Mouse parameter dynamic optimization method based on ai user behavior learning

Abstract

The application relates to the technical field of parameter optimization, in particular to a mouse parameter dynamic optimization method based on AI user behavior learning, which comprises the following steps: collecting the visual state of a screen and the key state of an input device to construct a state index, and determining the action value distribution under the current state in a reinforcement learning model; calculating the difference between the optimal action value and the average action value in the action value distribution, and mapping the difference into a confidence degree representing the proficiency of the reinforcement learning model; generating an exploration coefficient to determine a target action containing a target mouse parameter; monitoring the moving track of the mouse under the target action, calculating operation entropy and taking the operation entropy as a penalty term into a reward function; and updating the action value in the reinforcement learning model. The technical scheme can simultaneously consider the flexibility and stability of mouse operation, and improve the operation precision and user experience.

Description

Technical Field

[0001] This application relates to the field of parameter optimization technology, and in particular to a method for dynamic optimization of mouse parameters based on AI user behavior learning. Background Technology

[0002] With the booming development of competitive games, the real-time performance and accuracy of human-computer interaction have become key factors determining competitive performance. In gaming scenarios, users place extremely high demands on the control performance of input devices such as mice. Taking first-person shooters (FPS) as an example, not only is high sensitivity required in the macroscopic scene of searching for enemies to enable rapid turning, but also extremely high stability is needed in the microscopic scene of aiming with a scope to match fine-tuning hand movements and ensure operational stability.

[0003] Currently, patent application CN118796050A discloses a method, system, device, storage medium, and product for optimizing mouse operation parameters. This method monitors network latency data in real time and calculates the average value. Using preset DPI and IPS prediction models, it establishes an inverse relationship between parameters and network latency. When increased latency is detected due to decreased network bandwidth or increased packet loss rate, the system automatically reduces the mouse's DPI and IPS parameter values. By limiting the mouse's sensitivity and movement speed, it masks the lag caused by network latency, thereby preventing accidental touches due to slow signal transmission.

[0004] However, the mouse parameter adjustments in the above methods rely solely on network communication quality rather than the user's actual operational intent. As a result, when the network is stable, the mouse parameters remain static and cannot adapt to changes in the scenario, thus failing to meet user needs and leading to inaccurate mouse parameter optimization results. Summary of the Invention

[0005] To address the technical problem of inaccurate mouse parameter optimization results, this application provides a dynamic mouse parameter optimization method based on AI user behavior learning, which can simultaneously take into account the flexibility and stability of mouse operation, thereby improving the user's operational accuracy and user experience.

[0006] In a first aspect, this application provides a method for dynamically optimizing mouse parameters based on AI user behavior learning. The optimization method includes: real-time acquisition of the visual state of the screen and the button state of the input device; constructing a state index based on the visual state and button state; using the state index to determine the action value distribution in the current state within a reinforcement learning model; calculating the difference between the optimal action value and the average action value in the action value distribution; mapping the difference to a confidence level representing the proficiency of the reinforcement learning model; generating an exploration coefficient negatively correlated with the confidence level; dynamically determining a target action containing target mouse parameters based on the exploration coefficient; sending the target mouse parameters to the mouse for execution; monitoring the mouse's movement trajectory under the target action; calculating the operational entropy representing the degree of disorder in the trajectory direction distribution; and substituting the operational entropy as a penalty term into the reward function; and updating the action values ​​in the reinforcement learning model for the next optimization.

[0007] By constructing a state index based on visual and button states to locate the distribution of action value, and using the difference between the maximum and average action value to calculate confidence and dynamically adjust the exploration coefficient, and introducing operation entropy as a penalty term into the reward function to update action value, the reinforcement learning model can actively explore new parameters in unfamiliar scenarios and stably lock in the optimal parameters in familiar scenarios. Furthermore, the operation entropy penalty mechanism forces the model to avoid parameter configurations that, although they can generate interaction, will cause the cursor to shake violently, thereby achieving accurate and dynamic matching of mouse parameters with user operation intentions and hand stability.

[0008] Preferably, the steps for acquiring the visual state and the button state include: acquiring an image of the central area of ​​the screen, identifying the current interactive field of view level through a hash algorithm, quantizing the interactive field of view level into an integer as the visual state, wherein the interactive field of view level includes macroscopic field of view, standard field of view, and microscopic field of view; listening to input events of the keyboard and mouse side buttons, identifying the user's operation intention, and quantizing the operation intention into an integer as the button state, wherein the operation intention includes acceleration intention, walking intention, and crouching intention.

