A personalized interactive interface self-adaption method based on reinforcement learning

Abstract

The application relates to the technical field of computer systems based on specific calculation models, and discloses a personalized interactive interface self-adapting method based on reinforcement learning, which comprises the following steps: setting complexity cost for interface elements and maintaining a complexity total budget, purifying user operation signals by fusing physical motion data of the device, and then calculating objective behavior efficiency indexes; the method also utilizes the decision confidence of the reinforcement learning model itself, cooperates with the behavior efficiency indexes, and adjusts the visibility of the interface elements in a closed loop under the constraint of the total budget; the application converts the interface layout problem into a budget management process oriented to objective operation efficiency, establishes a continuous self-adapting adjustment relationship between interface complexity and user skill level, enables the interface to automatically and seamlessly evolve along with the user skill evolution, intelligently perceives and supports the learning and exploration process of the user, and improves the reliability in real complex application scenarios.

Description

TECHNICAL FIELD

[0001] The application relates to a personalized interactive interface adaptive method based on reinforcement learning, and belongs to the technical field of computer systems based on specific computing models. BACKGROUND

[0002] At present, in order to improve the experience of human-computer interaction, providing a personalized interactive interface for a user has become a widely accepted technical consensus. The current mainstream technical method is usually to allow the user to manually set the interface according to personal preferences, or to allow the system to analyze the historical operation data of the user, for example, to place frequently used functions in a more accessible position. These methods meet the personalized needs of the user in a certain stage to some extent and constitute the basis of the existing adaptive interaction technology. However, when the perspective is placed on a longer interaction time scale, especially when observing the complete life cycle of a user from a novice to an expert, the inherent design philosophy shortcoming of the above-mentioned static or historical fixed repetition-based methods begins to appear. A typical scenario is that a new employee needs an extremely simple interface with guidance to complete the most basic tasks when using a complex enterprise management software. However, as the user's skills improve rapidly, the interface designed for novices becomes a serious obstacle to improving work efficiency because of the nested hidden advanced functions. The system not only fails to help the user grow, but also fixes the initial novice state, forming an adaptation gap.

[0003] To solve this problem, an obvious improvement idea is to provide a manual switching option between a novice mode and an expert mode for the user. However, this method ignores the fact that the same user is a novice when facing new functions introduced by the software, but is an expert when dealing with routine tasks that he is familiar with. This mixed state makes any single fixed mode unable to truly match the dynamic changing needs of the user. Another seemingly feasible path is to further enhance the analysis of user behavior. However, if the analysis is only based on the analysis of surface data such as click frequency, it will fall into a deeper logical dilemma because high-frequency clicking of a function may be due to the user's proficiency and preference, or may be due to the unreasonable design of the function, which forces the user to repeatedly operate to complete the task. The existing technology lacks an internal mechanism to effectively distinguish between these two types of high-frequency behaviors with completely different natures.

[0004] Specifically, the prior art mainly has the following fundamental constraints in realizing truly dynamic and efficient interface adaptation: 1. The skill level of the user is assumed to be a stable and single state, which cannot adapt to the reality that the user's skill dynamically switches between learning new functions and handling routine tasks; 2. The core of its personalization logic is the solidification and reproduction of historical behavior, rather than the real assessment and support of user operation efficiency, which thus limits the growth space of user productivity. Therefore, how to build a computational model that does not rely on the direct judgment of whether the user is a novice or an expert, but can establish a quantifiable resource management mechanism for the complexity of the interface itself, and dynamically allocate this resource based on the objective assessment of the user's real operation efficiency, so that the interactive interface can seamlessly and automatically evolve with the evolution of the user's skill, becomes the technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a personalized interactive interface adaptation method based on reinforcement learning, which mainly aims to solve the problem that the prior art cannot effectively assess the user's real operation efficiency and cannot adapt to the dynamic evolution of the user's skill level, thereby making it difficult to continuously provide an interactive interface that can balance exploration and efficiency to match the user's growth throughout the complete life cycle.

[0006] To achieve the above-mentioned purpose, the present application provides a personalized interactive interface adaptation method based on reinforcement learning, which comprises:

[0007] Setting respective basic complexity costs for interface elements in the interactive interface;

[0008] Parallelly acquiring user operation trajectory data originating from the human-machine interface event stream and device physical motion data originating from the motion sensor; if the motion state indicated by the device physical motion data exceeds the motion threshold value stored in the user device, filtering the user operation trajectory data based on the frequency characteristics of the device physical motion data to generate purified operation trajectory data, otherwise, directly using the user operation trajectory data as the purified operation trajectory data;

[0009] Calculating a behavior efficiency indicator based on the purified operation trajectory data;

[0010] Using a reinforcement learning model, calculating an information entropy value based on the expected return value distribution of the reinforcement learning model itself within a decision cycle; if the information entropy value is not lower than the confidence threshold value stored in the user device, updating a complexity total budget of an interface based on the behavior efficiency indicator, and under the constraint of not exceeding the complexity total budget, selecting a visible set of interface elements by the reinforcement learning model, and adjusting the interface presentation according to the visible set.

[0011] Preferably, the step of calculating the behavior efficiency indicator based on the purified operation trajectory data comprises: obtaining two adjacent interactive operations in the purified operation trajectory data; measuring a time interval between the two adjacent interactive operations; determining whether an invalid exploration behavior occurs in the purified operation trajectory data, wherein the invalid exploration behavior refers to a behavior that a user opens a menu and does not select any option in the menu before closing the menu; and generating the behavior efficiency indicator based on the time interval and the occurrence of the invalid exploration behavior.

