AI-based game performance analysis and operational stability prediction system

Abstract

This invention discloses an AI-based game performance analysis and operational stability prediction system, relating to the interdisciplinary fields of machine learning and game technology. The system includes a data acquisition and preprocessing module, a causal feature extraction module, a stability prediction module, and a causal root cause localization module. The data acquisition and preprocessing module collects game performance data from multiple channels and performs standardized processing to generate frame-level feature vectors. The causal feature extraction module constructs a causal graph using an improved PC algorithm, filters features, and adaptively adjusts feature weights. The stability prediction module uses a causal temporal convolutional network to predict anomaly probabilities and trigger warnings. The causal root cause localization module relies on virtual causal intervention to complete root cause identification and hierarchical judgment. This invention replaces traditional experience-based feature selection and correlation analysis with causal logic, eliminating the tedious manual selection of indicators and enabling the layer-by-layer localization of multi-level performance triggers, thus capturing potential game lag and crashes early.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of machine learning and game technology, specifically to a game performance analysis and operational stability prediction system based on artificial intelligence. Background Technology

[0002] With the continuous development of the gaming industry, technologies such as open-world design, ray tracing, physics simulation, and real-time global illumination are gradually becoming more widespread in various games, significantly improving game graphics precision, scene scale, and interactive complexity. The operation of current mainstream games involves the coordinated work of multiple stages, including CPU logic calculation, GPU rendering, memory scheduling, disk I / O, and network transmission. The operational status of each stage directly determines the overall performance of the game. The industry's requirements for game performance and stability are constantly increasing, making performance analysis and optimization one of the core aspects of the entire game development process.

[0003] Existing machine learning-based game performance analysis methods suffer from two core drawbacks. First, feature engineering relies entirely on expert experience, requiring developers to manually select potentially performance-related metrics. This is not only time-consuming and labor-intensive but also prone to overlooking cross-functional coupling features, leading to insufficient model accuracy. Second, all methods are based on correlation analysis, which can only discover the relationships between metrics but cannot distinguish between cause and effect, let alone locate indirect root causes. The accuracy of root cause location is generally low, failing to provide developers with effective optimization guidance. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based game performance analysis and operational stability prediction system. This system breaks down game frame generation into six mutually exclusive operational stages: CPU computation, GPU rendering, video memory scheduling, disk I / O, and network transmission. It then uses an improved PC algorithm incorporating domain priors to construct a full-link causal graph, automatically selecting feature subsets with real causal impact, eliminating the hassle of manually selecting indicators. The prediction part uses a causal temporal convolutional network, which can output the probability of stuttering and crashing tens of seconds in advance. Once an anomaly is triggered, virtual causal intervention is used to trace back the root cause layer by layer, forming a multi-level causal transmission chain, and performing graded processing based on confidence level.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a game performance analysis and operational stability prediction system based on artificial intelligence, the system comprising: Data Acquisition and Preprocessing Module: This module is used to collect full-link performance data of the game through three independent channels: engine layer, hardware layer and application layer. It performs four types of standardization preprocessing on the raw data in sequence: missing value handling, outlier handling, data standardization and time sequence alignment. It generates frame-level feature vectors for each frame, unifies the data format of different sources, and eliminates data disorder caused by time difference in acquisition. Causal Feature Extraction Module: Receives frame-level feature vectors, which are used to divide the game frame generation process into six mutually exclusive running stages. An improved PC algorithm with constraints from prior knowledge of the game domain is used to construct a full-link directed acyclic causal graph. The average causal effect of each index on the target performance variable is calculated, a subset of causal features is selected and redundant features are removed. At the same time, the game scene is identified in real time and the feature weights are dynamically adjusted. The scene-adaptive causal feature subset is output, which can reduce the consumption of invalid data in subsequent computing resources and adapt to the changing patterns of game performance under different scenarios. Stability prediction module: Receives a subset of scene-adaptive causal features, which is used to employ a causal temporal convolutional network. Inputting a historical causal feature sequence, it outputs the probability of stuttering and crashing at multiple future time points. When preset conditions are met, it triggers an early warning to expose potential operational failures in advance, giving developers ample time to handle them. Causal root cause localization module: Receives warning signals and frame-level feature vectors and scene-adaptive causal feature subsets for the corresponding time period. It is used to extract performance data for the corresponding time period when an anomaly is triggered, calculate the average causal effect through virtual causal intervention, locate the direct root cause and construct a multi-order causal transmission chain, calculate the confidence score and perform hierarchical processing, distinguish between related phenomena and real fault causes, and avoid blind operation during maintenance and rectification.