[0009] Preferably, the step of constructing a state index based on visual state and button state includes: obtaining the total number of categories of the button state; calculating the product of the integer value corresponding to the visual state and the total number of categories, adding the product to the integer value corresponding to the button state, and using the calculated result as the state index.

[0010] This ensures that each specific "screen-action" combination corresponds to a unique storage row in the Q-value table, enabling fast table lookup and independent storage of action value distribution, and significantly improving the retrieval efficiency and state discrimination of the reinforcement learning model during real-time operation.

[0011] Preferably, mapping the difference to a confidence level representing the proficiency of the reinforcement learning model includes: obtaining the maximum action value and the average action value in the action value distribution; calculating the difference between the maximum action value and the average action value; and using the ratio of the difference to the maximum action value as the confidence level.

[0012] It quantifies the degree of certainty of the reinforcement learning model in the current scenario; it distinguishes whether the model is in the exploration phase or the stable phase, providing objective data support for the adaptive adjustment of the exploration coefficients in the future, and avoiding meaningless random exploration by the model after it has learned the optimal policy.

[0013] Preferably, generating an exploration coefficient that is negatively correlated with the confidence level includes: in response to the confidence level being greater than a confidence level threshold, setting the exploration coefficient to zero to prohibit attempts at new parameter configurations; in response to the confidence level not being greater than the confidence level threshold, the exploration coefficient is the product of a baseline exploration rate and an adjustment coefficient, the adjustment coefficient having a value range of 0 to 1 and being negatively correlated with the confidence level.

[0014] When the reinforcement learning model is sufficiently proficient in the current scene, it strictly executes the optimal mouse parameters to ensure competitive stability. However, when the reinforcement learning model is unfamiliar with the scene, it automatically increases the probability of random attempts to escape local optima, thereby improving the convergence speed of parameter optimization and the robustness of the final result.

[0015] Preferably, dynamically determining the target action containing the target mouse parameters based on the exploration coefficient includes: generating a random number between 0 and 1; in response to the random number being greater than the exploration coefficient, determining the action corresponding to the maximum action value in the action value distribution as the target action; and in response to the random number not being greater than the exploration coefficient, randomly selecting an action from all available actions as the target action.

[0016] By transforming the exploration coefficients into decision-making criteria for action selection, dynamic adjustment of the greedy strategy in reinforcement learning is realized. This ensures that the selection of mouse parameters can not only follow the best configuration in historical experience with a high probability, but also retain a certain probability of discovering potential better configurations, thus guaranteeing the ability of the reinforcement learning model to continuously evolve.

[0017] Preferably, the operational entropy for calculating the degree of disorder in the trajectory direction distribution includes: collecting the direction distribution of the mouse movement trajectory vector within a preset evaluation time window, wherein the direction distribution includes the probability of occurrence of each direction; and using the information entropy of the direction distribution as the operational entropy.

[0018] Operational entropy is used to objectively measure the degree of disorder in a user's hand movements. It can identify poor parameter configurations that cause minor, disordered cursor jitter or frequent direction corrections. Even if these configuration parameters achieve the desired result, they can still be identified as unstable controls, thus providing a quantitative evaluation standard for the smoothness of parameter optimization.

[0019] Preferably, substituting the operational entropy as a penalty term into the reward function includes: obtaining the result reward value generated by the mouse operation, wherein the result reward value is positively correlated with the effective interaction event; calculating the product of the operational entropy and a preset entropy penalty weight as a penalty term; and using the difference between the result reward value and the penalty term as the output value of the reward function.

[0020] When guiding reinforcement learning models to find parameters, they not only pursue the success rate of interaction, but also take into account the smoothness and stability of cursor movement, effectively filtering out the uncontrollable jitter caused by high sensitivity and improving the user experience.

[0021] Preferably, updating the action value in the reinforcement learning model includes: calculating the target value of the current state based on the Bellman equation, using the output value of the reward function, the maximum action value of the next time step, and the discount factor; and updating the target value and the original action value of the current state using a weighted average with the learning rate.

[0022] Preferably, the target mouse parameters include DPI value and LOD value.