[0012] Preferably, the method further comprises: obtaining, by the ambient light sensor, an ambient light intensity in which the device is located; calculating a context influence factor based on the device physical motion data and the ambient light intensity; and combining the base complexity cost of the interface element and the context influence factor to generate an effective complexity cost. When the reinforcement learning model makes a selection, the constraint followed by the reinforcement learning model is that the sum of the effective complexity costs of the interface elements in the visible set does not exceed the complexity total budget.

[0013] Preferably, the step of updating the complexity total budget based on the behavior efficiency indicator comprises: increasing the value of the complexity total budget when the value of the behavior efficiency indicator is continuously higher than the efficiency positive threshold; and decreasing the value of the complexity total budget when the value of the behavior efficiency indicator is continuously lower than the efficiency negative threshold and the information entropy value is not lower than the confidence threshold.

[0014] Preferably, the step of calculating an information entropy value based on the distribution of the expected return values of the reinforcement learning model itself comprises: after the reinforcement learning model calculates the respective expected return values of all candidate interface element visibility adjustment actions; forming a probability distribution by collecting the values of all the expected return values; and calculating the information entropy value based on the probability distribution.

[0015] Preferably, when the information entropy value is lower than the confidence threshold, the method triggers a learning incentive state, and in the learning incentive state, the operation of decreasing the complexity total budget due to the value of the behavior efficiency indicator being lower than the efficiency negative threshold is suspended.

[0016] Preferably, the step of filtering the user operation trajectory data based on the frequency characteristics of the device physical motion data comprises: performing a fast Fourier transform on the device physical motion data to determine a main frequency of shaking; constructing a digital band-stop filter, wherein the center frequency of the digital band-stop filter is set according to the main frequency of shaking; and processing the user operation trajectory data by using the digital band-stop filter to generate the purified operation trajectory data.

[0017] Preferably, the calculation of the effective complexity cost follows the following rules: wherein, is the base complexity cost, to quantify the generated motion impact factor according to the amount of device physical motion data, to quantify the generated lighting impact factor according to the amount of ambient lighting intensity, with to store in the user device the sensitivity coefficients respectively characterizing the interface elements' sensitivity to motion and lighting.

[0018] Preferably, the method further comprises: triggering the implicit cognitive load detection step when no user operation trajectory data is acquired within a continuous time window; the implicit cognitive load detection step comprises: generating a human eye sub-threshold visual stimulation signal through the interactive interface, and capturing the user's pupil micro-jitter signal via the image sensor; calculating the mutual information value between the pupil micro-jitter signal and the sub-threshold visual stimulation signal; and providing a negative reward to the reinforcement learning model when the mutual information value exceeds a cognitive overload threshold.

[0019] Preferably, all the steps of acquisition, calculation, judgment, selection and adjustment contained in the method are executed locally on the user device, without relying on data communication with a cloud server to complete the decision-making of the reinforcement learning model.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] 1. The present application quantifies the time and sequence characteristics of user interaction behavior into a behavior efficiency indicator as a reward signal by pre-defining the complexity cost of interface elements and introducing a total complexity budget of the interface for dynamic maintenance, so that the reinforcement learning model makes decisions within a constrained resource system; this transformation of interface layout problem into a closed-loop budget management process oriented to behavior efficiency establishes a continuous self-adaptive adjustment relationship between the interactive interface and the user's skill level, and the evolution of interface form is no longer dependent on direct guessing of user intent or solidification of historical behavior, but becomes an objective and automatic mapping of changes in user operation proficiency, so that the need for simplicity when exploring new functions and the need for efficiency when performing routine tasks can coexist dynamically in the same interaction session.

[0022] 2. By further analyzing the internal state of the reinforcement learning model when making decisions in real time to determine its decision confidence, the system provides a basis for distinguishing different causes behind low behavior efficiency; when the model presents a continuously low decision confidence because the user's behavior enters its insufficiently learned field, the system temporarily suspends the adjustment path of reducing the total complexity budget due to negative incentives, which is not simply tolerating inefficiency, but rather builds a protected learning environment for the temporary period of operation inexperience that inevitably accompanies the user's exploration of new functions or adaptation to new interfaces, avoiding the system's blocking of the user's skill growth path by mistakenly punishing positive exploration as negative loss, and evolving human-machine collaboration from pure efficiency matching to intelligent perception and support for the user's growth process.

[0023] 3、The application also introduces the physical motion state information of the device itself to filter the trajectory data of the user in the frequency domain, and when it is determined that the user is in an operation vacuum state caused by information overload, the model is provided with a preset negative reward through the analysis of the synchronization of sub-threshold visual stimulation and physiological response. This mechanism that fuses and decouples the noise signals from the physical environment and the cognitive state signals from the user's subconscious with the explicit operation signals of the user ensures that the calculation basis of the behavior efficiency index can reflect the user's real interaction intention and cognitive load, so that the decision of the whole adaptive system not only adapts to the change of the user's skill, but also adapts to the change of the physical environment and the internal cognitive state, thereby improving the reliability and environmental resilience of the adaptive method in real complex application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1 The figure is a closed-loop control function architecture diagram of the adaptive method of the application.