[0006] Furthermore, in the causal feature extraction module, the six mutually exclusive operation stages are CPU logic calculation, GPU vertex shading, pixel rendering, video memory scheduling, disk I / O, and network transmission; the prior knowledge constraints added to the improved PC algorithm include: indicators within the same stage can have causal relationships; upstream stage indicators can be the cause of downstream stage indicators, but not vice versa; disk I / O and network transmission stage indicators can have bidirectional causal relationships with all other stage indicators, constraining the drawing range according to the game's native operation logic, and reducing the number of invalid causal connections generated.

[0007] Furthermore, the formula for calculating the average causal effect by the causal feature extraction module is as follows: ; in, Indicators For target performance variables Average causal effect Represents the mathematical expectation. This indicates a causal intervention operation. Indicators The 95th percentile under normal operating conditions, Indicators The quintiles under normal operating conditions; The feature selection rules are as follows: retain indicators with an average causal effect greater than 0.05; for two indicators A and B that have a direct causal relationship, if A is the direct cause of B and the average causal effect of A is greater than the average causal effect of B, then retain A and remove B, and rely on quantitative values ​​to eliminate parameters with weak correlation and simplify the size of the feature set.

[0008] Furthermore, the causal feature extraction module uses a lightweight convolutional neural network with 2.3M parameters to identify game scenes and outputs the probability distributions of five scenes: combat, loading, cutscenes, open-world exploration, and menu interface. When the probability of a scene is greater than 0.9, the module switches to the causal graph and feature weight coefficients corresponding to that scene. When the scene probability is between 0.5 and 0.9, the module uses a weighted average of multiple scene causal graphs and weight coefficients, with the weights being the probability values ​​of each scene. The single-frame inference latency of the lightweight convolutional neural network is less than 1 millisecond, and the lightweight structure does not consume additional computing power from the game device.

[0009] Furthermore, the causal temporal convolutional network in the stability prediction module consists of an input layer, three sequentially connected causal convolutional blocks, a global average pooling layer, and two parallel output heads. Each causal convolutional block contains two causal convolutional layers, a residual connection, and a layer normalization layer. The dilation coefficients of the three causal convolutional blocks are 1, 2, and 4, respectively. The causal convolutional layers only use the input data from the current time and previous times for calculation, without revealing data information from future times. The two parallel output heads output the stuttering probability and the crash probability, respectively. When the predicted stuttering probability is greater than 0.8 or the crash probability is greater than 0.5, an early warning signal is triggered.

[0010] Furthermore, the specific steps of the causal root cause localization module in performing virtual causal intervention are as follows: replacing all values ​​of the indicator to be intervened within the abnormal time period with the average value of the indicator under normal operating conditions within 1 minute before the abnormality occurred; inputting the feature sequence after intervention into a pre-trained frame rate prediction model to obtain the predicted frame rate after intervention; calculating the difference between the predicted frame rates before and after intervention as the average causal effect of the indicator on this performance abnormality; replacing the abnormal field with normal values ​​to simulate the operational changes brought about by the change of a single parameter.