[0023] The technical solution of this application has the following beneficial technical effects: By constructing a state index through real-time acquisition of screen visual and button states, and combining this with the action value distribution to calculate the confidence level representing the proficiency of the model, the exploration coefficients are further adaptively adjusted based on the confidence level. Furthermore, operational entropy, representing the degree of trajectory disorder, is introduced into the reward function as a penalty term, achieving adaptive optimization of mouse parameters. This allows the system to automatically perceive and output the most suitable DPI and LOD parameters for different game interaction scenarios. Simultaneously, operational entropy effectively suppresses hand tremors caused by improper mouse parameters, while also balancing the flexibility and stability of mouse operation, thus improving the user's operational accuracy and user experience in esports scenarios. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for dynamically optimizing mouse parameters based on AI user behavior learning, according to an embodiment of this application.

[0025] Figure 2 This is a schematic diagram illustrating the evolution trend of confidence in reinforcement learning models according to embodiments of this application. Detailed Implementation

[0026] According to a first aspect of this application, this application provides a method for dynamically optimizing mouse parameters based on AI user behavior learning, applicable to e-sports game scenarios. In game scenarios, users' mouse control needs vary significantly at different times. For example, in a first-person shooter game, when a user is in a macroscopic field of view quickly turning to search for enemies, higher mouse sensitivity is needed to achieve rapid turning; while when the user is in a microscopic field of view aiming with a high-powered scope, lower sensitivity is needed to match fine-tuning hand movements and to filter out minor hand tremors caused by hand tension. This application achieves dynamic optimization of mouse parameters through a reinforcement learning model. Figure 1 This is a flowchart of a method for dynamically optimizing mouse parameters based on AI user behavior learning, according to an embodiment of this application. Figure 1 As shown, the method for dynamic optimization of mouse parameters based on AI user behavior learning includes steps S101 to S104, which are described in detail below.

[0027] S101, Real-time acquisition of the visual state of the screen and the key state of the input device, construction of a state index based on the visual state and key state, and determination of the action value distribution in the current state in the reinforcement learning model using the state index.

[0028] In one embodiment, visual state refers to the level of interaction granularity corresponding to the current screen display content; key state refers to the category of operation intent conveyed by the user through input devices such as keyboard or mouse side buttons; and state index is a unique identifier generated by combining the above two, used to quickly locate the corresponding policy row in the database of the reinforcement learning model, i.e., the Q-value table. The reinforcement learning model is a Q-learning model or a DQN model; this embodiment uses a Q-learning model as an example for illustration.

[0029] The steps for acquiring the visual state and the button state include: acquiring an image of the central area of ​​the screen, identifying the current interactive field of view level through a hash algorithm, and quantizing the interactive field of view level into an integer as the visual state, wherein the interactive field of view level includes macroscopic field of view, standard field of view, and microscopic field of view; listening to input events of the keyboard and mouse side buttons, identifying the user's operation intention, and quantizing the operation intention into an integer as the button state, wherein the operation intention includes acceleration intention, walking intention, and crouching intention.

[0030] Specifically, for the visual state, image data from the central area of ​​the screen is captured at a high frame rate. To reduce computational load, a hash algorithm is used to calculate the image fingerprint, which is then matched against a pre-defined scene fingerprint database. The matching result is quantified into an integer to determine the visual state. For example, when a wide map scene is matched, it is determined to be a macroscopic field of view, and the visual state value is 0; when a regular combat scene with a crosshair is matched, it is determined to be a standard field of view, and the visual state value is 1; when a partial scene with a scope or a complex UI is matched, it is determined to be a microscopic field of view, and the visual state value is 2.

[0031] For key press states, a global hook is used to listen for input events and match them with preset shortcut keys. For example, when the user presses the Space or Alt key, it is identified as an acceleration intention, and the key press state value is 2; when no special key is pressed, it is identified as a walking intention, and the key press state value is 0; when the user presses the Ctrl or Shift key, it is identified as a crouching intention, and the key press state value is 1.

[0032] After acquiring two independent visual states and button states, they need to be dimensionality-reduced to determine the state index via a lookup table. The state index is a combination of the visual state and the button state. Specifically, constructing the state index based on the visual state and the button state includes: obtaining the total number of categories for the button states; calculating the product of the integer value corresponding to the visual state and the total number of categories; adding the product to the integer value corresponding to the button state; and using the result as the state index. In this way, the two-dimensional state space is mapped to a one-dimensional continuous integer space.

