An artificial intelligence-based virtual character real-time perception and action decision method

Abstract

The application discloses a virtual character real-time perception and action decision method based on artificial intelligence, relates to the technical field of artificial intelligence and virtual character control, and comprises the following steps: acquiring a multi-modal perception state matrix of a virtual character, performing unit root test to form a significant stimulus source set; calculating a time series mutual information distance, calculating a perception information density, performing diffusion clustering by using the perception information density, and generating a cognitive environment slice set; constructing an error correction model, calculating the absolute value of a long-term error correction coefficient change quantity between perception variables, and calculating the absolute value of a short-term Granger causality index change quantity; calculating a macro stress score, fusing the macro stress score and a micro stress score to obtain a core decision driving force score; calculating an attention weight matrix, and iteratively and re-weighting to generate a final decision action and a push explanation chain. The application greatly improves the rationality, explainability and self-adaptive survival ability of the virtual character action in a complex and sudden environment.

Description

Technical Field

[0001] The present invention relates to the technical field of a method for real-time perception and action decision-making of virtual characters based on artificial intelligence. Background Technology

[0002] In computer-generated 3D virtual environments or embodied intelligent scenarios, virtual characters need to continuously interact with the outside world and react in real time. With the exponential increase in the complexity of application environments, the multimodal perception data received by virtual characters exhibits a trend of massive volume and redundancy. Existing methods for perception and decision-making in complex environments for virtual characters have significant limitations. Traditional behavior trees and finite state machines rely on pre-written rules and lack flexible contextual association capabilities when facing unknown and sudden events, failing to accurately identify the true threat sources in massive amounts of perception data. Existing deep reinforcement learning decision-making methods are black-box models, consuming significant computational resources during the inference phase and struggling to generate logically consistent reasoning explanations after decision-making. Current technologies cannot distinguish between macro-level global threats and micro-level local lethal threats in real-time perception, cannot perform motivational inference analysis based on underlying error correction data, and cannot adaptively reweight attention for non-stationary abrupt changes, leading to rigid character decisions, untraceable actions, and a lack of intelligent decision-making and action tracing mechanisms that balance interpretability and real-time performance.

[0003] Currently, Chinese invention patent CN116236783B discloses a method, device, and storage medium for a virtual character to follow movement, including the following steps: Step S1, constructing a virtual space, a virtual character, and a user character; obtaining the real-time position between the user character and the virtual character; and calculating the straight-line distance between the user character and the virtual character; Step S2, determining whether the straight-line distance exceeds a preset value; when the straight-line distance exceeds the preset value, adjusting the state of the virtual character to a moving state, and the virtual character moves towards the target position; Step S3, when the virtual character is at the target position, detecting the user's current moving state, adjusting the state of the virtual character to a standby state, and following the user character. In this invention, the user can see and perceive the existence of the virtual character and interact with it. However, existing technologies cannot distinguish between macro-level global threats and micro-level local fatal threats in real-time perception, cannot infer motivations based on error correction, cannot adaptively reweight attention for non-stationary abrupt changes, and lack intelligent decision-making and action tracing mechanisms that balance interpretability and real-time performance. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies cannot distinguish between macroscopic global threats and microscopic local fatal threats in real-time perception, cannot infer motivations based on error correction, cannot perform adaptive attention reweighting for non-stationary mutation data, and lack intelligent decision-making and action tracing mechanisms that take into account both interpretability and real-time performance.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for real-time perception and action decision-making of virtual characters based on artificial intelligence, comprising the following steps:

[0006] Step S1: Obtain the multimodal perception state matrix of the virtual character, divide the multimodal perception state matrix into historical stationary state windows and current abrupt state windows, perform unit root tests on the data in the current abrupt state window, extract the perceptual variables for which the first hypothesis is true, and form a set of significant stimulus sources.

[0007] Step S2: Calculate the temporal mutual information distance between perceptual variables in the set of salient stimulus sources, construct a perceptual nearest neighbor network based on the temporal mutual information distance and calculate the perceptual information density, perform diffusion clustering using the perceptual information density, and generate a set of cognitive environment slices;

[0008] Step S3: Construct an error correction model within each cognitive context slice queue, calculate the absolute value of the change in the long-term error correction coefficient between perceptual variables, and calculate the absolute value of the change in the short-term Granger causality index.

[0009] Step S4: Calculate the macro stress score based on the two absolute values ​​in step S3, and obtain the micro stress score by averaging the number of connection edges in the perceptual nearest neighbor network. Combine the macro stress score and the micro stress score to obtain the core decision-making driving force score.

[0010] Step S5: Calculate the attention weight matrix based on the core decision-driving force score, use the attention weight matrix to iteratively reweight the macro stress score, output the converged core decision-driving force score ranking, and generate the final decision action and reasoning explanation chain based on the perceptual variable corresponding to the top-ranked core decision-driving force score.

[0011] Preferably, step S1 includes:

[0012] The behavior time step and columns of the multimodal perception state matrix are feature types, and the feature types include the virtual character's visual range perception, auditory volume perception, self-health value state perception, and cooldown time state perception.

[0013] The multimodal sensing state matrix is ​​divided along the time axis to obtain a historical steady state window representing the memory of calm baseline and a current abrupt state window representing the memory of sudden events.

[0014] Within the current mutation state window, a unit root test is performed on each sensing feature time series in the multimodal sensing state matrix, and the test statistic is obtained by calculating the difference equation of the sensing feature time series.

[0015] The extended Dickie-Fowler (ADF) test was used to test the first and second hypotheses. The first hypothesis was that the time series of target-sensing features has a unit root, and the second hypothesis was that the time series of target-sensing features does not have a unit root.

[0016] The determination is made based on the comparison between the significance probability of the test statistic and the initial judgment criterion:

[0017] When the saliency probability is greater than or equal to the initial judgment criterion, the first hypothesis is determined to be true, the perceptual feature time series is marked as an abnormal fluctuation state, and all perceptual feature time series in the abnormal fluctuation state are merged to form a saliency stimulus source set.

[0018] When the significance probability is less than the initial judgment criterion, the second hypothesis is determined to be true, and the perceived feature time series is marked as maintaining a stable baseline state.

[0019] Preferably, step S2 includes:

[0020] In a set of salient stimulus sources, the temporal mutual information distance, which measures the degree of overlap of sequence information, is obtained by calculating the probability distribution of any two perceptual variable sequences and obtaining their mutual information.

[0021] Based on the temporal mutual information distance, the K nearest neighbor nodes of each perceptual variable are extracted to form a nearest neighbor node set. When the first perceptual variable is located in the nearest neighbor node set of the second perceptual variable, and the second perceptual variable is also located in the nearest neighbor node set of the first perceptual variable, a nearest neighbor bidirectional connection edge is constructed between the first perceptual variable and the second perceptual variable. All nearest neighbor bidirectional connection edges are merged to obtain the perceptual nearest neighbor network.

[0022] The number of nearest bidirectional connections included in each sensing variable in the statistical sensing nearest neighbor network is used as the sensing information density of that sensing variable.

[0023] Preferably, the mean of the perceived information density of all perceived variables in the set of salient stimulus sources is calculated. When the perceived information density of a perceived variable is greater than the mean of the perceived information density, it is determined that the perceived variable is in a high-density complex environment and a first diffusion boundary value is assigned to it.

[0024] When the perceived information density of a perceived variable is less than or equal to the mean perceived information density, the perceived variable is determined to be in a low-density isolated environment, and a second diffusion boundary value is assigned to it.

[0025] The first diffusion boundary value is smaller than the second diffusion boundary value;

[0026] The system searches for shared neighbor nodes outward from the center of the perceptual variable. When the number of shared neighbor nodes is greater than the corresponding diffusion boundary value, the neighbor nodes are grouped into the same cognitive environment slice queue until the search process ends. The output is a set of cognitive environment slices composed of interconnected perceptual variables.

[0027] Preferably, step S3 includes:

[0028] Extract all perceptual variables from a single cognitive situation slice queue in the cognitive environment slice set, and establish a vector autoregressive equation based on all perceptual variables;

[0029] Extract the cointegration relationship equation from the vector autoregression equation, calculate the change in the long-term error correction coefficient between the current abrupt state window and the historical stationary state window, and take the absolute value to obtain the first absolute value intensity.

[0030] Significance verification calculations are performed on the lagged terms in the vector autoregression equation. The short-term Granger causality index change caused by the introduction of the current mutation state window data is calculated and its absolute value is taken to obtain the second absolute value strength.

[0031] Preferably, step S4 includes:

[0032] The formula for calculating the macro-stress score is as follows:

[0033] ;

[0034] in, For the perceived variable i, the macro-stress score. Let i be the set of neighbor nodes connected to the sensing variable i in the nearest neighbor network. Let be the intensity of the first absolute value of the perception variable i for neighbor node j. Let be the intensity of the second absolute value of the perception variable i of neighbor node j. The initial macro- and micro-distribution coefficients are taken as constants between 0 and 1.

[0035] The formula for calculating the micro-stress score is:

[0036] ;

[0037] in, For the micro-stress score of perceived variable i, It is a scalar representing the total number of nodes in the neighbor node set.

