 75} } } Further dimensionality reduction and visualization can be performed to create a policy space distribution map, which clearly shows that this method can more comprehensively and objectively characterize the capability boundary and effectively solve the problem of evaluation distortion caused by style mixing.

[0119] refer to Figure 4 , Figure 4 This is a distribution diagram of the strategy capability space analysis results provided by the present invention.

[0120] The style-aware game strategy evaluation and calibration device provided by the present invention will be described below. The style-aware game strategy evaluation and calibration device described below can be referred to in correspondence with the style-aware game strategy evaluation and calibration method described above.

[0121] refer to Figure 5 , Figure 5 This is a schematic diagram of the module of the style-aware game strategy evaluation and calibration device provided by the present invention.

[0122] The acquisition module 501 is used to acquire the evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include the win rate and ELO score extracted from the game evaluation system. Clustering module 502 is used to perform unsupervised clustering based on the evaluation feature vector to obtain multiple style clusters, where each style cluster corresponds to a behavioral pattern of a game strategy. The re-evaluation module 503 is used to perform style-conditional evaluation recalculation on each game strategy within each style cluster to obtain calibrated evaluation indicators. The calibrated evaluation indicators include the percentile of the comprehensive score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster. The generation module 504 is used to generate a structured evaluation report of the target game strategy based on the calibrated evaluation index, wherein the structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

[0123] Specifically, the style-aware game strategy evaluation and calibration device provided by the present invention can implement all the method steps implemented in the above-described style-aware game strategy evaluation and calibration method embodiments, and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiments and the beneficial effects will not be described in detail.

[0124] Figure 6 This is a schematic diagram of the physical structure of the electronic device provided by the present invention, such as... Figure 6As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communications bus 640. The processor 610 can call logic instructions in the memory 630 to execute a style-aware game strategy evaluation calibration method. This method includes: acquiring evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include win rates and ELO scores extracted from a game evaluation system; performing unsupervised clustering based on the evaluation feature vectors to obtain multiple style clusters, wherein each style cluster corresponds to a behavioral pattern of a game strategy; performing style-conditional evaluation recalculation on each game strategy within each style cluster to obtain calibrated evaluation indicators, wherein the calibrated evaluation indicators include the percentile of the overall score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster; and generating a structured evaluation report of the target game strategy based on the calibrated evaluation indicators, wherein the structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

[0125] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the style-aware game strategy evaluation and calibration method provided by the above methods. The method includes: obtaining evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include win rate and ELO score extracted from the game evaluation system; performing unsupervised clustering based on the evaluation feature vectors to obtain multiple style clusters, wherein each style cluster corresponds to a behavioral pattern of a game strategy; performing style-conditional evaluation recalculation on each game strategy within each style cluster to obtain calibrated evaluation indicators, wherein the calibrated evaluation indicators include the percentile of the comprehensive score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster; and generating a structured evaluation report of the target game strategy based on the calibrated evaluation indicators, wherein the structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

[0127] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the style-aware game strategy evaluation and calibration method provided by the above methods. This method includes: acquiring evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include win rate and ELO score extracted from a game evaluation system; performing unsupervised clustering based on the evaluation feature vectors to obtain multiple style clusters, wherein each style cluster corresponds to a behavioral pattern of a game strategy; performing style-conditional evaluation recalculation on each game strategy within each style cluster to obtain calibrated evaluation indicators, wherein the calibrated evaluation indicators include the percentile of the overall score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster; and generating a structured evaluation report of the target game strategy based on the calibrated evaluation indicators, wherein the structured evaluation report includes a style cluster identifier and relative level information within the style cluster.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A style-aware game strategy evaluation and calibration method, characterized in that, include: Obtain evaluation feature vectors for multiple game strategies, wherein the evaluation feature vectors include win rate and ELO score extracted from the game evaluation system; Unsupervised clustering is performed based on the evaluation feature vectors to obtain multiple style clusters, where each style cluster corresponds to a behavioral pattern of a game strategy. For each game strategy within each style cluster, a style-conditional evaluation recalculation is performed to obtain a calibrated evaluation index. The calibrated evaluation index includes the percentile of the comprehensive score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster. Based on the calibrated evaluation indicators, a structured evaluation report of the target game strategy is generated, wherein the structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

2. The style-aware game strategy evaluation and calibration method according to claim 1, characterized in that, The process of obtaining the evaluation feature vectors of multiple game strategies includes: The original evaluation index of each game strategy in multiple game strategies is extracted from the preset game evaluation system. The original evaluation index includes ability value, win rate and ELO score. Obtain the environmental parameters for each game strategy, wherein the environmental parameters include map complexity, venue type, and game format; The original evaluation index is fused with the environmental parameters to obtain the initial feature vector of each game strategy; The initial feature vectors are standardized to obtain the evaluation feature vectors for each game strategy.