[0033] By using a shift-weighted approach, the two-dimensional state features are flattened into a one-dimensional array of indices. For example, the total number of button state categories. The value is 3, when the user is in the microscopic field of vision, i.e., the visual state. The value is 2, and the character is currently crouching, indicating a button press state. When the value is 1, the calculated state index is 7. Based on this state index, the row of the Q-value table can be directly located, and the entire row of data corresponding to the state index can be extracted as the action value distribution. This ensures that each specific game scene can correspond to an independent set of action value distributions, achieving fine-grained differentiation of game scenes.

[0034] In this way, before the user actually moves the mouse, the system can predict their operation needs based on the visual environment and button actions, and convert them into a unique state index, thus solving the problem of lag in mouse parameter optimization.

[0035] S102, calculate the difference between the optimal action value and the average action value in the action value distribution, map the difference to a confidence level representing the proficiency of the reinforcement learning model; generate an exploration coefficient that is negatively correlated with the confidence level, and dynamically determine the target action containing the target mouse parameters based on the exploration coefficient.

[0036] In one embodiment, the confidence score is used to evaluate the proficiency of the reinforcement learning model in the game scene corresponding to the state index. Mapping the difference to a confidence score characterizing the proficiency of the reinforcement learning model includes: obtaining the maximum action value and the average action value in the action value distribution; calculating the difference between the maximum action value and the average action value; and using the ratio of the difference to the maximum action value as the confidence score.

[0037] The confidence level satisfies the following relationship: ; In the formula, This represents the confidence level, with a value ranging from 0 to 1. This represents the maximum action value in the current action value distribution; The average action value representing the current action value distribution; The default positive number is 0.001, for example, to prevent the denominator from being zero. Since the maximum action value and the average action value may be negative, the confidence level is truncated to ensure that the confidence level is within the range of 0 to 1, and is forced to be a value between 0 and 1.

[0038] When the maximum motion value is significantly higher than the average motion value A value approaching 1 indicates that the reinforcement learning model clearly knows which action is best in the game scenario corresponding to that state index, and is in a stable period; conversely, if the value differences between actions are not significant, A lower value indicates that the reinforcement learning model is still in the exploratory stage and cannot know which action is best in the game scenario corresponding to that state index.

[0039] Furthermore, to avoid interference from transient noise, an exponential moving average algorithm is used to smooth the confidence scores of the reinforcement learning model for any number of training steps. Please see [link / reference]. Figure 2This diagram illustrates the evolution trend of the confidence level of the reinforcement learning model according to the embodiments of this application. The horizontal axis represents the number of training steps, and the vertical axis represents the confidence level. In the early stage of training, the confidence level of the reinforcement learning model remains at a low level, mainly conducting random exploration. Subsequently, as data accumulates, the reinforcement learning model learns the optimal strategy for the current scenario, and the confidence level fluctuates upward and stabilizes at a high level. At this time, the optimal parameters are locked to ensure stable operation. When the game scenario changes, the confidence level declines, indicating that the reinforcement learning model is not yet familiar with the new game scenario. At this time, the exploration coefficient is automatically increased to learn the optimal strategy under the new game scenario. In this way, the dynamic adaptive adjustment of the reinforcement learning model is realized.

[0040] Based on the calculated confidence level, the learning strategy needs to be dynamically adjusted. The exploration coefficient represents the probability that the reinforcement learning model will attempt new parameter configurations. The exploration coefficient of the reinforcement learning model is adjusted according to the confidence level. Generating an exploration coefficient negatively correlated with the confidence level includes: in response to the confidence level being greater than a confidence threshold, setting the exploration coefficient to zero to prohibit attempts at new parameter configurations; and in response to the confidence level not being greater than the confidence threshold, the exploration coefficient is the product of the baseline exploration rate and the adjustment coefficient, where the adjustment coefficient ranges from 0 to 1 and is negatively correlated with the confidence level.

[0041] Specifically, let the baseline exploration rate be... For example, a value of 0.2 is used; the confidence threshold is... For example, the value is 0.8. When Greater than At that time, the exploration coefficient It is 0. When Not greater than At that time, the exploration coefficient The calculation satisfies the following relation: ; In the formula, The constant represents the decay rate, exemplarily set to 2; e is the base of the natural logarithm. As the model's proficiency, or confidence, increases, the exploration coefficient decreases smoothly, allowing the reinforcement learning model to remain stable at high confidence levels. However, when the confidence is low, i.e., when the reinforcement learning model is less familiar, it is more necessary to try new parameter configurations.