[0038] Preferably, the expression for calculating the core decision-driving force score is:

[0039] ;

[0040] in, The core decision-driving force score for perceived variable i. The logistic smoothing activation function maps real numbers to the interval from constant 0 to constant 1. To perform the macro stress score after minimax normalization, To perform the micro-stress score after minimax normalization, and These are the first mapping weight and the second mapping weight, respectively, and B is the mapping bias constant.

[0041] Preferably, step S5 includes:

[0042] Calculate the attention weight matrix based on the core decision-driving force scores of neighboring nodes of the perceived variable i.

[0043] The attention weight is substituted into the macro stress score for iterative reweighting calculation, the macro stress score is updated, and the core decision-making driving force score is updated based on the updated macro stress score.

[0044] Preferably, the core decision-driving force score is continuously updated by iterative reweighting calculation. After each iteration, the absolute value of the difference between the current iteration's core decision-driving force score and the previous iteration's core decision-driving force score is calculated, and the sum of all absolute values ​​of the differences during the iteration process is calculated.

[0045] When the sum of the absolute values ​​of the differences is less than the initial convergence minimum constant, stop the iterative calculation, extract the core decision driving force scores of all perceptual variables and sort them in descending order, and output the ranking of the converged core decision driving force scores.

[0046] Extract the perceptual variable corresponding to the top-ranked core decision-making driving force score, and set the perceptual variable as the final inference root cause;

[0047] The motion engine module of the virtual character is invoked, and the corresponding physical avoidance response coordinates are output based on the final inference root cause. The virtual character is then controlled to move towards the safe coordinate area to generate the final decision action.

[0048] Preferably, the inference explanation chain generated based on the perceptual variable corresponding to the top-ranked core decision-making driving force score includes:

[0049] The steps of source tracing reasoning are constructed using a natural language processing model, and these steps specifically include:

[0050] Step S51: Concatenate the unit root test state reversal information to generate the perception trigger paragraph, and output the first interpretation paragraph;

[0051] Step S52: Concatenate the perceived information density value and the target diffusion boundary value to generate a context-related paragraph, and output the second explanatory paragraph;

[0052] Step S53: Concatenate the changes in long-term error correction coefficients to generate the expected deviation paragraph, and output the third explanatory paragraph;

[0053] Step S54: Concatenate the short-term Granger causality index change to generate the primacy reflex trigger paragraph, and output the fourth explanatory paragraph;

[0054] Step S55: After converging, the attention weights approach a constant of 1 to generate a closed-loop locking paragraph, and output the fifth explanatory paragraph;

[0055] Step S56: Combine all explanatory paragraphs to output the inference chain that follows the chronological causal order.

[0056] The beneficial effects of this invention are as follows: By performing unit root tests on the multimodal perception state matrix, this invention accurately filters out non-stationary salient stimulus sources, breaking the limitations of traditional static rule-based information filtering and avoiding the waste of computational resources on useless environmental noise for virtual characters. By constructing a perceptual nearest neighbor network and diffusion clustering based on perceptual information density, isolated perceptual variables are combined into cognitive context slices with topological relationships, enabling virtual characters to reason in conjunction with context. This invention utilizes an error correction model to decompose the calculation of underlying features into the intensity of strategic motivation transfer and the intensity of tactical reflex stimulus, creatively distinguishing the driving differences between the disruption of long-term environmental equilibrium and the short-term instinctive impact. The dual-path mechanism of macro-stress scoring and micro-stress scoring designed in this invention eliminates the network degree bias phenomenon caused by a single measurement system, neither ignoring threats of widespread global pressure nor neglecting threats of direct lethality at a single point. By iteratively reweighting the macro-stress scoring through the attention weight matrix, closed-loop feedback is achieved, forcing the system's attention to be locked on the only true root cause variable. The final output reasoning and explanation chain makes the decision-making process of the virtual character completely transparent and highly similar to the judgment logic of humans when facing a crisis, which greatly improves the rationality, explainability and adaptive survival ability of the virtual character's actions in complex and sudden environments. Attached Figure Description

[0057] Figure 1 This is a basic flowchart illustrating a method for real-time perception and action decision-making of virtual characters based on artificial intelligence, provided as an embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0059] Reference Figure 1 As an embodiment of the present invention, a method for real-time perception and action decision-making of virtual characters based on artificial intelligence is provided, comprising the following steps:

[0060] Step S1: Obtain the multimodal perception state matrix of the virtual character, divide the multimodal perception state matrix into historical stationary state windows and current abrupt state windows, perform unit root tests on the data in the current abrupt state window, extract the perceptual variables for which the first hypothesis is true, and form a set of significant stimulus sources.

[0061] Step S2: Calculate the temporal mutual information distance between perceptual variables in the set of salient stimulus sources, construct a perceptual nearest neighbor network based on the temporal mutual information distance and calculate the perceptual information density, perform diffusion clustering using the perceptual information density, and generate a set of cognitive environment slices;

[0062] Step S3: Construct an error correction model within each cognitive context slice queue, calculate the absolute value of the change in the long-term error correction coefficient between perceptual variables, and calculate the absolute value of the change in the short-term Granger causality index.

[0063] Step S4: Calculate the macro stress score based on the two absolute values ​​in step S3, and obtain the micro stress score by averaging the number of connection edges in the perceptual nearest neighbor network. Combine the macro stress score and the micro stress score to obtain the core decision-making driving force score.

[0064] Step S5: Calculate the attention weight matrix based on the core decision-driving force score, use the attention weight matrix to iteratively reweight the macro stress score, output the converged core decision-driving force score ranking, and generate the final decision action and reasoning explanation chain based on the perceptual variable corresponding to the top-ranked core decision-driving force score.

[0065] The dual-path mechanism of macro-level and micro-level stress scoring designed in this invention eliminates the network degree bias caused by a single measurement system, ensuring that neither threats exerted globally are overlooked nor those that are directly lethal at a single point are ignored. The final output reasoning and explanation chain makes the decision-making process of the virtual character completely transparent and highly replicates the judgment logic of humans facing crises, greatly improving the rationality, explainability, and adaptive survivability of the virtual character's actions in complex and sudden environments.

[0066] Step S1 includes:

[0067] The behavior time step of the multimodal perception state matrix is ​​listed as the feature type. The feature types include the virtual character's visual range perception, auditory volume perception, self-health status perception, and cooldown time status perception.

[0068] The multimodal sensing state matrix is ​​divided along the time axis to obtain a historical steady state window representing the memory of calm baseline and a current abrupt state window representing the memory of sudden events.

[0069] Within the current mutation state window, a unit root test is performed on each sensing feature time series in the multimodal sensing state matrix, and the test statistic is obtained by calculating the difference equation of the sensing feature time series.

[0070] The extended Dickie-Fowler (ADF) test was used to test the first and second hypotheses. The first hypothesis was that the time series of target-sensing features has a unit root, and the second hypothesis was that the time series of target-sensing features does not have a unit root.

[0071] The determination is made based on the comparison between the significance probability of the test statistic and the initial judgment criterion:

[0072] When the significance probability is greater than or equal to the initial judgment criterion, the first hypothesis is determined to be true, the perceptual feature time series is marked as an abnormal fluctuation state, and all perceptual feature time series in the abnormal fluctuation state are merged to form a significant stimulus source set.

[0073] When the significance probability is less than the initial judgment criterion, the second hypothesis is determined to be true, and the perceived feature time series is marked as maintaining a stable baseline state.

[0074] In this embodiment, the process of dividing the window into historical stationary state windows and current abrupt change state windows includes:

[0075] Set the current sensing moment as the time axis cursor, extract the first preset time step as the mutation evaluation period, and extract the second preset time step as the stable baseline period;

[0076] The value of the second preset time step is greater than that of the first preset time step;

[0077] Using the current sensing moment as the end point of the time axis, backtrack along the time axis for the first preset time step, extract the multimodal sensing state matrix data within this time interval, and construct the current abrupt change state window;

[0078] Taking the starting time node of the current mutation state window on the time axis as the new endpoint, backtrack along the time axis for the second preset time step, extract the multimodal sensing state matrix data within this time interval, and construct the historical stable state window;

[0079] As the virtual character's behavior time step progresses, the timeline cursor slides forward synchronously, continuously looping data capture operations to form a dynamically updated data stream of adjacent double sliding time windows in memory.

[0080] The first preset time step is preferably a frame scalar that includes the virtual character's perception behavior in the last 0.5 to 2 seconds. This value can ensure the real-time performance of edge computing while fully accommodating the minimum sample size required to perform unit root testing, so as to keenly capture minute sudden anomalies.

[0081] The second preset time step is preferably a frame scalar that includes the perceived behavior of the virtual character in the past 5 to 15 seconds. This value can effectively filter out short-term random environmental noise and provide a statistically stable and reliable calm baseline memory for subsequent comparisons.

[0082] The time sequence of sensing features is the column vector of the multimodal sensing state matrix;

[0083] The specific process of unit root testing includes:

[0084] The processing logic for performing unit root tests and generating a set of significant stimulus sources includes:

[0085] Based on any perceptual feature time series within the current mutation state window, an extended Dickie-Fowler ADF unit root test difference equation with lag terms is constructed, and its mathematical expression is:

[0086] ;

[0087] in, To perceive the first-order difference of the time series sequence at time step t, where c is a constant drift term, For time trend items, Let be the autoregressive coefficient to be tested. The system automatically selects the lag order based on the Akaike Information Criterion (AIC). The coefficient of the lagged term, It is a white noise residual sequence.