[0011] Furthermore, the steps of the causal root cause localization module in constructing a multi-level causal transmission chain are as follows: the indicator with the largest average causal effect is identified as the direct root cause; all parent nodes pointing to the direct root cause are traced backward along the causal graph, and the average causal effect of each parent node on the direct root cause is calculated; the parent node with the largest average causal effect is identified as the next-level root cause, and the tracing process is repeated until the original cause without a parent node or the causal effect is less than 0.05 is reached; all the identified root causes are connected in causal order to form a multi-level causal transmission chain, and the source is traced layer by layer along the data influence link to completely restore the fault's layer-by-layer propagation path.

[0012] Furthermore, the formula for calculating the confidence score by the causal root cause localization module is as follows: ; in, This represents the confidence score for the root cause. This represents the average causal effect of the root cause. This represents the weight of the corresponding edge in the causal graph. This represents the data quality coefficient of the indicator, with a value ranging from 0 to 1, and is calculated based on the indicator's sampling rate and data integrity. The tiered processing rules are as follows: when the confidence level is greater than 0.8, the root cause localization result is output directly; when the confidence level is between 0.5 and 0.8, the sampling frequency is increased to 1000Hz and the root cause analysis is re-executed; when the confidence level is less than 0.5, the developers are prompted to conduct manual review; the credibility of the fault is determined by a combination of three parameters to avoid misjudging the source of the fault based on a single data point.

[0013] Furthermore, the root cause localization module is also used to prioritize multiple root causes, with the following ranking dimensions and corresponding weights: average causal effect 0.4, occurrence frequency 0.3, impact range 0.2, and repair difficulty 0.1. A comprehensive priority score is calculated based on the sum of the products of the scores of each dimension and their corresponding weights, and the results are sorted from high to low. Based on the root cause-solution knowledge base, executable optimization suggestions are generated to facilitate staff to rectify faults one by one according to their importance.

[0014] Furthermore, it also includes a closed-loop feedback module, which receives root cause information, optimization suggestions, and operational data from the stability prediction module output by the root cause localization module. This module receives confirmation and correction information from developers regarding the root cause localization results, receives performance change data after the implementation of optimization measures, monitors the performance of the prediction model in real time, and updates the causal graph structure, root cause-solution knowledge base, and prediction model parameters based on feedback information and monitoring results. The update rules for the closed-loop feedback module are as follows: the prediction model is fine-tuned weekly using newly collected player data; when the difference between the causal structure of 1000 consecutive frames of data and the existing causal graph exceeds 10%, an incremental update of the causal graph is triggered; when the accuracy of the prediction model is below 85%, a model retraining process is automatically triggered, continuously optimizing the entire analysis logic along with game version iterations to maintain long-term system adaptability.

[0015] Compared with existing technologies, this AI-based game performance analysis and operational stability prediction system has the following advantages: I. This invention divides the game frame generation process into multiple mutually exclusive stages, and uses an improved PC algorithm with constraints from prior knowledge in the game domain to construct a full-link directed acyclic causal graph. It calculates the average causal effect of each index on the target performance variable, selects a subset of causal features and removes redundant features. This method does not rely on the experience of game developers, can automatically discover coupling features across stages, reduces feature dimensionality, improves feature extraction efficiency, and solves the problems of traditional feature engineering being time-consuming, labor-intensive, and prone to missing key features.

[0016] Second, this invention calculates the average causal effect of each indicator on performance anomalies using virtual causal intervention technology, locates direct root causes and constructs multi-level causal transmission chains, and performs hierarchical processing in conjunction with confidence scores. This method can accurately distinguish between causal and correlational relationships between performance indicators, improve the accuracy of root cause location, and locate multi-level indirect root cause chains. At the same time, it uses causal temporal convolutional networks for stability prediction, effectively extending the early warning time of stability problems and providing developers with clear optimization directions.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a diagram showing the overall module architecture and data flow of the system of the present invention; Figure 2 This is a flowchart of the causal feature extraction and dynamic screening process of the present invention; Figure 3 This is a flowchart of the causal root cause localization and hierarchical processing of the present invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This embodiment discloses a specific implementation of an AI-based game performance analysis and operational stability prediction system. After system startup, initial configuration is completed first, loading a pre-trained scene recognition model, a causal temporal convolutional network model, and an initial causal graph structure. Simultaneously, the parameters of each acquisition channel and the data cache queue are initialized. During game execution, the system sequentially completes data acquisition, preprocessing, feature extraction, stability prediction, root cause localization, and result output according to a fixed temporal logic. Modules interact with each other via memory sharing; the output of one module serves as the input of the next, ensuring real-time data transmission and low latency.