[0025] Fig. 2 The figure is a signal purification and reinforcement learning decision core flowchart of the application.

[0026] Fig. 3 The figure is a device end localization system deployment architecture diagram of the application. DETAILED DESCRIPTION

[0027] In order to make the technical solutions and advantages of the application clearer, the technical solutions of the application will be described in detail below with reference to the drawings, but those skilled in the art can understand that the embodiments described below and the features in the embodiments can be combined with each other without conflict, and the various technical solutions formed thereby are also within the protection scope of the application.

[0028] This invention discloses a personalized interactive interface adaptive method based on reinforcement learning. Its computational model, at the system architecture level, includes an interface element cost definition module, a multi-source interaction signal acquisition and purification module, a behavior efficiency evaluation and budget update module based on dual-path logic, and a reinforcement learning execution module for constraint-based interface decision-making. These modules operate collaboratively on the user device, forming a closed-loop control system that drives the interactive interface to continuously evolve according to the user's actual capabilities and environment. In application scenarios such as new employees using complex enterprise resource planning systems for the first time, this method aims to resolve the long-standing contradiction of a generic interface being too complex for beginners and inefficient for the same user who later becomes an expert due to deeply hidden functions. It transforms the adjustment of the interface layout into a computational problem of dynamically managing the interface complexity budget, guided by calculable operational efficiency, thereby establishing a continuous adaptive relationship between interface form and user skill level. To transform the interface design object into an entity manageable by the computational model, this method first sets a basic complexity cost for each interface element in the interactive interface. The interface element here refers to any unit in the interface that can be displayed or hidden independently, such as a button or a menu item, whose basic complexity cost... It is predefined as a dimensionless integer representing its inherent cognitive load and interaction difficulty; for example, the cost of a button in a primary menu. It can be denoted as 1, while the cost of an option within a secondary menu that requires two clicks to access is... It can be calibrated as 3; however, in interactive scenarios such as users operating devices on moving vehicles or viewing screens under strong outdoor light, the actual interactive difficulty of an interface element is not solely determined by its design. Therefore, the system also introduces an effective complexity cost. The method employs a dynamic modulation mechanism; specifically, it acquires the ambient light intensity of the device through the ambient light sensor built into the mobile device, and simultaneously acquires the physical motion data of the device through a motion sensor. Based on these two environmental signals, the system jointly calculates a context influence factor and follows a specific calculation rule to generate an effective complexity cost. The rule is defined deterministically as follows: In this formula, This refers to the aforementioned basic complexity cost. It is a motion influence factor generated by quantifying physical motion data of the equipment, such as the root mean square value of angular velocity. This is a light influence factor generated based on the ambient light intensity, such as the lux value. and are respectively stored in the pre-calibration of the user device, representing the sensitivity coefficients of the specific interface element to motion and illumination, for example, a physical smaller button, its sensitivity coefficient to motion will be set higher, while a color contrast lower text label, its sensitivity coefficient to illumination will be set higher; in this way, a basic cost of 1 small button, in the device stable indoor environment ( ), its effective complexity cost is still 1, but in the bumpy environment, for example is quantified as 0.8, 0.5, its effective complexity cost will be dynamically adjusted to , this effective complexity cost after dynamic modulation will be followed by the real-time price of the interface resources when the subsequent reinforcement learning model makes decisions, so that the model's decisions can adapt to changes in the physical environment.

[0029] After establishing the cost system for interface elements, the system enters the continuous interactive signal acquisition and processing phase. The challenge lies in the fact that users' raw operation signals are often contaminated by non-intentional physical noise. For example, invalid touch trajectories generated by a user's movement in a bumpy vehicle, if directly used for analysis, can mislead the system's judgment of the user's true intentions. To address this challenge, the system is configured to acquire two data streams in parallel: one is user operation trajectory data from the human-machine interface event stream, and the other is device physical motion data from the motion sensor. Internally, the system maintains a motion threshold stored in the user device. The calibration process for this threshold is as follows: offline, motion data of the device is collected under typical static conditions (e.g., stationary on a desktop) and dynamic conditions (e.g., handheld walking). The 95th percentile of the motion data amplitude under static conditions is used as the threshold to distinguish between device jitter and the user's intentional movement. During system operation, if the motion state indicated by the real-time monitored physical motion data of the device exceeds a certain motion threshold, the system determines that there is significant environmental noise and triggers a filtering process. Specifically, this process first performs a Fast Fourier Transform on the current physical motion data of the device to determine the jitter frequency with the most concentrated energy. Subsequently, the system dynamically constructs a digital band-stop filter with this jitter frequency as the center frequency, and uses this filter to process the user operation trajectory data acquired in the same time period to filter out noise components introduced by environmental jitter. The trajectory data after this processing is defined as purified operation trajectory data. If the motion state of the device does not exceed the motion threshold, the original user operation trajectory data is directly used as purified operation trajectory data. This mechanism enables the signals input to subsequent analysis modules to more accurately reflect the user's interaction intent.