[0088] The parameters of the difference equation for the unit root test of the Extended Dickie-Fowler ADF were estimated using ordinary least squares (OLS), and the autoregressive coefficients were extracted. The estimated value and its corresponding standard error are obtained. Dividing the estimated value by the standard error yields the test statistic for the time series of the perceived feature, namely the ADF statistic.

[0089] The calculated test statistic is mapped and compared with a pre-set McKinnon critical value distribution table to calculate and extract the corresponding significance probability. The calculation process includes:

[0090] In the ADF unit root test, since its statistic does not follow a standard normal distribution or t-distribution, but rather a special Dickie-Fowler distribution, the probability must be accurately calculated using the McKinnon critical value table.

[0091] Obtain the effective sample size scalar for the difference equation calculation of the extended Dickie-Fowler unit root test within the current mutation state window;

[0092] The system calls the pre-set McKinnon critical value distribution table in the system memory. The distribution table stores the response surface regression coefficient matrix with the sample size as the independent variable under different model types.

[0093] Based on the effective sample size scalar and the response surface regression coefficient matrix, the dynamic McKinnon critical values ​​corresponding to multiple standard significance levels at the current sample scale are calculated using the finite sample adjustment formula, thus forming multiple discrete significance test threshold points.

[0094] By comparing the test statistic with the significance test threshold, the target critical value range in which the test statistic falls can be determined.

[0095] The standard significance levels corresponding to both ends of the target critical value interval are extracted as probability boundaries. A log-linear interpolation algorithm is used, combined with the proportional weight of the absolute offset distance of the test statistic within the target critical value interval and the total length of the interval, to perform approximation calculations. Discrete test threshold points are mapped to continuous probability values ​​in the interval [0,1]. These continuous probability values ​​are the corresponding significance probabilities.

[0096] The significance probability is used to measure the probability of observing the current ADF statistic under the condition that the first hypothesis of unit root is true. The larger the significance probability value, the weaker the counter-proof of the hypothesis that the current data has a unit root or is non-stationary, and the more likely the series is to show a non-stationary or structural drift tendency.

[0097] The smaller the significance probability, the stronger the evidence for the second hypothesis, and the more likely the sequence is to be in a stationary baseline state.

[0098] Referring to Table 1, the pre-set McKinnon critical value distribution table in memory, i.e., the response surface regression coefficient matrix, is shown in Table 1:

[0099] Table 1, calculated according to econometric standards and stored in read-only memory, contains three regression coefficients used to calculate critical values ​​for finite samples at different standard significance levels: the infinite sample coefficient, the first-order adjustment coefficient, and the second-order adjustment coefficient.

[0100] The specific steps for calculation and mapping comparison based on Table 1 are as follows:

[0101] Step S11: Assume that the time step of the current mutation state window captures the data of the most recent second. Calculated at a system sampling rate of 60 frames / second, the effective sample capacity scalar T for participating in the difference equation calculation is 60.

[0102] The system calls the coefficients from Table 1 and substitutes them into the finite sample adjustment formula. ,

[0103] Calculate the three dynamic critical points for the current sample size:

[0104] Critical value at the 1% significance level:

[0105] ;

[0106] Critical value at the 5% significance level:

[0107] ;

[0108] Critical value at the 10% significance level:

[0109] ;

[0110] Step S12: Assuming the test statistic of the perceived feature sequence calculated by the autoregressive difference equation is -3.2500, compare the test statistic -3.2500 with the three critical value points mentioned above. Since -3.4843 < -3.2500 < -3.1838, the target critical value range in which the test statistic falls is locked between 1% and 5%.

[0111] Step S13: Extract the probabilities of 5% and 10% at both ends of the interval, and the corresponding critical values ​​of -3.4843 and -3.1838. Use a linear interpolation algorithm to calculate the absolute offset position of the test statistic -3.2500 within this interval, and output the final continuous probability value. The expression for calculating the significance probability value is:

[0112] ;

[0113] The final significance probability extracted by the system is 0.089, or 8.9%. When 0.089 is compared with the initial judgment standard obtained by dynamic calculation, such as 0.05, since 0.089≥0.05, the first hypothesis of the feature is determined to be true and it is marked as abnormal fluctuation.

[0114] It should be noted that by disclosing the response surface regression matrix in Table 1 in the specification, this application demonstrates that it is not merely an abstract theoretical concept, but possesses low-level, executable logic. Traditional table lookup methods use static tables with a fixed sample size, which leads to significant errors when encountering sample sizes in the middle range. This application, however, uses a solidified... , and This set of response surface parameters enables the system to resolve precise critical values ​​with very small errors in a very short time, under any fluctuating mutation window length, perfectly matching the constantly fluctuating frame rate and data inflow of virtual characters in real-time dynamic rendering.

[0115] The initialization criterion is calculated based on the statistical characteristics of historical stationary state windows, and its acquisition logic includes:

[0116] Basic judgment coefficients are pre-assigned to different feature types of virtual characters, wherein the basic judgment coefficients for visual range perception and auditory volume perception are preferably 0.05, and the basic judgment coefficients for self-health status perception are preferably 0.01.

[0117] Extract the historical variance values ​​of the corresponding perceptual features from the historical stationary state window, and substitute these historical variance values ​​into the smoothing correction function to dynamically fine-tune the basic determination coefficients. The calculation process includes:

[0118] When the historical variance value is greater than the preset noise threshold, it means that there is a noisy background. The basic judgment coefficient is increased according to the preset amplification ratio to make the judgment conditions more stringent to prevent accidental touch.

[0119] When the historical variance value is less than or equal to the preset noise threshold, it means that there is an absolutely quiet background. The basic judgment coefficient is reduced according to the preset attenuation ratio to relax the judgment conditions and improve the sensitivity.

[0120] The fine-tuned value is output as the initialization criterion for this perceptual feature at the current moment.

[0121] The logical premise for hypothesis testing is that the time series is in a stationary state.

[0122] Referring to Table 2, the dynamic calculation parameter allocation table for the initial judgment criteria of virtual character perception features is used to extract the variance values ​​of each feature in real time from the historical stable state window before each round of perception judgment and to execute a smoothing correction strategy.

[0123] The calculation logic based on the smoothing correction function in Table 2 includes:

[0124] False alarm suppression in noisy environments, taking visual range perception as an example:

[0125] In the current sliding time step, the visual feature stream of the virtual character in the past stable state window within 10 seconds is extracted. Assuming that the character is in an environment with wind and rain and the grass is shaking violently, the measured value of the visual variance in this historical window is calculated to be 15.5.

[0126] Compare it with the preset noise threshold of 10.0 for visual features. Since 15.5 > 10.0, the current visual environment is determined to be in a noisy and disturbed state.

[0127] The preset amplification rule in the smoothing correction function is triggered, with a preferred amplification ratio of 1.5, to fine-tune the basic judgment coefficient:

[0128] ;

[0129] The final initial criterion for this visual feature was tightened, increasing from 0.05 to 0.075. This means that in subsequent comparisons, only visual sequences with a significance probability greater than or equal to 0.075 will be considered abnormal. The system thus successfully filters out common visual noise from windy and rainy environments, avoiding frequent false alarms from the virtual character.

[0130] It should be noted that traditional fixed-threshold testing systems, such as rigidly using 0.05 as the significance threshold for all features, cannot adapt to the complex scenarios of rapid transitions in games or embodied intelligence, easily falling into logical dead ends like being paranoid in a bustling city or reacting sluggishly in a dark room. This application, by disclosing a variance-driven dynamic testing criterion adjustment matrix as shown in Table 2, mathematically closely fits the psychological alertness compensation mechanism of the real biological brain. This mechanism allows each independent feature in the multimodal perception state matrix to adaptively adjust the boundary of abnormal triggering based on its recently experienced historical noise baseline. This completely breaks free from the constraints of static thresholds on real-time decision-making systems, endowing virtual characters with strong environmental adaptability.

[0131] The significance probability of the perceived feature time series is compared with the calculated initial judgment criteria:

[0132] When the significance probability is greater than or equal to the initial judgment criterion, the first hypothesis is valid in a statistical sense. The target's perceptual characteristics have been impacted by an external sudden event and exhibit non-stationary characteristics, and are marked as an abnormal fluctuation state.

[0133] When the significance probability is less than the initial judgment criterion, the data evolution trend of the target perception feature within the current mutation state window is judged. Compared with the historical stable state window, there is no structural mutation or significant mean drift, and it tends to be in a stable baseline state.

[0134] The temporal sequence of the perceived features is defined as safe environmental noise or meaningless redundant data, and it is filtered out from the current processing link and not included in the set of salient stimulus sources.

[0135] The system directly releases the computational resources for subsequent cognitive slice clustering and inference related to the perceived feature, while continuing to perform silent listening and routine standby patrol actions.

[0136] Traverse all column features in the multimodal perception state matrix, extract all perception features marked as abnormal fluctuation states, merge them to form a set of salient stimulus sources, and input them into the next cognitive environment slice clustering step.