[0022] like Figure 1 As shown, the system consists of a data acquisition and preprocessing module, a causal feature extraction module, a stability prediction module, a causal root cause localization module, and a closed-loop feedback module. These modules are connected in sequence to form a closed-loop optimization.

[0023] The first step in system operation is data acquisition and preprocessing, which is completed in parallel through three independent acquisition channels. The engine-layer channel directly interfaces with the game engine's underlying native interface, requiring no modification to the game's core business logic. Data acquisition is achieved solely through the engine's plugin mechanism, collecting information such as the rendering pipeline's time consumption at each stage, the number of drawing calls, particle system running status, physics engine collision detection counts, object pool utilization, garbage collection trigger time and duration, etc. The hardware-layer channel uses the performance counter interface provided by the operating system kernel to collect hardware operation data such as the CPU's core utilization, thread execution time, cache hit rate, GPU load, video memory usage and bandwidth consumption, disk read / write speed and IOPS, and network latency and jitter. The application-layer channel collects upper-layer application data such as the game's current scene identifier, player action sequences, system-triggered event logs, real-time frame rate, and frame time. All collected data is appended with a unified high-precision timestamp, with microsecond-level accuracy. A hardware clock synchronization algorithm aligns data from different channels to the same timeline, ensuring that the time synchronization error between data from different sources does not exceed 1 millisecond.

[0024] The preprocessing process sequentially performs missing value handling, outlier handling, data standardization, and temporal alignment. For missing values ​​in continuous data, linear interpolation is used for imputation, calculating the missing value based on the values ​​of the two valid data points before and after the missing value and the time interval. For missing values ​​in discrete data, forward imputation is used, replacing the missing value with the value of the previous valid data point. The 3σ principle is used to identify outliers; data points with values ​​exceeding three times the standard deviation of the mean are identified as outliers and removed to prevent them from interfering with subsequent causal analysis results. Z-score standardization is applied to all continuous indicators, converting each indicator's value into a standard distribution with a mean of 0 and a standard deviation of 1, eliminating the impact of dimensional differences between different indicators on the model's calculation results. Finally, using the game's frame time as a benchmark, all data with different sampling frequencies are aligned to each game frame, generating a frame-level feature vector for each frame, which is then input into the subsequent causal feature extraction module for processing.

[0025] The causal feature extraction module receives the aforementioned frame-level feature vectors and first divides the game frame generation process into six mutually exclusive operational stages: CPU logic computation, GPU vertex shading, pixel rendering, video memory scheduling, disk I / O, and network transmission. Each stage contains several performance indicators with clearly defined physical meanings. This division strictly follows the physical flow of game frame generation, ensuring clear logical relationships between indicators in each stage, and clear boundaries between different stages without overlap or omissions. The causal feature extraction module uses an improved PC algorithm with pre-defined prior knowledge constraints from the game domain to construct a full-link directed acyclic causal graph. Before the algorithm runs, predefined prior knowledge constraints are loaded. These constraints are determined based on the inherent logic of game frame generation, specifically including the possibility of causal relationships between indicators within the same stage, the possibility that indicators in upstream stages can cause indicators in downstream stages, and the possibility that indicators in downstream stages cannot cause indicators in upstream stages. Indicators in the disk I / O and network transmission stages can have bidirectional causal relationships with indicators in all other stages. After adding these prior knowledge constraints, the number of erroneous edges in the causal graph construction process is significantly reduced, the construction time can be shortened by more than half, and the accuracy of the causal graph is also improved.