[0030] Based on the purification operation trajectory data, the system then calculates a behavioral efficiency index that serves as a reward signal for the reinforcement learning model. The calculation logic of this indicator aims to objectively quantify the user's operational proficiency. Specifically, the system continuously measures the time interval between two adjacent interactions in the cleanup operation trajectory data, and simultaneously determines whether any invalid exploration behavior occurred—such as the user opening a menu but not selecting any option before closing it. Based on these two dimensions of observation, the system generates a behavioral efficiency indicator. For example, a single fast, continuous operation of less than 200 milliseconds can provide... A contribution earns a +0.2 incentive, while a pause exceeding 3 seconds incurs a -0.1 penalty. Each invalid exploration action results in a direct penalty. Apply a penalty of -0.5; thus, the more skilled and direct the user's operation, the higher their behavioral efficiency index. The higher the cumulative value, the better, and vice versa. become the basis for adaptive adjustment of interface complexity; the core decision of this method is driven by a reinforcement learning model, whose goal is to maximize the long-term cumulative behavioral efficiency indicator expectation under a constraint called the total budget of interface complexity ; however, inefficient behavior (i.e., low ) itself has semantic ambiguity, which can either result from user disorientation due to interface clutter, or from user exploration for learning new functionality, and indiscriminately penalizing all inefficient behavior can hinder the learning process; to resolve this issue, this method introduces a mechanism that inverts the reinforcement learning model's own decision confidence; specifically, in each decision cycle of the reinforcement learning model, after it calculates the expected return value for each candidate interface element visibility adjustment action (i.e., showing or hiding an element) such as the Q value in Q-Learning, the system does not directly select the optimal action, but first constructs a probability distribution from the numerical set of all these expected return values; then, the system calculates the information entropy value based on this probability distribution; this information entropy value, as a quantitative indicator of model decision confidence, to further describe the quantification process, assume that the model needs to choose from actions A, B, and C, when the user is proficient in operating a certain routine task, the calculated expected return values (Q values) can be {Q(A)=5.0, Q(B)=1.2, Q(C)=1.0}, after the system converts them into a probability distribution using the Softmax function, it gets the execution probabilities of each action as {0.978, 0.012, 0.010}, at this time, the information entropy value calculated based on this distribution is about 0.14, which is a value close to zero, indicating that the model is highly confident in choosing action A, in contrast, when the user explores new functionality, the model calculates the Q values as {Q(A)=2.1, Q(B)=2.0, Q(C)=1.9}, the converted probability distribution is about {0.355, 0.322, 0.323}, at this time, the calculated information entropy value is about 1.58, which is a higher value, indicating that the model is very hesitant about which action to choose, therefore, by comparing the calculated information entropy value with a pre-set confidence threshold (e.g., 1.0), the system can accurately distinguish between high-confidence routine operation and low-confidence exploratory learning.

[0031] In view of this, the system sets a confidence threshold, which is calibrated according to the distribution of information entropy values presented by the model under typical routine tasks and exploration tasks, and according to the comparison result of the currently calculated information entropy value and the threshold, the budget update and interface adjustment of the dual-channel logic are executed; if the currently calculated information entropy value is not lower than the confidence threshold, it indicates that the user is in a routine operation or lost state, at this time the system executes the standard budget update logic; the logic is specifically: when the value of the behavior efficiency index continuously exceeds a preset positive efficiency threshold, the system determines that the user has mastered the interface of this complexity, and accordingly increases the value of the complexity total budget ; while the value of the behavior efficiency index continuously falls below a negative efficiency threshold, the system determines that the current interface may be too complex for the user, and reduces the value of the complexity total budget ; the logic is specifically: assuming that the initial value of the system preset complexity total budget is 50, the positive efficiency threshold is +0.8, the negative efficiency threshold is -0.5, and the confidence threshold for distinguishing the exploration state is 1.0, when the system monitors that the value of the user behavior efficiency index is continuously +1.2, because it is higher than the positive threshold, the system determines that the user has mastered the current interface, and increases the total budget to 55, on the contrary, if the behavior efficiency index falls to -0.7, which is lower than the negative threshold, the system will further analyze the decision confidence: if the currently calculated decision information entropy value is 1.6, which is higher than the confidence threshold 1.0, the system determines that the user is lost due to interface overload, and immediately reduces the total budget to 45 to simplify the interface; but if the information entropy value is only 0.8 under the same low efficiency of -0.7, which is lower than the confidence threshold, it indicates that the user is in a protected learning exploration period, the system will suspend the operation of reducing the budget, so that the total budget remains 50 unchanged, this adjustment mechanism containing specific numerical judgment ensures that the interface adaptation can match the user's proficiency and intelligently support the growth process; on the contrary, when the information entropy value is lower than the confidence threshold, the system determines that the user has entered the exploration period, and triggers a learning incentive state; in this state, the system will temporarily suspend the operation of reducing the complexity total budget due to the value of the behavior efficiency index being lower than the negative efficiency threshold, which provides a protected learning sandbox for the user to temporarily and unpunishedly be inefficient when exploring new functions, thereby supporting the user's skill growth process; after the budget update is completed, the final one is executed by the reinforcement learning model without exceeding the updated complexity total budget Under the constraints, select an optimal set of visible interface elements and adjust the interface presentation in real time based on this set of visible elements.