[0137] It should be noted that traditional technologies, when handling anomaly perception in virtual characters, typically employ a crude judgment method based on the absolute magnitude of the current frame's feature value exceeding a fixed preset threshold. This approach is prone to frequent false alarms in complex environments with significant background noise, such as near waterfalls or in wind and rain, while in quiet environments, it may miss subtle danger signals, such as extremely faint footsteps. This application creatively introduces unit root tests from econometrics and dynamic calculation initialization criteria. Its core advantage lies in focusing not on the absolute magnitude of values ​​but on the disruption of the stationarity of the data sequence's evolutionary trend. Even against a background of violently shaking grass—i.e., a high-variance stationary sequence—as long as no new structural anomalies appear, a high significance probability can be calculated to filter it out. Combined with an initialization criterion that adaptively adjusts historical variance, the model is endowed with a human-like attentional alertness regulation mechanism, perfectly solving the industry challenge of noise stripping and real threat extraction in complex multimodal environments.

[0138] It should be noted that traditional techniques for anomaly detection in time series often simply apply the standard normal distribution or Student's t-distribution to obtain probabilities when statistical tests are involved. However, in non-stationary time series data, this traditional approach produces severe spurious regression, resulting in artificially low probabilities and consequently frequent erroneous actions by virtual characters. The method adopted in this application, which dynamically calculates the McKinnon critical value based on a finite sample scalar and performs interpolation mapping, strictly adheres to the true probability density function of non-stationary time series. It adaptively adjusts the rigor of the test according to the physical length of the current mutation window and the sample size. This logic effectively eliminates the statistical bias caused by small sample windows, ensuring that the extracted significance probabilities maintain high mathematical rigor even within extremely short time windows (e.g., 0.5 seconds), greatly improving the accuracy of the model in identifying sudden events.

[0139] It should be noted that traditional technologies for processing the perceived state of virtual characters typically rely on fixed-length queue caching along a single straight line or truncated updates based on fixed thresholds. This approach suffers from limitations such as the inability to create dynamic references over time and the difficulty in distinguishing between gradual changes in the overall environment and sudden, localized stimuli. In contrast, the update method employed in this application, which uses a time-axis cursor to backtrack and segment adjacent double sliding windows in real time, adaptively constructs a data topology structure of a long-period stable background and a short-period dynamic foreground without incurring additional data storage overhead. This structure provides a rigorous reference benchmark for subsequent unit root tests, effectively addressing the problem of misjudgment and overreaction in traditional virtual characters under complex and continuous visual or auditory perturbations. Furthermore, it accurately isolates regular background fluctuations, demonstrating a significant improvement in the anti-interference capability and robustness of the real-time perception mechanism in responding to unexpected events.

[0140] Existing technologies typically use fixed data boundary values ​​to segment and filter perceived information, but this method cannot cope with complex and ever-changing environmental background noise disturbances. This application uses unit root tests to monitor the stationarity of continuous time series in real time. Stationarity indicates that the current variable fluctuates within a normal background range, while rejection of the stationarity hypothesis indicates that the variable has encountered sudden anomalies. This processing method eliminates the limitations of static judgment and implements an adaptive event-triggered perception mechanism, including only signals that truly cause anomalies in the set of salient stimuli.

[0141] Step S2 includes:

[0142] In a set of salient stimulus sources, the temporal mutual information distance, which measures the degree of overlap of sequence information, is obtained by calculating the probability distribution of any two perceptual variable sequences and obtaining their mutual information.

[0143] Based on the temporal mutual information distance, the K nearest neighbor nodes of each perceptual variable are extracted to form a nearest neighbor node set. When the first perceptual variable is located in the nearest neighbor node set of the second perceptual variable, and the second perceptual variable is also located in the nearest neighbor node set of the first perceptual variable, a nearest neighbor bidirectional connection edge is constructed between the first perceptual variable and the second perceptual variable. All nearest neighbor bidirectional connection edges are merged to obtain the perceptual nearest neighbor network.

[0144] The number of nearest-neighbor bidirectional connections included in each sensing variable in the statistical sensing nearest neighbor network is used as the sensing information density of that sensing variable.

[0145] In this embodiment, the specific steps for calculating the temporal mutual information distance include:

[0146] For any two perceptual variable sequences X and Y in the set of salient stimulus sources, extract all feature observations within the current mutation state window; the perceptual variables are the perceptual feature time series sequences.

[0147] The histogram interval partitioning method is used to divide the numerical distribution interval of the feature observations into several discrete data buckets;

[0148] The frequency of the characteristic observations of the perceived variable sequences X and Y falling into each data bucket is counted separately. After frequency normalization, the marginal probability distributions p(x) of the perceived variable sequence X and p(y) of the perceived variable sequence Y are calculated.

[0149] Simultaneously, the frequency of the characteristic observations of both falling into the corresponding data bucket at the same time step is counted, and the joint probability distribution p(x,y) is calculated.

[0150] Based on the marginal probability distribution, the information entropy H(X) of the perceived variable sequence X and the information entropy H(Y) of the perceived variable sequence Y are calculated respectively to measure the amount of uncertain information contained in each sequence.

[0151] Based on marginal probability distribution and joint probability distribution, the temporal mutual information between two sequences is calculated according to Shannon's information theory criterion. Temporal mutual information is used to quantify the amount of information shared between two perceptual variable sequences. The larger the value, the more synchronized the mutation trend and distribution pattern of the two heterogeneous modal variables are in the time series.

[0152] In order to transform the temporal mutual information that measures similarity into a distance metric that measures the proximity of network topologies, the temporal mutual information is normalized and inversely mapped.

[0153] The temporal mutual information distance is obtained by calculating the ratio of the temporal mutual information to the sum of the information entropies of the two sequences, and then subtracting twice this ratio from a constant of 1. Under this mapping, the temporal mutual information distance is strictly constrained to the interval between a constant of 0 and a constant of 1. The closer the distance value is to 0, the higher the degree of information overlap between the sequences, and the closer they are in the topological space; the closer the value is to 1, the more independent the information between the sequences.

[0154] By iterating through all perceptual variables in the set of salient stimulus sources, calculating the temporal mutual information distance pairwise, a globally symmetric temporal mutual information distance matrix is ​​constructed.

[0155] For any perceptual variable in the set, extract the corresponding row distance vector from the temporal mutual information distance matrix;

[0156] Sort the temporal mutual information distance values ​​in the distance vector in ascending order and exclude elements whose distance to itself is 0;

[0157] Based on the initial hyperparameter K of the number of neighboring nodes, select the K perceptual variables with the smallest distance values ​​from the permutation sequence;

[0158] The K extracted perceptual variables are combined to form the nearest neighbor set of the perceptual variable.

[0159] It should be noted that traditional techniques for calculating the similarity of multimodal variables typically use Euclidean distance or Pearson correlation coefficient. However, in the complex perception of virtual characters, multimodal data not only have completely different dimensions, but also often exhibit complex nonlinear causal delay relationships due to differences in physical propagation characteristics. Linear correlation coefficients simply cannot capture this implicit association.

[0160] This application creatively introduces temporal mutual information distance as a criterion for constructing perceptual nearest neighbor networks. It completely breaks free from the limitations of absolute numerical magnitude and linear correlation. As long as two perceptual variables co-move along the time axis—for example, the visual frequency of swaying grass and the auditory rhythm of extremely faint footsteps, although different in absolute value, have overlapping information distributions—their mutual information in probability distribution will be extremely high, thus bringing them closer in the topological space and incorporating them into each other's K nearest neighbor sets. This mathematical mechanism breaks down the barriers between physical modalities, perfectly realizing the spontaneous semantic binding of isolated features such as swaying grass and local environmental darkening, laying an extremely reliable network topological foundation for subsequently generating intelligent cognitive environment slices that completely break away from rigid rules.

[0161] Calculate the mean of the perceived information density of all perceived variables in the set of salient stimulus sources. When the perceived information density of a perceived variable is greater than the mean of the perceived information density, it is determined that the perceived variable is in a high-density complex environment and a first diffusion boundary value is assigned to it.

[0162] When the perceived information density of a perceived variable is less than or equal to the mean perceived information density, the perceived variable is determined to be in a low-density isolated environment, and a second diffusion boundary value is assigned to it.

[0163] The first diffusion boundary value is less than the second diffusion boundary value;

[0164] The system searches for shared neighbor nodes outward from the center of the perceptual variable. When the number of shared neighbor nodes is greater than the corresponding diffusion boundary value, the neighbor nodes are grouped into the same cognitive environment slice queue until the search process ends. The output is a set of cognitive environment slices composed of interconnected perceptual variables.

[0165] In this embodiment, the dynamic calculation and processing logic for the neighbor node number hyperparameter K and the diffusion boundary value includes:

[0166] The adaptive calculation of the number of nearest neighbor nodes K includes:

[0167] Extract the scalar N representing the total number of perceptual variables contained in the set of salient stimulus sources within the current time step;

[0168] To avoid excessively large local neighborhoods when there are many variables, or neighborhood failures when there are few variables, the hyperparameter K is defined as a dynamic parameter that has a non-linear proportional relationship with the total size of the set. Its calculation expression is as follows:

[0169] ;

[0170] in, As a floor operator, when N < 2, K = 0 and the nearest neighbor network construction step is skipped; this formula ensures a reasonable number of basic neighbor nodes for each perceived variable in the network, and smoothly expands the nearest neighbor search range by the square root as the total number of environmental anomaly variables N increases, while avoiding calculation overruns when the number of nodes is too small.

[0171] A shared neighbor node is the intersection of nodes in the nearest neighbor network that simultaneously exist in the nearest neighbor set of the first perceptual variable M and the nearest neighbor set of the second perceptual variable A.