[0042] After determining the exploration coefficient, specific action selection can be performed. Specifically, dynamically determining the target action containing the target mouse parameters based on the exploration coefficient includes: generating a random number between 0 and 1; in response to the random number being greater than the exploration coefficient, determining the action corresponding to the maximum action value in the action value distribution as the target action; in response to the random number not being greater than the exploration coefficient, randomly selecting an action from all available actions as the target action. It should be noted that the target mouse parameters include DPI value and LOD value.

[0043] For example, if the current state is high confidence and the exploration coefficient is 0, the optimal action will be locked, and the action corresponding to the maximum action value will be directly determined as the target action. For example, if action A3 is selected, its corresponding DPI value, i.e., sensitivity value, is 800 (dots per inch), and its LOD value, i.e., the distance to stillness, is 1.0 mm.

[0044] In this way, an adaptive exploration mechanism based on confidence is constructed, which can boldly try new parameter configurations when the reinforcement learning model is unfamiliar, and lock in the optimal strategy after the model matures.

[0045] S103, the target mouse parameters are sent to the mouse for execution, the movement trajectory of the mouse under the target action is monitored, the operation entropy representing the degree of disorder in the trajectory direction distribution is calculated, and the operation entropy is substituted into the reward function as a penalty term.

[0046] In one embodiment, operational entropy is a concept borrowed from Shannon entropy in information theory, used to measure the degree of disorder and chaos in a user's hand movements; the reward function is the core evaluation criterion used to guide the evolution of reinforcement learning models.

[0047] The DPI and LOD values ​​from the target mouse parameters are sent to the mouse for execution. Subsequently, when the user performs an operation, the movement trajectory of the mouse cursor under the target action is collected. The calculation of the operational entropy, which characterizes the degree of disorder in the trajectory direction distribution, includes: collecting the direction distribution of the mouse movement trajectory vector within a preset evaluation time window, wherein the direction distribution includes the probability of occurrence of each direction; and using the information entropy of the direction distribution as the operational entropy.

[0048] Specifically, the evaluation time window is set to 200 milliseconds. During this period, the mouse cursor position is collected at a polling rate of 1000 Hz, and the 360-degree planar space is divided into... A discrete directional interval, for example The interval is 36, meaning each interval is 10 degrees. The frequency with which the trajectory vector falls into each interval is used to obtain the set of directional probability distributions. Wherein, the trajectory vector is the tangent direction of the mouse cursor position along the movement trajectory. Operational entropy. Satisfying the relation: ; In the formula, This indicates that the mouse movement trajectory falls on the . The probability of each directional interval. Let represent the total number of directional intervals. Understandably, when a user operates decisively and the trajectory is smooth, the trajectory vector is concentrated in a few directional intervals, the probability distribution is sharp, and the calculated operational entropy is low; when a user is hesitant, their hand will unconsciously tremble slightly, causing the trajectory vector to be randomly distributed in various directions, and the operational entropy will increase significantly.

[0049] To further improve the stability of user operations after dynamic optimization of mouse parameters, a reward mechanism incorporating a penalty term was constructed. The step of substituting the operation entropy as a penalty term into the reward function includes: obtaining the result reward value generated by the mouse operation, wherein the result reward value is positively correlated with the effective interaction event; calculating the product of the operation entropy and a preset entropy penalty weight as the penalty term; and using the difference between the result reward value and the penalty term as the output value of the reward function.

[0050] Specifically, the reward function Satisfying the relation: ; In the formula, This represents the output value of the reward function; The reward value represents the result, which is awarded when a valid interaction event such as a kill or skill hit occurs. Take a positive value, such as 100; otherwise, take 0. and These are the result weight and the entropy penalty weight, respectively. For example, =1, It is 5; Let this be the operational entropy. Understandably, even if the user hits the target and generates a valid interaction event, if their action is accompanied by violent, disordered jitter, the final reward will be affected. It will also be significantly reduced.

[0051] In this way, the operational entropy, as a negative feedback signal, can effectively filter out the noise data generated by disordered jitter, forcing the reinforcement learning model to find the parameter configuration that can both complete the task and allow the user's hand control to be most stable.

[0052] S104, Update the action values ​​in the reinforcement learning model for the next optimization.

[0053] In one embodiment, updating the action value in the reinforcement learning model includes: calculating the target value of the current state based on the Bellman equation, using the output value of the reward function, the maximum action value of the next time step, and a discount factor; and updating the target value and the original action value of the current state using a learning rate.