[0032] Furthermore, to address situations where users experience an operational vacuum due to information overload, this method also integrates an implicit cognitive load detection mechanism. The specific steps are as follows: when no user action is detected within a preset time window (e.g., 3 seconds), the system generates a subthreshold visual stimulus signal sequence based on the screen background. It synchronously captures the user's pupil micro-pulse signal sequence via a standard image sensor at a typical frequency (e.g., 60Hz). Subsequently, within a sliding time window (e.g., 2 seconds), the system constructs the joint probability distribution of the two sequence signals by quantizing and binning them. And calculate the mutual information value according to the information theory formula. If the value exceeds the cognitive overload threshold calibrated by the user baseline, the system determines that the user is in a cognitive deadlock and provides a preset negative reward to the reinforcement learning model to force interface simplification, thereby actively intervening to break the user's cognitive deadlock. This detection method, which includes clear technical parameters and calculation formulas, constitutes an important technical feature of this invention, ensuring an objective assessment of the user's cognitive state. As a redundancy guarantee for this mechanism, the interface always provides a fixed one-click back button, allowing users to reset the interface to the initial simplified mode at any time through active interaction, thereby ensuring the availability of the system in various scenarios. It should be noted that the adaptive method disclosed in this invention includes all steps of acquisition, calculation, judgment, selection, and adjustment, all of which are designed to be executed locally on the user's device. The entire decision-making closed loop can be completed without relying on data communication with the cloud server. This local execution configuration not only eliminates the impact of network latency on the interactive experience but also ensures the privacy of user interaction data.

[0033] Example 1: In a large enterprise resource planning (ERP) system, a newly hired financial analyst logs into the system for the first time without any external guidance to handle their first task: submitting a travel expense reimbursement. At this moment, the system interface presented to the user is not the default full-featured view with hundreds of function options, but an initial form generated by the method of this invention. Due to this being the first use, the system's internal efficiency indicators regarding the user's behavior... Since there is no historical data on the model's decision states, the total complexity budget of the system is... It is initialized to a low value of 20; under this budget constraint, the reinforcement learning model is based on the effective complexity cost of each interface element. Only core interface elements directly related to the expense reimbursement task were selected and displayed, such as the "Create Expense Form" button, the amount input box, the explanation of the reason, and the submit button. The total cost did not exceed the budget. while other high-level functional modules such as vendor management, asset depreciation, quarterly report generation, etc. remain invisible.

[0034] In the first few operations, the analyst completes several reimbursements, whose operation characteristics are characterized by direct point-and-click and short interaction time interval, based on which the system calculates a positive and high value of the behavior efficiency indicator ; in response, the reinforcement learning model continuously receives positive rewards and gradually increases the value of the complexity total budget according to the budget update logic in the above ; accordingly, some functions adjacent to the core task and with moderate cost begin to be selectively presented, such as the historical reimbursement query button and the preset template menu; further, after a month, when the analyst needs to perform a monthly report generation task, the interface has evolved into a medium-level form containing a series of medium-complexity functions such as pivot table chart generation tools, without relying on any preset novice mode or expert mode switching, but by quantifying the user's operation efficiency as a reward signal and using the signal to drive a budget management process targeting interface complexity, so that the form of the interface matches the evolution of the user's skill level.

[0035] At this time, the analyst faces a new task: completing a non-standardized quarterly financial risk assessment report that requires calling multiple data sources; he begins to exhibit searching and trying behaviors on the interface, such as long pauses in the cursor or repeatedly opening and closing multiple irrelevant drop-down menus, which are judged as multiple invalid exploration behaviors under the rules defined in the above , resulting in a significant and continuous decrease in the value of the behavior efficiency indicator ; at the same time, as the user's behavior sequence enters an area that the reinforcement learning model has not fully learned, the expected return values calculated by the model for different subsequent actions become flat and close to each other, and thus, according to the calculation procedure, the system calculates that the model decision information entropy value is continuously higher than the preset confidence threshold; in this state, a synergistic effect is formed between the behavior efficiency indicator calculated based on user operations and the decision confidence calculated based on the internal state of the model; instead of immediately adjusting only when the behavior efficiency indicator is negative, the system forms a double verification: a high information entropy value indicates that the user is in an exploration state, under which a continuously negative behavior efficiency indicator triggers the operation of reducing the complexity total budget Higher elements, thus reducing the presentation of non-core information on the interface, help users focus on core task areas in an information overload state.

[0036] After the analyst's skills eventually cover this advanced task and can complete the operation in a skilled and efficient manner, the behavior efficiency index Again, the complexity budget is converted into a sustained positive high value Also rises to a new stable high, and the interface eventually evolves into an efficient form with a comprehensive layout; during this process, even if a behavior efficiency index occasionally drops momentarily due to a false touch, the system will trigger a learning incentive state when it detects that the decision information entropy value of the model at this time is at a low level, temporarily suspending the adjustment of the budget due to a single negative incentive, in order to maintain the stability of the user's skilled workflow; the operation of this method is not based on the abstract concept of directly modeling user skills, but rather it builds a computable interface resource management closed loop driven by objective efficiency; it treats the user's growth, exploration, and even temporary loss during use as specific states in the computational model, and through the adjustment of a single variable, it continuously and seamlessly reconstructs the interaction between man and machine, transforming the interface itself from a static tool into a computational entity that changes with the user's ability.