[0172] When performing diffusion clustering search, the search extends outward from the first perceptual variable M to the second perceptual variable A, which has a direct connection edge with it. The number of shared neighbor nodes between the two is calculated. The number of shared neighbor nodes is used to measure the degree of environmental coupling between the two perceptual variables in the local network topology space.

[0173] Since the first diffusion boundary value is assigned to the perceptual variable in a high-density complex situation, the second diffusion boundary value is assigned to the perceptual variable in a low-density isolated situation, and the first diffusion boundary value is smaller than the second diffusion boundary value.

[0174] The first diffusion boundary value is dynamically calculated based on the number of neighboring nodes K. With the second diffusion boundary value Its mathematical expression is:

[0175] ;

[0176] ;

[0177] in, The first polymerization coefficient, The second polymerization coefficient is preferred; the preferred range for the first polymerization coefficient is 0.2 to 0.4, and the preferred range for the second polymerization coefficient is 0.6 to 0.8, because... This ensures that, mathematically, it is strictly guaranteed. .

[0178] When searching outwards to absorb nodes, in a high-density and complex situation where the perceived variable is, the number of shared neighbor nodes only needs to be greater than the smaller one. It can be absorbed;

[0179] When the perceived variable is in a low-density, isolated context, the number of shared neighbor nodes must be greater than a large number. Only then can it be absorbed.

[0180] It should be noted that traditional unsupervised clustering algorithms, such as K-Means or standard DBSCAN, typically rely on globally fixed distance thresholds or fixed density requirements. When processing virtual character perception data, this approach often tears apart subtle features in high-density areas or incorrectly merges irrelevant noise in low-density areas. This application creatively introduces an adaptive clustering topology based on a shared neighbor node structure and a dual-path dynamic diffusion boundary. When perceptual variables are in high-density, complex scenarios, such as the center of a battlefield or the site of an explosion, the system automatically assigns a smaller first diffusion boundary value, lowering the node absorption threshold and making aggregation requirements more lenient. This encourages a large number of fragmented sensory anomalies to quickly aggregate into a single, large, and complete slice of the cognitive environment, preventing the crisis scene from being fragmented and analyzed. When in low-density, isolated scenarios, the system assigns a larger second diffusion boundary value, making the node absorption threshold extremely high and the aggregation requirements extremely stringent. Isolated variables must have a very high proportion of shared neighbors to be allowed to be associated, thus perfectly severing occasional false connections between sporadic environmental noises. This design completely frees the model from manual parameter tuning and gives it a strong intelligent analytical and adaptive capability to complex and ever-changing 3D virtual environments.

[0181] Step S3 includes:

[0182] Extract all perceptual variables from a single cognitive situation slice queue in the cognitive environment slice set, and establish a vector autoregressive equation based on all perceptual variables;

[0183] Extract the cointegration relationship equation from the vector autoregression equation, calculate the change in the long-term error correction coefficient between the current abrupt state window and the historical stationary state window, and take the absolute value to obtain the first absolute value intensity.

[0184] Significance verification calculations are performed on the lagged terms in the vector autoregression equation. The short-term Granger causality index change caused by the introduction of the current mutation state window data is calculated and its absolute value is taken to obtain the second absolute value strength.

[0185] In this embodiment, the processing logic for constructing an error correction model and calculating the absolute value intensity within a single cognitive context slice queue includes:

[0186] Extract all perceptual variables from a single cognitive context slice queue within the cognitive environment slice set, and construct a multidimensional time series vector. ;

[0187] Based on the Akaike Information Criterion (AIC), the optimal lag order p is determined, and a standard vector autoregressive equation is constructed using the least squares method. Its mathematical expression is as follows:

[0188] ;

[0189] in, Let be the vector of perceived variables at time t. It is a constant intercept vector. Let be the lag coefficient matrix of order k. It is a white noise random error vector.

[0190] The process of calculating the constant intercept vector C and the k-th order hysteresis coefficient matrix includes:

[0191] Multidimensional perceptual feature time-series observation data are extracted from the cognitive context slice queue, and a matrix-form multiple linear regression equation is constructed. Its mathematical expression is as follows:

[0192] ;

[0193] in, For the dependent variable matrix, This includes the constant term 1 and lag terms of each order. The independent variable design matrix is ​​given by U, the residual matrix is ​​given by B, and the parameter matrix to be estimated is given by B.

[0194] The parameter estimates that minimize the sum of squared residuals are obtained using ordinary least squares (OLS). The calculation expression is as follows:

[0195] ;

[0196] The obtained parameter matrix The slice is segmented, and its first column is the constant intercept vector C, corresponding to each... The dimensional matrix block is the lag coefficient matrix of each order. , where k=1,2,...,p.

[0197] The optimal parameter estimate Design matrix with independent variables By performing matrix multiplication, the theoretical prediction matrix of the model for multidimensional perceptual features is obtained. ,Right now Subtract the theoretical prediction matrix from the actual obtained dependent variable observation matrix Y. The difference matrix obtained by subtracting the two is the residual matrix U.

[0198] Perform the Johansen cointegration test on the multidimensional time series vector to verify whether there is a long-term equilibrium state among the perceived variables;

[0199] Given the existence of cointegration, a difference transformation is performed on the vector autoregressive equation to extract the cointegration equation for the vector error correction model. Its mathematical expression is as follows:

[0200] ;

[0201] in, It is a first-order difference vector. The long-term impact matrix is ​​decomposed into: , This is the error correction coefficient matrix, representing the adjustment and recovery rate when variables deviate from long-term equilibrium. The cointegration vector matrix represents the long-term equilibrium relationship. This is a short-term dynamic adjustment matrix;

[0202] The lag coefficient matrices of each order are transformed by polynomial algebraic transformation. Transform from a level value model to a difference model. Long-term influence matrix. The calculation logic is the sum of all lag coefficient matrices minus the identity matrix I, and its expression is:

[0203] ;

[0204] Short-term dynamic adjustment matrix The calculation logic is the negative sum of the lag coefficient matrices from the (k+1)th order to the highest order p, and its expression is:

[0205] ;

[0206] The long-term influence matrix was estimated using the Johansen maximum likelihood estimation method. The process of performing a reduced-rank decomposition to calculate E and R specifically includes:

[0207] The first-order difference vectors are respectively and lag level vector An auxiliary regression is performed on the short-term lagged difference terms to extract the residual vectors from the two regressions, and the residual covariance matrix of the residuals is constructed from these vectors. , , and ;

[0208] Based on the residual covariance matrix, construct and solve the generalized eigenvalue equation:

[0209] ;

[0210] in, Let be the eigenvalues ​​to be solved;

[0211] Sort the solved eigenvalues ​​that are greater than zero in descending order to obtain Extract the eigenvectors corresponding to these r significant eigenvalues, and combine these eigenvectors column by column to form the cointegration vector matrix R representing the long-term equilibrium relationship of the system.

[0212] Using the obtained cointegration vector matrix R and residual covariance matrix, the error correction coefficient matrix E, representing the adjustment recovery rate when deviating from equilibrium, is calculated by matrix multiplication. Its calculation expression is as follows:

[0213] ;

[0214] And mathematically, it strictly satisfies the decomposition relation. .

[0215] Data from the historical stationary state window and the current abrupt state window are extracted separately, and the cointegration relationship equation is estimated independently. The historical error correction coefficient matrix and the current error correction coefficient matrix are extracted separately.

[0216] Extract the corresponding matrix elements of perceived variable j to perceived variable i, calculate the difference between the current error correction coefficient and the historical error correction coefficient, and take the absolute value to obtain the long-term error correction coefficient change, i.e. the first absolute value intensity. The closer the first absolute value intensity is to zero, the more it represents that the long-term survival equilibrium has not been affected. The larger the value, the more it is assessed that the perceived variable has a deeper degree of damage to the survival equilibrium of the virtual character over the time span.

[0217] In the short-run dynamic adjustment term of the cointegration equation (i.e. In the calculation, the Wald joint significance test is performed to determine the predictive contribution of the lagged difference term of the perceived variable j to the current state of the perceived variable i.

[0218] The chi-square statistic value of the output of the Wald joint significance test is extracted and used as the short-term Granger causality index, which is used to measure the ability of lagged data to directly predict the current state in a very short time.

[0219] Similarly, the historical short-term Granger causality index was calculated independently using historical stationary state window data and current abrupt change state window data, respectively. With the current short-term Granger causality index ;

[0220] Calculate the current exponent of perceived variable j pointing to perceived variable i. Compared with historical index The difference between the two values ​​is taken as the absolute value to obtain the short-term Granger causality index change, i.e., the second absolute value intensity. The second absolute value intensity is used to represent the degree of instantaneous pulse impact and instinctive reflex triggering caused by each perceptual variable to the virtual character's state within a preset extremely short time window.

[0221] In the difference equation for the target perceived variable i, the historical data coefficients of all short-term dynamic adjustment terms originating from the perceived variable j are extracted. ;

[0222] The null hypothesis is that all the above lag coefficients are combined to zero, which means that variable j has no short-term predictive power for i.