[0026] The improved PC algorithm first constructs a complete undirected graph containing all performance metrics, where each node represents a performance metric, and there is an undirected edge between any two nodes. Then, the algorithm progressively removes edges without causal relationships through conditional independence checks. For each pair of metrics X and Y, the algorithm sequentially checks whether X and Y are independent under different sizes of variable sets Z. In this embodiment, partial correlation testing is used as the method for conditional independence testing, which is suitable for continuous performance data. If the absolute value of the partial correlation coefficient between X and Y is less than a preset threshold under the given variable set Z, then X and Y are determined to be independent under the given Z, and the edge between X and Y is deleted. The size of the variable set Z gradually increases from 0 until all possible condition sets have been tested. After edge deletion, the algorithm determines the direction of the remaining edges based on the V-structure principle and prior knowledge constraints, ultimately obtaining a directed acyclic graph for the game frame generation.

[0027] Based on the constructed causal graph, the causal feature extraction module calculates the average causal effect of each indicator on the target performance variable. The calculation formula is as follows: ; in, Indicators For target performance variables Average causal effect Represents the mathematical expectation. This indicates a causal intervention operation. Indicators The 95th percentile under normal operating conditions, Indicators The quintiles under normal operating conditions. In this embodiment, the backdoor adjustment method is used to calculate the do operator. Since the constructed causal graph is a directed acyclic graph and satisfies the backdoor criterion, the expected value after intervention can be calculated by adjusting the set of parent nodes of X. The average causal effect reflects the degree of influence of changes in index X on the target performance variable Y; a larger value indicates a more significant impact of the index on game performance.

[0028] The causal feature extraction module filters causal feature subsets based on average causal effect. First, it retains all indicators with an average causal effect greater than a preset threshold, forming an initial feature subset. Then, it removes redundant features from the initial feature subset. For two indicators A and B with a direct causal relationship, if A is a direct cause of B, and the average causal effect of A is greater than the average causal effect of B, then A is retained and B is removed. This is because if A is a cause of B, then changes in A will lead to changes in B. Retaining A allows us to capture the impact of B on performance, while also reducing feature dimensionality and the computational load of subsequent models.

[0029] Meanwhile, the causal feature extraction module employs a lightweight convolutional neural network to identify the current game scene in real time. This network is an improvement on the MobileNetV2 architecture, removing the final classification layer and replacing it with a fully connected layer with an output dimension of 5, corresponding to five scene types: combat, loading, cutscenes, open-world exploration, and menu interface. The network input is a grayscale image of the current game frame, and the output is the probability distribution of the five scenes. When the probability of a scene is greater than 0.9, the system switches to the causal graph and feature weight coefficients corresponding to that scene; when the scene probability is between 0.5 and 0.9, a weighted average of multiple scene causal graphs and weight coefficients is used, with the weights being the probability values ​​of each scene. This dynamic adjustment mechanism allows the system to adapt to the performance differences under different game scenes, improving the accuracy of subsequent analysis results. Figure 2 As shown, the above-described causal feature extraction and dynamic filtering process starts with multi-source performance data input, and sequentially goes through process segmentation, causal graph construction, average causal effect calculation and filtering, and redundant feature removal, ultimately completing dynamic scene recognition and feature weighting. The module finally outputs a subset of scene-adaptive causal features, which is then input into the subsequent stability prediction module for processing.

[0030] The stability prediction module receives the adaptive causal feature subset of the aforementioned scenario and employs a causal temporal convolutional network as the prediction model. This model consists of an input layer, three sequentially connected causal convolutional blocks, a global average pooling layer, and two parallel output heads. Each causal convolutional block contains two causal convolutional layers, a residual connection, and a layer normalization layer. The dilation coefficients of the three causal convolutional blocks are 1, 2, and 4, respectively, capable of capturing temporal dependencies at short, medium, and long time scales. The causality of the causal convolutional layers is achieved by masking the convolutional kernels, setting the weights corresponding to future time moments to zero. This ensures that the model only uses input data from the current time moment and earlier during computation, without leaking data information from future time moments, thus meeting the causality requirements of temporal prediction.