[0037] Example 2: To objectively verify the actual effect of the technical solution of the present application in improving the long-term interaction efficiency of users, this example constructs and performs a controlled test aimed at quantitatively comparing the operation efficiency and task completion time of an interactive interface using the adaptive method of the present application and a traditional static full-featured interface during the process of a user from a novice to a skilled user; the test platform is built in an isolated network environment, using a workstation with standard input devices, which runs an application program developed for this test to simulate a complex professional-level data analysis software; the application program contains more than 200 independent functions, covering data import, cleaning, statistical modeling, visualization generation, etc.; the test recruited 20 participants with no experience using the specific software but with equivalent computer basic operation ability, and randomly assigned them to the control group and the test group, 10 people in each group; the application program interface used by the control group is a fixed static layout that presents all functions through multiple menus and toolbars; the application program used by the test group integrates the personalized interactive interface adaptive method based on reinforcement learning in the specific implementation; the core task flow of the test is designed to include 10 operation tasks with increasing complexity, starting from simple CSV file import, to medium-complexity sales monthly bar chart generation, to constructing a customer churn prediction logistic regression model that requires the combined use of multiple advanced functions; during the test, the system automatically records the time consumed by each participant to complete each task.

[0038] During the experiment, if the change trend of the task completion time of the two groups with the increase of the task complexity is compared, it can be observed that the two curves show significant differentiation; in the initial simple operation of task 1, the average task completion time of the two groups is close, which is 25.4 seconds and 24.8 seconds respectively; but since the third task, the curve corresponding to the test group starts to rise with a more gentle slope, while the curve of the control group becomes steeper; for example, in task 5, the average task completion time of the control group is 112.5 seconds, while that of the test group has dropped to 75.4 seconds; this gap is further expanded in subsequent high complexity tasks, and by task 9, the completion time of the control group reaches 355.2 seconds, while that of the test group is 180.1 seconds, which is less than half of the former; the efficiency improvement shown by the test group corresponds to the dynamic evolution of the complexity budget of the system, which steadily increases from the initial 20 to 190 at the end of task 10, which shows that the system actively places high-level function modules in an easily accessible location according to the continuous positive behavior efficiency index feedback obtained by the user, thereby reducing the search cost of the user; the test results confirm that the interactive interface using the method of the present application can shorten the time required to complete complex tasks compared with the traditional static interface during the skill growth period of the user; this shows that the behavior efficiency oriented interface complexity budget closed-loop management mechanism established by the method can effectively deal with the contradiction between the initial ease of use and the long-term efficiency of the interactive interface, and provides an interactive environment that can evolve together with the user's ability.

[0039] Embodiment 3: This embodiment describes a kind of individualized interactive interface adaptive method based on reinforcement learning, as Figs. 1 to 3 Fig. 1 ​​As shown, it depicts a closed-loop control system centered on a reinforcement learning decision model, which obtains multi-source input signals including user operation trajectory, device physical movement and environmental light intensity from external environment and user interaction, these signals are first input processed by a multi-source signal acquisition and purification module, the processed purified signals are sent to a behavior efficiency evaluation module for state quantization to generate a quantitative behavior efficiency index E, the index E is transmitted to a complexity budget dynamic updating module, which simultaneously receives a learning incentive state signal from a decision confidence analysis module, and based on the two inputs, the complexity total budget B is adjusted in strategy, the decision confidence analysis module identifies the user's exploration period by analyzing the internal state of the model output by the reinforcement learning decision model, the updated complexity total budget B is sent to the core decision unit of the reinforcement learning decision model as a core constraint, the model also receives a preset negative reward provided by the implicit cognitive load detection module when the user's cognitive impasse is detected, and finally outputs an interface adjustment action, which acts on the adaptive interface presentation module to complete the system output, thereby forming a dynamic adaptive closed loop in the continuous interaction between the user and the interface.

[0040] As shown in Fig. 2 , first, the common entry signal purification processing link is filtered and denoised, based on the purified signal, the system calculates the behavior efficiency index E representing the user's operation proficiency, which is mainly generated according to the time interval of interactive operation and whether there is invalid exploration behavior, the index E is input into the core reinforcement learning model as a reward signal. The model internally evaluates the expected return of each action through Q value calculation, and quantifies its decision confidence through information entropy calculation, at the same time, a parallel implicit cognitive load detection module can judge the cognitive state of the user through pupil response detection when there is no operation, and provide auxiliary input to the model, the model adjusts the complexity budget B dynamically according to the received signals and internal state, and finally selects elements under the constraint of not exceeding the budget B to determine a visible set whose sum of effective complexity cost of all elements satisfies ≤B, and then adjusts the interactive interface finally presented to the user, and iterates this process through the user feedback loop continuously.

[0041] As shown in Fig. 3As shown, the whole adaptive mechanism is deployed in a local execution environment without relying on cloud communication, ensuring response speed and data privacy. The architecture runs on typical user equipment, such as mobile diagnostic terminals or portable devices. At the hardware level, the system interacts with the human-machine interface, including a touch screen and an image sensor as a camera, and a sensor group including a motion sensor and an ambient light sensor; at the software level, the core is a personalized interface adaptive application as an application, which embeds multiple functional units including a signal purification module, a behavior efficiency evaluation module, a decision and budget updating module, and a cognitive load detection module. The application calls a reinforcement learning model as data work, and runs according to a system configuration and threshold file as configuration work, thereby forming a complete system that runs autonomously on the device side with software and hardware cooperation.