[0223] The Wald test statistic matrix is ​​constructed based on the residual covariance matrix, and its mathematical expression is as follows:

[0224] ;

[0225] in, Let j be the column vector of short-term coefficients of the extracted variable j. Estimate its covariance matrix. The real value of the calculated Wald test statistic W represents the predictive contribution of the target feature dependent variable to the fluctuation of the source feature independent variable. The larger this value, the higher the probability of the null hypothesis being rejected, indicating a greater intensity of short-term impulse shocks.

[0226] It should be noted that existing models, when processing multimodal sensing data, often only perform simple linear smoothing or end-to-end black-box regression, ignoring the non-stationary nature of time series and easily producing serious spurious regressions. That is, two completely unrelated variables are incorrectly judged to be causally related simply because they both show a trend of increasing over time. The calculation process described in this application, through the reduced-rank decomposition of E and R, completely separates the random walk trend in the data sequence from the constraints of real physics at the underlying mathematical logic level. Matrix R accurately locks in the long-term binding relationship of simultaneous rises and falls between physical modes, while matrix E quantifies the strength and direction of the restoration efforts of each sensing variable to pull the system back to the equilibrium baseline when the system suffers sudden damage. This modeling method, which delves into the eigenvalue solution level, gives the absolute strength of the generated system absolute mathematical authority, fundamentally eliminating logical misjudgments by virtual characters in complex and noisy environments.

[0227] It should be noted that traditional artificial intelligence, when processing the correlation between multimodal sequences, typically uses attention mechanisms or simple Pearson correlation coefficients to directly calculate the weight of a single feature. This black-box approach confuses long-term strategic threats with instantaneous tactical shocks, causing virtual characters to easily make inverted judgments when faced with short-term impulse shocks and long-term equilibrium disruptions.

[0228] This application creatively introduces the error correction model VECM from econometrics, mathematically separating the causal relationships between perceptual variables through long-wavelength and short-wavelength dimensionality reduction. By using long-term error correction coefficients, variables that are usually inconspicuous but completely disrupt the underlying balance of the role system are identified and categorized as strategic motivations; by using short-term Granger causality indices, variables that suddenly explode in value and trigger neural reflexes are identified and categorized as tactical reflexes. This method breaks through the black-box limitations of end-to-end weights in deep learning, endowing virtual characters with a white-box cognitive deconstruction capability that combines human-like subconscious defense and stress reflexes.

[0229] Step S4 includes:

[0230] The formula for calculating the macro-stress score is:

[0231] ;

[0232] in, For the perceived variable i, the macro-stress score. Let i be the set of neighbor nodes connected to the sensing variable i in the nearest neighbor network. Let be the intensity of the first absolute value of the perception variable i for neighbor node j. Let be the intensity of the second absolute value of the perception variable i of neighbor node j. The initial macro- and micro-distribution coefficients are taken as constants between 0 and 1.

[0233] A macroeconomic stress score is calculated for any perceived variable. The macroeconomic stress score is used to quantify the global oscillation caused by the perceived variable.

[0234] The formula for calculating the micro-stress score is:

[0235] ;

[0236] in, For the micro-stress score of perceived variable i, It is a scalar representing the total number of nodes in the neighbor node set.

[0237] A micro-stress score is calculated for any perceived variable, which eliminates the influence of the number of neighboring nodes to quantify the lethality of a single causal edge.

[0238] When the set of neighbor nodes of the perceived variable i in the nearest neighbor network When empty, define and And in the calculation of the attention weight matrix, let A self-loop is formed to ensure that the iterative reweighted computation can be executed.

[0239] The formula for calculating the core decision-driving force score is:

[0240] ;

[0241] in, The core decision-driving force score for perceived variable i. The logistic smoothing activation function maps real numbers to the interval from constant 0 to constant 1. To perform the macro stress score after minimax normalization, To perform the micro-stress score after minimax normalization, and These are the first mapping weight and the second mapping weight, respectively, and B is the mapping bias constant.

[0242] In this embodiment, the initialized macroscopic and microscopic allocation coefficients First mapping weight Second mapping weight Mapping bias constant and temperature coefficient scalar The acquisition and dynamic calculation processing logic includes:

[0243] During the initial deployment phase of the virtual character, the system performs a constraint interval initialization operation on the aforementioned undetermined constants, which aligns with biological intuition:

[0244] Macro- and micro-distribution coefficients Constrained within the interval [0,1], it is used to characterize the personality perspective of a character. The closer the value is to 1, the more the character's personality focuses on global macro pressure. The closer the value is to 0, the more the character's personality focuses on single point micro pressure.

[0245] First mapping weight With the second mapping weight The constraint is a real number greater than zero, used to characterize the neural sensitivity to dual-path pressure;

[0246] The mapping bias constant B is constrained in the real domain and is used to set the activation baseline of the core decision-making drivers.

[0247] It should be noted that traditional graph node evaluation is easily affected by the degree bias of the number of nodes connected, resulting in nodes with more connections having inflated scores. The dual-path parallel macro and micro stress scoring design of this application not only identifies the range stress features with wide correlation, but also accurately locks the micro stress features with only a single connected edge but extremely high intensity, completely eliminating the degree bias loophole.

[0248] Step S5 includes:

[0249] The attention weight matrix is ​​calculated based on the core decision-driving force scores of neighboring nodes of perceived variable i. The calculation expression is as follows:

[0250] ;

[0251] in, Let be the attention weights of the perceived variable i pointing to its neighbor node j in the attention weight matrix. The score representing the core decision-making driving force of neighbor node j. Let e ​​be the core decision-driving force score of all nodes within the neighbor node set, and e be the natural constant. A temperature coefficient scalar that maximizes the smoothness of the distribution;

[0252] Substituting the attention weights into the macro stress score for iterative reweighting calculation, the macro stress score is updated. The mathematical expression for updating the macro stress score is:

[0253] ;

[0254] in, For the updated macro-stress score after one iteration, all perceived variables are... After performing max-min normalization, the calculation expression for the core decision-driving score is re-inputted to replace the macro-stress score, and the core decision-driving score is updated.

[0255] Construct a high-dimensional parameter vector P=[ , , , , ].

[0256] After the virtual character performs real-time perception and action decisions, the corresponding multimodal perception state matrix and actual survival result benefits are recorded. The actual survival result benefits include the percentage of damage avoided by the character, survival time, and energy consumption rate.

[0257] The actual survival outcome benefit is constructed as a target fitness function. The higher the value of this function, the more favorable the decision made by the character is to survive under the current parameter combination P.

[0258] During the inactive standby period of the virtual character, including the hibernation and background calculation phases, the Henry's Gas Solubility Optimization Algorithm (HGSO) is invoked to perform global optimization calculations on the parameter vector set. The process includes:

[0259] Different parameter combinations P are instantiated into multiple gas molecule populations under specific temperatures and pressures to simulate the evolutionary clusters of different character subclasses.

[0260] The solubility of each gas molecule population is calculated by substituting it into the target fitness function, and the gas is divided into multiple clusters according to the gas type. The local optimal molecule and the global optimal molecule in each cluster are extracted.

[0261] Based on the mathematical model of Henry's Law Where C is the solubility and H is the Henry's constant. To simulate the influence of external survival pressure on the evolution of the gas molecule, the temperature scalar T and Henry's constant H of each gas molecule are calculated and updated.

[0262] Based on a temperature scalar cooling decay strategy, the position transfer equation of gas molecules is executed, and poor parameter combinations are moved toward local and global optimal parameter combinations. Its iterative update expression covers the search step size adjustment in the exploration and development phases.

[0263] When the fitness function value does not improve continuously or reaches the preset maximum number of evolution iterations of 100, the calculation of the Henry's gas solubility optimization algorithm (HGSO) is stopped.

[0264] Extract the position coordinates of the gas molecule with the highest global fitness, and unpack it to restore it to a high-dimensional parameter vector P;

[0265] This set of optimal solutions is overwritten into the system memory and used as the macro- and micro-allocation coefficients, first mapping weights, second mapping weights, and mapping bias constants initialized at the start of the next active cycle.

[0266] Referring to Table 3, here is the data table of the evolution and iteration of the virtual character's background personality parameters:

[0267] It should be noted that traditional deep learning or reinforcement learning systems typically have fixed fusion weights after deployment, or require massive computing power for full backpropagation updates. This static mechanism leads to rigid characterization, making it impossible to adjust response strategies to dynamically changing threat patterns.

[0268] This application creatively utilizes the Henry's Gas Solubility Optimization Algorithm (HGSO) to ingeniously transform the calculation of hyperparameters into an adaptive evolutionary process of personality variation driven by environmental pressure. This mechanism eliminates the need for complex gradient differentiation, making it extremely suitable for lightweight background computations at the edge. The macroscopic and microscopic allocation coefficients and mapping biases continuously approach the optimal solution of the current environment through countless standby cycles, enabling the virtual character to exhibit highly embodied intelligence characteristics.

[0269] The core decision-driving force score is continuously updated by iterative reweighting calculation. After each iteration, the absolute value of the difference between the current iteration's core decision-driving force score and the previous iteration's core decision-driving force score is calculated, and the sum of all absolute values ​​of differences during the iteration process is calculated.

[0270] When the sum of the absolute values ​​of the differences is less than the initial convergence minimum constant, stop the iterative calculation, extract the core decision driving force scores of all perceptual variables and sort them in descending order, and output the ranking of the converged core decision driving force scores.