3. The style-aware game strategy evaluation and calibration method according to claim 1, characterized in that, The unsupervised clustering based on the evaluated feature vectors yields multiple style clusters, including: The K-Means clustering algorithm is used to cluster the evaluation feature vectors, and the optimal number of clusters is determined by the elbow rule. Based on the optimal number of clusters, the multiple game strategies are divided into multiple style clusters, and a style cluster identifier is assigned to each game strategy in the multiple game strategies.

4. The style-aware game strategy evaluation and calibration method according to claim 3, characterized in that, The process of recalculating the style-conditional evaluation for each game strategy within each style cluster to obtain calibrated evaluation metrics includes: For each style cluster, the following procedures are performed: Determine the overall score distribution, ELO score distribution, and win rate distribution of all game strategies within the style cluster; Based on the overall score distribution, determine the percentile of the overall score for each game strategy within the style cluster; Based on the ELO score distribution, determine the relative position of the ELO score of each game strategy within the style cluster; Based on the win rate distribution, determine the relative position of the win rate of each game strategy within the style cluster; Based on the percentile of the overall score, the relative position of the ELO score, and the relative position of the win rate, a calibrated evaluation index is determined for each game strategy.

5. The style-aware game strategy evaluation and calibration method according to claim 1, characterized in that, The process of generating a structured evaluation report of the target game strategy based on the calibrated evaluation indicators includes: The evaluation feature vector of the target game strategy is input into a pre-trained unsupervised clustering model to obtain the predicted style cluster of the target game strategy. Determine the overall score distribution, ELO score distribution, and win rate distribution of all game strategies corresponding to the predicted style cluster; By comparing and analyzing the evaluation feature vector of the target game strategy with the comprehensive score distribution, the ELO score distribution and the win rate distribution, the calibrated evaluation index of the target game strategy is obtained. Based on the calibrated evaluation indicators of the target game strategy, a structured evaluation report of the target game strategy is generated.

6. The style-aware game strategy evaluation and calibration method according to claim 1, characterized in that, The method further includes: Obtain the calibrated evaluation metrics for all game strategies and their corresponding style cluster identifiers; Dimensionality reduction is performed based on the calibrated evaluation metrics of all game strategies to obtain low-dimensional space coordinates. The low-dimensional spatial coordinates are visualized and rendered to generate a spatial distribution map of game strategy capabilities, wherein game strategies identified by different style clusters in the game strategy capability distribution map are distinguished by different colors or shapes. Based on the spatial distribution map of the strategy capabilities, capability boundary analysis is performed to obtain the clustering region and boundary characteristics of each style cluster.

7. A style-aware game strategy evaluation and calibration device, characterized in that, include: The acquisition module is used to acquire evaluation feature vectors of multiple game strategies, wherein the evaluation feature vectors include win rate and ELO score extracted from the game evaluation system; The clustering module is used to perform unsupervised clustering based on the evaluation feature vector to obtain multiple style clusters, wherein each style cluster corresponds to a behavioral pattern of a game strategy. The re-evaluation module is used to recalculate the style-conditional evaluation of each game strategy within each style cluster to obtain calibrated evaluation indicators. The calibrated evaluation indicators include the percentile of the comprehensive score within the style cluster, the relative position of the ELO score within the style cluster, and the relative position of the win rate within the style cluster. The generation module is used to generate a structured evaluation report of the target game strategy based on the calibrated evaluation index, wherein the structured evaluation report includes style cluster identifiers and relative level information within the style cluster.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the style-aware game strategy evaluation and calibration method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the style-aware game strategy evaluation and calibration method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the style-aware game strategy evaluation and calibration method as described in any one of claims 1 to 6.

Images