[0042] Embodiment 4: This embodiment provides a set of systematic, reproducible calibration and initialization procedures for key parameters and internal states in the foregoing technical solutions, to ensure that the core functions of the adaptive method can be deterministically implemented when deployed to a new target software application. In a specific deployment scenario, the adaptive method of the present application is applied to a third-party computer-aided design software. The CAD software interface is complex, and its end users often use portable devices in field environments including mobile vehicles and variable lighting. To enable the adaptive mechanism of the present application to function effectively in this specific application and its environment, an offline three-stage parameter calibration and model initialization process needs to be performed before delivery to the end user. The first stage is the environmental sensitivity coefficient calibration of interface elements. This process determines the sensitivity coefficients in the effective complexity cost calculation formula and . For a type of small-sized icon buttons in the CAD software interface, the calibration personnel repeatedly perform the clicking task on the button on a test platform that can simulate vibration. The platform generates vibrations with a controllable amplitude of 0-1.5g root mean square acceleration. The vibration is quantified as a motion influence factor . At the same time, adjustable light sources are used to change the ambient light, which is quantified as a light influence factor . During this process, the system records the average time and error rate of the user completing the clicking task under different and combinations, and combines them into an interaction difficulty index. Then, through curve fitting, the optimal and values are solved, so that the change trend of the value calculated according to the effective complexity cost calculation formula has the highest goodness of fit with the change trend of the measured interaction difficulty index. In a specific calibration step, a button with a diameter of 5mm has a calculated a value of 0.8, while a low-contrast text label has a value of 0.6, these coefficients are recorded as the attributes of the CAD software interface element library. a value of 0.6, these coefficients are recorded as the attributes of the CAD software interface element library.

[0043] The second stage is the determination of the system running threshold value; this procedure first calibrates the motion threshold value for distinguishing environmental noise, the procedure is to collect the motion sensor readings of the target portable device for one hour in the state of resting on the desktop, and collect the readings for one hour in the state of walking to simulate the on-site investigation, and calculate the probability distribution of the amplitude of the readings in the two states, and finally select the 99th percentile of the amplitude value in the resting state as the motion threshold value to distinguish between filtering jitter noise and retaining user intentional micro operations; in other words, this stage also needs to calibrate the decision confidence threshold value for distinguishing user states, the procedure is to invite a group of test users to use the CAD software to perform two types of tasks, one is repetitive drawing operation that has been mastered, and the other is exploratory task that needs to actively find and learn new commands, the system continuously records the information entropy value output by the reinforcement learning model during decision-making during this period, and forms two information entropy value distributions corresponding to the regular operation and the exploratory operation respectively, the confidence threshold value is finally set at the critical value that can maximize the distinction between the two distributions, the critical value is set at an entropy value point that can achieve a 95% discrimination accuracy by analyzing the overlap area of the two distributions; the third stage is the initialization of the reinforcement learning model; when the CAD software first loads the adaptive method of the application, the Q-Table of the Q-Learning reinforcement learning model used is initialized as a matrix with all state-action pairs having a Q value of zero; this zero initialization sets a starting point without preference, so that the model at the beginning of interaction with the first user, all its decisions completely depend on the immediate reward signal received, and any strategy formed thereafter is the result of learning from the actual interaction history of the user of this particular CAD software; by executing the complete procedure described above including the calibration of environmental sensitivity coefficients, the determination of running threshold value and the initialization of the model, the adaptive method of the application sets the specific numerical values of its internal key parameters based on objective data when it is deployed in a new target software, so that it can stably realize its adaptive function in different application ecosystems and hardware environments.

[0044] Embodiment 5: This embodiment presents the robustness of the core functional modules and the built-in risk avoidance mechanism of the aforementioned technical solutions in a complex high-interference real industrial environment; when the method of the present application is deployed on a mobile diagnostic terminal installed in a track detection vehicle, the system faces double and parallel environmental challenges; first, when the detection vehicle is driving on the track, the vehicle body and the equipment rack will produce multi-source non-stationary physical vibration, which will interfere with the accuracy of the trajectory data input by the operator through the touch screen; second, a fault warning light on the dashboard of the vehicle will flash at a fixed frequency of 2Hz, which will potentially interfere with the signal acquisition based on the visual channel; during the system operation, to cope with the physical vibration challenge, a motion sensor deployed on the terminal device collects real-time physical motion data of the device itself; the signal purification module in the system performs continuous short-time fast Fourier transform on the motion data stream; the analysis results show that there are wideband low-frequency vibrations corresponding to track joint impacts and a narrowband high-frequency vibration peak at 50Hz corresponding to the vehicle-mounted engine resonance in the frequency spectrum; accordingly, the system dynamically constructs and applies a composite filter composed of a wideband band-stop filter and a narrowband notch filter with a center frequency anchored at 50Hz; the composite filter is applied to the original user operation trajectory data, while retaining the user's intended non-periodic low-speed control signals, it suppresses the vibration noise from the above two sources, thereby extracting the purified operation trajectory data that reflects the user's operation intention for the subsequent behavior efficiency index calculation module.