[0271] Extract the perceptual variables corresponding to the top-ranked core decision-making driving force score, and set the perceptual variables as the root cause of the final inference;

[0272] The motion engine module of the virtual character is invoked, and the corresponding physical avoidance response coordinates are output based on the final inference root cause. The virtual character is then controlled to move towards the safe coordinate area to generate the final decision action.

[0273] In this embodiment, the initial convergence minimum constant is a pre-allocated floating-point precision threshold, which is dynamically calculated based on the underlying hardware precision and the local network topology scale.

[0274] Extract the minimum floating-point machine precision value of the underlying computing devices NPU and GPU running this virtual character instance. For single-precision floating-point numbers, it is typically approximately [value missing]. ;

[0275] Extract the total number of nodes in the neighbor node set of the current target perception variable as a scalar, multiply the machine precision minimum by the square root of the neighbor node number scalar and the system preset tolerance relaxation coefficient, preferably 10.0, and calculate the current convergence minimum constant.

[0276] After each iteration of the attention weight matrix reweighting calculation of the macro stress score, the sum of the absolute values ​​of the differences between the core decision-driving force scores of all nodes in the current iteration and the previous iteration is obtained.

[0277] When the sum of the absolute values ​​of the differences is less than the initial convergent minimum constant, it is determined that the network computing power has fully focused its attention, the mathematical evolution has reached its limit, and an iteration stop signal is triggered.

[0278] Extract the perceptual variable corresponding to the top-ranked core decision-driving force score after stopping iteration, set it as the final inference root cause, analyze the three-dimensional spatial attributes of the final inference root cause, and extract its pressure source location coordinate vector in the virtual world coordinate system.

[0279] Call the motion engine module of the virtual character to obtain the current position coordinate vector of the character, calculate the escape direction vector from the threat source to the character and normalize it;

[0280] By combining the virtual character’s current movement speed scalar with the navigation grid boundary constraints, the physical evasion response coordinates are calculated along the normalized escape direction vector.

[0281] The physical avoidance response coordinates are input into the character's underlying motion controller, driving the character to move towards the safe coordinate area and generating the final decision action.

[0282] The inference explanation chain generated based on the perceptual variables corresponding to the top-ranked core decision-making driver score includes:

[0283] The steps of source tracing reasoning are constructed using a natural language processing model, and these steps specifically include:

[0284] Step S51: Concatenate the unit root test state reversal information to generate the perception trigger paragraph, and output the first interpretation paragraph;

[0285] Step S52: Concatenate the perceived information density value and the target diffusion boundary value to generate a context-related paragraph, and output the second explanatory paragraph;

[0286] Step S53: Concatenate the changes in long-term error correction coefficients to generate the expected deviation paragraph, and output the third explanatory paragraph;

[0287] Step S54: Concatenate the short-term Granger causality index change to generate the primacy reflex trigger paragraph, and output the fourth explanatory paragraph;

[0288] Step S55: After converging, the attention weights approach a constant of 1 to generate a closed-loop locking paragraph, and output the fifth explanatory paragraph;

[0289] Step S56: Combine all explanatory paragraphs to output the inference chain that follows the chronological causal order.

[0290] The system invokes a pre-defined slot-filling natural language template, extracts core mathematical indicators from the end-to-end computation process, performs text entity mapping, and generates five explanatory paragraphs in sequence:

[0291] Extract the test statistic, significance probability, and dynamic initialization criterion from the unit root test, and fill them into the first explanatory paragraph: "At the current moment, the significance probability of feature [X] is [P-value], which is greater than or equal to the dynamic threshold [standard significance level]." "The first statistical hypothesis is valid, indicating that non-steady-state anomalies have been detected, triggering an abnormal fluctuation alarm."

[0292] Extract the temporal mutual information distance, the number of nearest neighbor nodes, and their perceived information density, and fill them into the second explanatory paragraph: "The perceived information density of this feature is [perceptual information density value], [greater than / less than] the global mean, and is judged as a [high-density complex / low-density isolated] environment, and a diffusion boundary value is assigned []. / [Value], and then incorporate shared neighbor nodes, binding them with [interconnected perceptual variables] into a slice of the cognitive environment.

[0293] Extract the first absolute value intensity and fill it into the third explanatory paragraph: "Based on the error correction model calculation, the long-term error correction coefficient of this slice abruptly changes to [ ] ." This indicates that the variable has caused profound strategic disruption to the long-term survival equilibrium of the roles.

[0294] Extract the second absolute value intensity and fill it into the fourth explanatory paragraph: "At the same time, the short-term Granger causality index surged to [ This created a strong, instantaneous pulse impact, directly triggering tactical instinctive reflexes.

[0295] Extract the convergent attention weights at the end of the iteration and fill them into the fifth explanatory paragraph: "After [number of iterations] stress score iterations and reweighting, the attention weights for this feature will converge to a value close to a constant 1." This completely shields off other redundant environmental noise, achieving a closed-loop lock-on against this single, lethal threat.

[0296] The entity segments of the above five dimensions are cascaded and combined in the order of occurrence and causal deduction. The final decision action "Therefore execute the instruction to move to the coordinate [physical evasion response coordinate]" is concatenated at the end through the action engine interface, outputting a complete and transparent reasoning explanation chain to humans.

[0297] It should be noted that existing embodied intelligence or game NPC action decisions largely rely on deep reinforcement learning models, which are essentially a huge black box. When a virtual character suddenly performs an action, neither the developers nor the human players can know whether it is because they heard a sound, saw a light or shadow, or because of a bug in the system itself.

[0298] The final step of this application creatively achieves a dimensionality reduction translation from white-box mathematical reasoning to natural language. Not only does it endow the character with physical evasion capabilities through an action engine, but more importantly, through steps S51 to S56, it seamlessly translates the previously obscure unit root tests, topological diffusion boundaries, and long-term error correction matrices—abstract mathematical variables—into human-readable text with rigorous temporal and causal logic. This innovative effect makes the virtual character seem to possess realistic cognitive feedback capabilities. After making a decision, it can provide humans with a perfect tracing report containing the cause, environmental context, long-term motivation, short-term reflexes, and final focus, greatly improving the transparency, trustworthiness, and efficiency of backend debugging in AI interaction.

[0299] It should be noted that this step introduces a closed-loop reweighting mechanism for the core decision-making drive. Through repeated mathematical iterations, the attention weight of environmental noise features is converged to zero, realizing a high-level biological cognitive process similar to how humans focus on the single fatal threat and block out irrelevant factors when facing a moment of fright. This gives the character generation a breakthrough ability to trace a long logical chain of origin.

[0300] In the technical architecture of the 3D virtual environment, the 3D virtual environment includes games and an embodied intelligent simulation platform. The motion engine module is deployed at the bottom layer of the 3D virtual environment. It includes a low-level physical framework built based on Unity, Unreal Engine and ROS simulation. It is used to receive upper-level artificial intelligence decision instructions and transform them into the execution control layer component that controls the physical displacement, skeletal dynamics performance and environmental interaction feedback of the virtual character.

[0301] The core processing logic and functions of the motion engine module include pathfinding and spatial navigation, kinematics and dynamics calculation, animation state machine scheduling, and collision detection feedback.

[0302] Pathfinding and spatial navigation include: the motion engine module integrates a navigation grid of the three-dimensional environment. When it receives the physical avoidance response coordinates from the aforementioned AI decision output, the engine automatically calculates the shortest unobstructed path from the current coordinates to the safe coordinates, avoiding static obstacles in the map, such as walls and cliffs.

[0303] The kinematics and dynamics calculations include: the engine calculates the theoretical escape direction vector into the actual displacement and force state for each frame based on the virtual character's physical attributes, including maximum movement speed, acceleration, turning radius and gravity parameters, to ensure that the movement process conforms to the laws of physics, rather than a rigid instantaneous teleportation.

[0304] Animation state machine scheduling includes: based on the calculated motion speed and direction, the motion engine module automatically schedules the skeletal animation system of the virtual character. For example, when an emergency pulse impact is detected, the short-term Granger causality index is extremely high. The engine not only controls the coordinate movement, but also forcibly interrupts the current standby animation, smoothly blends and triggers tactically expressive action postures such as rolling to dodge or raising a shield to retreat.

[0305] Collision detection feedback includes: real-time spatial ray detection and collision bounding box calculation during the movement to safe coordinates, ensuring that the character can correctly respond to ground undulations or other dynamically flying threats during evasive maneuvers.

[0306] The following are explanations of terms that are not technical terms used in this specification:

[0307] A virtual character is a digital entity with perception, decision-making, and action capabilities, driven by artificial intelligence in a computer-generated virtual environment.

[0308] A salient stimulus is a set of perceptual variables that disrupt the historical stationary state of a virtual character, representing a data stream in the environment that has strong perturbation or attraction characteristics.

[0309] Cognitive context slices are logical groups of multiple perceptual variables that are highly correlated in time, representing a combination of specific events in terms of physical or semantic aspects.

[0310] The error correction model is an econometric model used to analyze long-term cointegration relationships and short-term dynamic adjustment mechanisms between non-stationary time series. In this application, it is used to measure deep causal changes between features.

[0311] The unit root test is a statistical test method used to determine whether time series data has a unit root, thereby determining whether it is stationary.

[0312] Henry's gas solubility optimization algorithm is a global optimization heuristic algorithm that simulates the dissolution behavior of gases in liquids according to Henry's law.