[0054] The update process of Q-value satisfies the following relationship: ; In the formula, Represents the index of the current state after the update in the Q-value table. and target action The corresponding action value; This represents an assignment operation; The learning rate, exemplified by a value of 0.1, determines the degree to which new information overwrites old information. The output value of the reward function; This represents a discount factor, exemplified by a value of 0.9, used to measure the importance of future long-term returns. Represents the state at the next moment. The maximum action value of all actions; Index of the current state before updating the Q-value table and target action The corresponding action value.

[0055] The state index for the next time step is determined, and the updated reinforcement learning model is used to perform the next optimization to obtain the target mouse parameters corresponding to the state index at the next time step. Understandably, to reduce computational cost, the updated reinforcement learning model can also be used to perform the next optimization only when the state index changes.

[0056] In this way, dynamic optimization of mouse parameters in different game scenarios was achieved.

[0057] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A method for dynamic optimization of mouse parameters based on AI user behavior learning, characterized in that, The optimization method includes: real-time acquisition of the visual state of the screen and the key state of the input device, constructing a state index based on the visual state and key state, and using the state index to determine the action value distribution in the current state in the reinforcement learning model; Calculate the difference between the optimal action value and the average action value in the action value distribution, and map the difference to a confidence level that represents the proficiency of the reinforcement learning model; Generate an exploration coefficient that is negatively correlated with the confidence level, and dynamically determine the target action containing the target mouse parameters based on the exploration coefficient; The target mouse parameters are sent to the mouse for execution, the mouse movement trajectory under the target action is monitored, the operational entropy representing the degree of disorder in the trajectory direction distribution is calculated, and the operational entropy is substituted into the reward function as a penalty term; the action value in the reinforcement learning model is updated for the next optimization.

2. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, The steps for obtaining the visual state and the button state include: The image of the central area of ​​the screen is captured, the current interactive field of view level is identified by a hash algorithm, and the interactive field of view level is quantized into an integer as the visual state. The interactive field of view level includes macro field of view, standard field of view and micro field of view. Listen for input events from the keyboard and mouse side buttons, identify the user's operation intention, and quantify the operation intention into an integer as the key state. The operation intention includes acceleration intention, walking intention, and crouching intention.

3. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 2, characterized in that, The state index constructed based on visual state and key state includes: Obtain the total number of categories of the button states; Calculate the product of the integer value corresponding to the visual state and the total number of categories, add the product to the integer value corresponding to the button state, and use the result as the state index.

4. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, Mapping the differences to confidence levels representing the proficiency of reinforcement learning models includes: Obtain the maximum action value and the average action value in the action value distribution; Calculate the difference between the maximum action value and the average action value; The ratio of the difference to the maximum action value is used as the confidence level.

5. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, Generating exploration coefficients that are negatively correlated with the confidence level includes: In response to the confidence level being greater than a confidence threshold, the exploration coefficient is set to zero to prevent attempts at new parameter configurations. In response to the confidence level not being greater than the confidence level threshold, the exploration coefficient is the product of the baseline exploration rate and the adjustment coefficient, the adjustment coefficient being in the range of 0 to 1 and negatively correlated with the confidence level.

6. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, Dynamically determining the target action containing the target mouse parameters based on the exploration coefficient includes: generating a random number between 0 and 1; In response to the random number being greater than the exploration coefficient, the action corresponding to the maximum action value in the action value distribution is determined as the target action; In response to the random number not being greater than the exploration coefficient, an action is randomly selected from all available actions and determined as the target action.

7. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, The operational entropy used to calculate the degree of disorder in the trajectory direction distribution includes: Within a preset evaluation time window, the directional distribution of the mouse movement trajectory vector is collected, and the directional distribution includes the probability of occurrence of each direction; The information entropy of the directional distribution is used as the operational entropy.

8. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, The step of substituting operational entropy as a penalty term into the reward function includes: Obtain the reward value of the result generated by the mouse operation, which is positively correlated with the effective interaction event; calculate the product of the operation entropy and the preset entropy penalty weight as the penalty term; The difference between the reward value and the penalty term is used as the output value of the reward function.

9. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, Updating the action value in the reinforcement learning model includes: Based on the Bellman equation, the target value of the current state is calculated using the output value of the reward function, the maximum action value of the next time step, and the discount factor. The target value and the original action value of the current state are updated using a weighted average based on the learning rate.

10. The method for dynamic optimization of mouse parameters based on AI user behavior learning according to claim 1, characterized in that, The target mouse parameters include DPI value and LOD value.

Images