[0031] The model takes as input a causal feature sequence from the past 10 seconds and outputs the probabilities of game stuttering and crashing at three time points: 30, 45, and 60 seconds. During training, the model uses supervised learning with historical performance data labeled with stuttering and crash events, employing a binary cross-entropy loss function and incorporating L2 regularization to prevent overfitting. After training, the model processes the input causal feature sequence in real time, outputting the latest stability prediction every 100 milliseconds. When the predicted stuttering probability is greater than 0.8 or the crash probability is greater than 0.5, the stability prediction module triggers an early warning signal and simultaneously sends the warning signal, the corresponding frame-level feature vector, and a scene-adaptive causal feature subset to the causal root cause localization module.

[0032] The causal root cause localization module receives the aforementioned warning signals and corresponding time-period feature data. Upon receiving a warning signal, detecting a frame rate below a preset threshold for more than one second, or a game crash, the root cause analysis process is initiated. The module first extracts all performance data from 30 seconds before the anomaly to the time of the anomaly, using this as input data for root cause analysis. Then, it performs virtual causal intervention on each indicator in the causal feature subset, replacing all values ​​of the indicator to be intervened with the average value of that indicator under normal operating conditions within one minute before the anomaly. The intervened feature sequence is then input into a pre-trained frame rate prediction model. This model uses the same causal temporal convolutional network structure as the stability prediction model, with the causal feature sequence as input and the predicted frame rate as output. The difference between the predicted frame rates before and after intervention is calculated as the average causal effect of that indicator on the current performance anomaly.

[0033] The causal root cause localization module identifies the indicator with the largest average causal effect as the direct root cause. It then traces backward along the causal graph, linking all parent nodes pointing to the direct root cause and calculating the average causal effect of each parent node on the direct root cause. The parent node with the largest average causal effect is identified as the next higher-level root cause, and this tracing process is repeated until an original cause with no parent node or a causal effect less than 0.05 is found. All identified root causes are connected sequentially according to their causal relationships, forming a multi-level causal transmission chain that clearly demonstrates the complete transmission path of performance abnormalities from the original cause to the final manifestation. Figure 3 As shown, this module calculates the average causal effect through virtual causal intervention, locates the direct root cause and traces back to the parent node, constructs a multi-level causal transmission chain, and then performs hierarchical processing based on the confidence score, finally outputting the root cause location result or prompting manual review.

[0034] The causal root cause localization module calculates the confidence score for each root cause using the following formula: ; in, This represents the confidence score for the root cause. This represents the average causal effect of the root cause. This represents the weight of the corresponding edge in the causal graph, which is calculated based on the conditional independence test results during the causal graph construction process. This represents the data quality coefficient of the indicator, ranging from 0 to 1. The data quality coefficient is the product of the indicator's sampling rate compliance coefficient and the data integrity coefficient. Data integrity is the ratio of the number of valid data points to the total number of data points. Root cause analysis results are graded based on the confidence score: if the confidence score is greater than 0.8, the root cause analysis result is output directly; if the confidence score is between 0.5 and 0.8, the sampling frequency is increased to 1000Hz, and 10 seconds of data are collected again before re-performing the root cause analysis; if the confidence score is less than 0.5, the developers are prompted to conduct a manual review.

[0035] For cases with multiple root causes, the root cause localization module prioritizes these root causes. The ranking dimensions include average causal effect, frequency of occurrence, scope of impact, and difficulty of remediation, with corresponding weights of 0.4, 0.3, 0.2, and 0.1, respectively. A comprehensive priority score is calculated based on the sum of the products of each dimension's score and its corresponding weight, and the root causes are ranked from highest to lowest. The module generates corresponding actionable optimization suggestions based on a root cause-solution knowledge base. This knowledge base is a structured database built upon historical performance remediation cases and developer experience, storing solutions, implementation steps, and historical optimization effect data for common performance root causes. Finally, the module synchronously outputs the root cause analysis results and optimization suggestions to the closed-loop feedback module.