[0313] Degree bias is a mathematical phenomenon in graph data processing where nodes with more connected edges tend to gain an absolute numerical advantage in summation calculations, thus masking the true causal relationship.

[0314] This invention performs unit root tests on the multimodal perception state matrix to accurately filter out non-stationary salient stimuli, breaking the limitations of traditional static rule-based information filtering and avoiding the waste of computational resources on useless environmental noise for virtual characters. By constructing a perceptual nearest neighbor network and diffusion clustering based on perceptual information density, isolated perceptual variables are combined into topologically related cognitive context slices, enabling virtual characters to reason in conjunction with context. This invention utilizes an error correction model to decompose the calculation of underlying features into the intensity of strategic motivation transfer and the intensity of tactical reflexive stimuli, creatively distinguishing the driving differences between the disruption of long-term environmental equilibrium and the short-term instinctive shock. The dual-path mechanism of macro-stress scoring and micro-stress scoring designed in this invention eliminates the network degree bias phenomenon caused by a single measurement system, ensuring that neither the threat of widespread global pressure is overlooked nor the threat of direct lethality at a single point is ignored. Through iterative reweighting of macro-stress scoring using the attention weight matrix, closed-loop feedback is achieved, forcing the system's attention to be locked on the single true root variable. The final output reasoning and explanation chain makes the decision-making process of the virtual character completely transparent and highly similar to the judgment logic of humans when facing a crisis, which greatly improves the rationality, explainability and adaptive survival ability of the virtual character's actions in complex and sudden environments.

[0315] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0316] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for real-time perception and action decision-making of virtual characters based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Obtain the multimodal perception state matrix of the virtual character, divide the multimodal perception state matrix into historical stationary state windows and current abrupt state windows, perform unit root tests on the data in the current abrupt state window, extract the perceptual variables for which the first hypothesis is true, and form a set of significant stimulus sources. Step S1 includes: The behavior time step and columns of the multimodal perception state matrix are feature types, and the feature types include the virtual character's visual range perception, auditory volume perception, self-health value state perception, and cooldown time state perception. The multimodal sensing state matrix is ​​divided along the time axis to obtain a historical steady state window representing the memory of calm baseline and a current abrupt state window representing the memory of sudden events. Within the current mutation state window, a unit root test is performed on each sensing feature time series in the multimodal sensing state matrix, and the test statistic is obtained by calculating the difference equation of the sensing feature time series. The extended Dickie-Fowler (ADF) test was used to test the first and second hypotheses. The first hypothesis was that the time series of target-sensing features has a unit root, and the second hypothesis was that the time series of target-sensing features does not have a unit root. The determination is made based on the comparison between the significance probability of the test statistic and the initial judgment criterion: When the saliency probability is greater than or equal to the initial judgment criterion, the first hypothesis is determined to be true, the perceptual feature time series is marked as an abnormal fluctuation state, and all perceptual feature time series in the abnormal fluctuation state are merged to form a saliency stimulus source set. When the significance probability is less than the initial judgment criterion, the second hypothesis is determined to be true, and the perceived feature time series is marked as maintaining a stable baseline state. Step S2: Calculate the temporal mutual information distance between perceptual variables in the set of salient stimulus sources, construct a perceptual nearest neighbor network based on the temporal mutual information distance and calculate the perceptual information density, perform diffusion clustering using the perceptual information density, and generate a set of cognitive environment slices; Step S3: Construct an error correction model within each cognitive context slice queue, calculate the absolute value of the change in the long-term error correction coefficient between perceptual variables, and calculate the absolute value of the change in the short-term Granger causality index. Step S4: Calculate the macro stress score based on the two absolute values ​​in step S3, and obtain the micro stress score by averaging the number of connection edges in the perceptual nearest neighbor network. Combine the macro stress score and the micro stress score to obtain the core decision-making driving force score. Step S5: Calculate the attention weight matrix based on the core decision-driving force score, use the attention weight matrix to iteratively reweight the macro stress score, output the converged core decision-driving force score ranking, and generate the final decision action and reasoning explanation chain based on the perceptual variable corresponding to the top-ranked core decision-driving force score.

2. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, Step S2 includes: In a set of salient stimulus sources, the temporal mutual information distance, which measures the degree of overlap of sequence information, is obtained by calculating the probability distribution of any two perceptual variable sequences and obtaining their mutual information. Based on the temporal mutual information distance, the K nearest neighbor nodes of each perceptual variable are extracted to form a nearest neighbor node set. When the first perceptual variable is located in the nearest neighbor node set of the second perceptual variable, and the second perceptual variable is also located in the nearest neighbor node set of the first perceptual variable, a nearest neighbor bidirectional connection edge is constructed between the first perceptual variable and the second perceptual variable. All nearest neighbor bidirectional connection edges are merged to obtain the perceptual nearest neighbor network. The number of nearest bidirectional connections included in each sensing variable in the statistical sensing nearest neighbor network is used as the sensing information density of that sensing variable.

3. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, Calculate the mean of the perceived information density of all perceived variables in the set of salient stimulus sources. When the perceived information density of a perceived variable is greater than the mean of the perceived information density, it is determined that the perceived variable is in a high-density complex environment and a first diffusion boundary value is assigned to it. When the perceived information density of a perceived variable is less than or equal to the mean perceived information density, the perceived variable is determined to be in a low-density isolated environment, and a second diffusion boundary value is assigned to it. The first diffusion boundary value is smaller than the second diffusion boundary value; The system searches for shared neighbor nodes outward from the center of the perceptual variable. When the number of shared neighbor nodes is greater than the corresponding diffusion boundary value, the neighbor nodes are grouped into the same cognitive environment slice queue until the search process ends. The output is a set of cognitive environment slices composed of interconnected perceptual variables.

4. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, Step S3 includes: Extract all perceptual variables from a single cognitive situation slice queue in the cognitive environment slice set, and establish a vector autoregressive equation based on all perceptual variables; Extract the cointegration relationship equation from the vector autoregression equation, calculate the change in the long-term error correction coefficient between the current abrupt state window and the historical stationary state window, and take the absolute value to obtain the first absolute value intensity. Significance verification calculations are performed on the lagged terms in the vector autoregression equation. The short-term Granger causality index change caused by the introduction of the current mutation state window data is calculated and its absolute value is taken to obtain the second absolute value strength.

5. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, Step S4 includes: The formula for calculating the macro-stress score is as follows: ； in, For the perceived variable i, the macro-stress score. Let i be the set of neighbor nodes connected to the sensing variable i in the nearest neighbor network. Let be the intensity of the first absolute value of the perception variable i for neighbor node j. Let be the intensity of the second absolute value of the perception variable i of neighbor node j. The initial macro- and micro-distribution coefficients are taken as constants between 0 and 1. The formula for calculating the micro-stress score is: ； in, For the micro-stress score of perceived variable i, It is a scalar representing the total number of nodes in the neighbor node set.

6. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, The formula for calculating the core decision-driving force score is: ； in, The core decision-driving force score for perceived variable i. The logistic smoothing activation function maps real numbers to the interval from constant 0 to constant 1. To perform the macro stress score after minimax normalization, To perform the micro-stress score after minimax normalization, and These are the first mapping weight and the second mapping weight, respectively, and B is the mapping bias constant.

7. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, Step S5 includes: Calculate the attention weight matrix based on the core decision-driving force scores of neighboring nodes of the perceived variable i. The attention weight is substituted into the macro stress score for iterative reweighting calculation, the macro stress score is updated, and the core decision-making driving force score is updated based on the updated macro stress score.

8. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, The core decision-driving force score is continuously updated by iterative reweighting calculation. After each iteration, the absolute value of the difference between the current iteration's core decision-driving force score and the previous iteration's core decision-driving force score is calculated, and the sum of all absolute values ​​of differences during the iteration process is calculated. When the sum of the absolute values ​​of the differences is less than the initial convergence minimum constant, stop the iterative calculation, extract the core decision driving force scores of all perceptual variables and sort them in descending order, and output the ranking of the converged core decision driving force scores. Extract the perceptual variable corresponding to the top-ranked core decision-making driving force score, and set the perceptual variable as the final inference root cause; The motion engine module of the virtual character is invoked, and the corresponding physical avoidance response coordinates are output based on the final inference root cause. The virtual character is then controlled to move towards the safe coordinate area to generate the final decision action.

9. The method for real-time perception and action decision-making of virtual characters based on artificial intelligence as described in claim 1, characterized in that, The inference explanation chain generated based on the perceptual variables corresponding to the top-ranked core decision-making driver score includes: The steps of source tracing reasoning are constructed using a natural language processing model, and these steps specifically include: Step S51: Concatenate the unit root test state reversal information to generate the perception trigger paragraph, and output the first interpretation paragraph; Step S52: Concatenate the perceived information density value and the target diffusion boundary value to generate a context-related paragraph, and output the second explanatory paragraph; Step S53: Concatenate the changes in long-term error correction coefficients to generate the expected deviation paragraph, and output the third explanatory paragraph; Step S54: Concatenate the short-term Granger causality index change to generate the primacy reflex trigger paragraph, and output the fourth explanatory paragraph; Step S55: After converging, the attention weights approach a constant of 1 to generate a closed-loop locking paragraph, and output the fifth explanatory paragraph; Step S56: Combine all explanatory paragraphs to output the reasoning explanation in chronological causal order.

Images