[0036] The closed-loop feedback module receives the root cause analysis results and optimization suggestions, while simultaneously collecting real-time operational data from the stability prediction module. It is responsible for gathering various feedback data during system operation to achieve continuous system optimization. The closed-loop feedback module also receives confirmation and correction information from developers regarding the root cause localization results. This information is used to adjust the weights of corresponding edges in the cause-effect graph, optimize the graph's structure, and improve the accuracy of subsequent root cause localization. The module also receives performance change data after the implementation of optimization measures, which is used to update the optimization effect score of the corresponding solution in the root cause-solution knowledge base.

[0037] The closed-loop feedback module monitors the performance of the prediction model in real time, including metrics such as accuracy, false positive rate, and false negative rate. The prediction model is fine-tuned weekly using newly collected player data to adapt to performance differences caused by game version updates and changes in the operating environment. When the causal structure of 1000 consecutive frames differs from the existing causal graph by more than 10%, an incremental update of the causal graph is triggered. The difference in causal structure is calculated by comparing the existence and weight differences of edges in the two causal graphs. The incremental update only adjusts the changed parts, without retraining the entire causal graph. When the accuracy of the prediction model falls below 85%, a model retraining process is automatically triggered, retraining the model using the latest dataset.

[0038] The system described in this embodiment can be applied to the development, testing, and launch phases of various PC games, mobile games, and console games. It can help developers quickly identify and locate performance issues, shorten the optimization cycle, and improve the game's operational stability and player experience.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An artificial intelligence-based game performance analysis and running stability prediction system, characterized in that, The system includes: Data acquisition and preprocessing module: used to collect full-link performance data of game operation through three independent channels: engine layer, hardware layer and application layer, to standardize and preprocess the raw data, and to generate frame-level feature vectors corresponding to each frame; Causal Feature Extraction Module: Receives frame-level feature vectors, which are used to divide the game frame generation process into six mutually exclusive running stages. An improved PC algorithm with constraints from prior knowledge of the game domain is used to construct a full-link directed acyclic causal graph. The average causal effect of each index on the target performance variable is calculated. A subset of causal features is selected and redundant features are removed. At the same time, the game scene is identified in real time and the feature weights are dynamically adjusted. The scene-adaptive causal feature subset is output. Stability prediction module: Receives a subset of scene-adaptive causal features, which is used to employ a causal temporal convolutional network, inputting a historical causal feature sequence, and outputting the probability of stuttering and crashing at multiple future time points, triggering an early warning when preset conditions are met; Causal root cause localization module: Receives warning signals and frame-level feature vectors and scene-adaptive causal feature subsets for the corresponding time period. It is used to extract performance data for the corresponding time period when an anomaly is triggered, calculate the average causal effect through virtual causal intervention, locate the direct root cause and construct a multi-order causal transmission chain, calculate the confidence score and perform hierarchical processing. 2.The AI-based game performance analysis and operation stability prediction system according to claim 1, wherein, In the causal feature extraction module, the six mutually exclusive operation stages are CPU logical calculation, GPU vertex shading, pixel rendering, video memory scheduling, disk I / O, and network transmission. The prior knowledge constraints added to the improved PC algorithm include: indicators within the same stage can have causal relationships; upstream stage indicators can be the cause of downstream stage indicators, but not vice versa; disk I / O and network transmission stage indicators can have bidirectional causal relationships with all other stage indicators.

3. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The formula for calculating the average causal effect by the causal feature extraction module is as follows: ； in, Indicators For target performance variables Average causal effect Represents the mathematical expectation. This indicates a causal intervention operation. Indicators The 95th percentile under normal operating conditions, Indicators The quintiles under normal operating conditions; The feature selection rules are as follows: retain indicators with an average causal effect greater than 0.05; for two indicators A and B that have a direct causal relationship, if A is the direct cause of B and the average causal effect of A is greater than the average causal effect of B, then retain A and remove B.

4. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The causal feature extraction module uses a lightweight convolutional neural network to identify game scenes and outputs the probability distribution of five scenes: combat, loading, cutscenes, open world exploration, and menu interface. When the probability of a scene is greater than 0.9, the module switches to the causal graph and feature weight coefficients corresponding to that scene. When the scene probability is between 0.5 and 0.9, the module uses a weighted average of multiple scene causal graphs and weight coefficients, with the weights being the probability values ​​of each scene. The single-frame inference latency of the lightweight convolutional neural network is less than 1 millisecond.

5. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The causal temporal convolutional network in the stability prediction module consists of an input layer, three sequentially connected causal convolutional blocks, a global average pooling layer, and two parallel output heads. Each causal convolutional block contains two causal convolutional layers, a residual connection, and a layer normalization layer. The dilation coefficients of the three causal convolutional blocks are 1, 2, and 4, respectively. The causal convolutional layers only use the input data from the current time and previous times for calculation, without revealing data information from future times. The two parallel output heads output the stuttering probability and the crash probability, respectively. When the predicted stuttering probability is greater than 0.8 or the crash probability is greater than 0.5, an early warning signal is triggered.

6. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The specific steps for the virtual causal intervention performed by the causal root cause localization module are as follows: replace all values ​​of the indicator to be intervened within the abnormal time period with the average value of the indicator under normal operating conditions within 1 minute before the abnormality occurred; input the feature sequence after intervention into the pre-trained frame rate prediction model to obtain the predicted frame rate after intervention; calculate the difference between the predicted frame rate before and after intervention as the average causal effect of the indicator on this performance abnormality.

7. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The steps for constructing a multi-level causal transmission chain by the causal root cause localization module are as follows: the indicator with the largest average causal effect is identified as the direct root cause; all parent nodes pointing to the direct root cause are traced backward along the causal graph, and the average causal effect of each parent node on the direct root cause is calculated; the parent node with the largest average causal effect is identified as the next-level root cause, and the tracing process is repeated until the original cause without a parent node or the causal effect is less than 0.05 is found; all the identified root causes are connected in causal order to form a multi-level causal transmission chain.

8. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The formula for calculating the confidence score by the causal root cause localization module is as follows: ； in, This represents the confidence score for the root cause. This represents the average causal effect of the root cause. This represents the weight of the corresponding edge in the causal graph. This represents the data quality coefficient of the indicator, with a value ranging from 0 to 1, and is calculated based on the indicator's sampling rate and data integrity. The tiered processing rules are as follows: when the confidence level is greater than 0.8, the root cause localization result is output directly; when the confidence level is between 0.5 and 0.8, the sampling frequency is increased to 1000Hz and the root cause analysis is re-executed; when the confidence level is less than 0.5, the developers are prompted to conduct manual review.

9. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, The root cause localization module is also used to prioritize multiple root causes, with the following ranking dimensions and corresponding weights: average causal effect 0.4, occurrence frequency 0.3, impact range 0.2, and repair difficulty 0.

1. The module calculates a comprehensive priority score based on the sum of the products of the scores of each dimension and their corresponding weights, sorts the root causes from high to low, and generates executable optimization suggestions based on the root cause-solution knowledge base.

10. The AI-based game performance analysis and operational stability prediction system according to claim 1, characterized in that, It also includes a closed-loop feedback module, which receives root cause information, optimization suggestions, and operational data from the stability prediction module output by the root cause localization module. This module receives confirmation and correction information from developers regarding the root cause localization results, receives performance change data after the implementation of optimization measures, monitors the performance of the prediction model in real time, and updates the causal graph structure, root cause-solution knowledge base, and prediction model parameters based on feedback information and monitoring results. The update rules for the closed-loop feedback module are as follows: the prediction model is fine-tuned weekly using newly collected player data; when the difference between the causal structure of 1000 consecutive frames of data and the existing causal graph exceeds 10%, an incremental update of the causal graph is triggered; and when the accuracy of the prediction model is below 85%, the model retraining process is automatically triggered.

Images