[0045] During this period, if the system monitors that there is no any interaction behavior within the preset time window, the implicit cognitive load detection step is triggered; when starting the step, the system first performs a pre-emptive ambient light condition self-check procedure; through the ambient light sensor and image sensor of the device, the system performs spectral analysis on the ambient light signal, and identifies the 2Hz frequency flicker signal caused by the fault warning light, which is energy concentrated; due to the fact that the external interference frequency is close to the default frequency 3Hz of the system built-in and used to generate sub-threshold visual excitation signal, there is a risk of harmonic or mutual information calculation error, the control logic of the system automatically dynamically migrates the generation frequency of the excitation signal to a 5.5Hz frequency point far from the interference source and non-harmonic; then, the system generates a human eye sub-threshold visual excitation signal at the 5.5Hz frequency selected dynamically, and performs subsequent pupil response capture and mutual information value calculation; the procedure actively identifies and avoids potential environmental signal conflicts, so that the implicit cognitive load detection function can still maintain its accuracy in a complex environment with light interference; through the above-mentioned built-in response mechanism for physical vibration and visual interference, the method of the present application shows the ability of the core adaptive logic chain to reliably operate under multiple concurrent environmental pressures; the system dynamically adjusts its internal signal processing and signal generation parameters through signal level perception of the external environment, rather than relying on ideal operating conditions, thereby maintaining the stability and accuracy of its decision-making in a changing real industrial scene.

[0046] Embodiment 6: The present embodiment provides a pre-emptive, individualized physiological baseline calibration and judgment threshold determination procedure for the implicit cognitive load detection step in the foregoing technical solutions; in the scenario where a user first uses the air traffic control simulation training system loaded with the method of the present application, in order to enable the subsequent implicit cognitive load detection function to judge the cognitive state of the user, the system automatically performs a one-time user physiological baseline calibration program with a duration of 60 seconds before the user starts the formal training task; the first stage of the program is a 30-second natural pupil activity baseline collection; during this stage, the system interface presents a gray static screen without interactive elements, and instructs the user to naturally gaze at the center of the screen; at this time, the system does not emit any sub-threshold visual excitation signal, and its image sensor is only used to continuously capture the micro-jitter signal of the user's pupil and calculate the time series variance and basic spectral characteristics of the signal, which is stored as the individualized physiological noise baseline of the specific user for processing the captured pupil signal in subsequent real-time detection.

[0047] The second stage of the procedure is a 30-second stimulus-response mapping acquisition, which is further divided into two consecutive 15-second task windows; in the first task window, the system presents a low cognitive load smooth pursuit task to the user, and simultaneously applies a preset subthreshold visual stimulus signal to the screen background area, the system records the pupil response during this period, and calculates a set of mutual information values corresponding to the focused state; in the second task window, the system presents a high cognitive load mathematical task requiring the user to complete a multi-step mental calculation in short-term memory, and also applies a stimulus signal, the system records the pupil response during this period, and calculates another set of mutual information values corresponding to the cognitive overload state; wherein the calculation of each independent mutual information value follows the following procedure: the preset stimulus signal time series is time-aligned with the synchronously captured pupil microsaccade speed signal time series , and in a 2-second sliding window, first the amplitude of the two signal series is uniformly quantized and binned, specifically, the system determines the maximum and minimum values of the signal amplitude range collected during the entire calibration period, and divides this range into a preset number of equal-width quantization levels (e.g. 16 levels), thereby mapping the continuous signal to discrete levels, based on this quantization result, the system constructs its joint probability distribution and the respective marginal probability distributions and , and then calculates the mutual information value in this time window according to the information theory formula ; after completing the data acquisition of the above two stages, the system obtains the mutual information value probability distribution of the user in the focused and cognitive overload states, in order to determine the personal cognitive overload threshold of the user, the system uses the receiver operating characteristic (ROC) curve analysis method to find the optimal segmentation point, specifically, the system takes different mutual information values as candidate thresholds, calculates and draws an ROC curve, with the false positive rate on the horizontal axis and the true positive rate on the vertical axis, and finds a working point on the curve that can control the false positive rate within 2% while ensuring that the true positive rate is not less than 98%, the value corresponding to this working point is the final determined personal cognitive overload threshold, thus the one-time individual calibration procedure is completed, all key parameters of the built-in implicit cognitive load detection function in the system have been adapted to the physiological characteristics of the specific user, and are ready for use to provide cognitive state judgment in the subsequent actual training process.

[0048] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0049] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. A method for personalized interactive interface adaptation based on reinforcement learning, characterized in that, The method includes: Assign a base complexity cost to each UI element in the interactive interface. The system acquires user operation trajectory data from the human-machine interface event stream and device physical motion data from the motion sensor in parallel. If the motion state indicated by the device physical motion data exceeds the motion threshold stored in the user device, the user operation trajectory data is filtered based on the frequency characteristics of the device physical motion data to generate purified operation trajectory data. Otherwise, the user operation trajectory data is directly used as purified operation trajectory data. Calculate behavioral efficiency indicators based on purification operation trajectory data; Using a reinforcement learning model, an information entropy value is calculated based on the expected return value distribution of the reinforcement learning model itself within the decision cycle. If the information entropy value is not lower than the confidence threshold stored in the user device, the total complexity budget of an interface is updated based on the behavioral efficiency index. Under the constraint of not exceeding the total complexity budget, the reinforcement learning model selects a visible set of interface elements and adjusts the interface presentation according to the visible set. The method also includes: obtaining the ambient light intensity of the device through an ambient light sensor; calculating a context influence factor based on the device's physical motion data and the ambient light intensity; combining the basic complexity cost of the interface elements with the context influence factor to generate an effective complexity cost; and the constraint followed by the reinforcement learning model when making selections is that the sum of the effective complexity costs of each interface element in the visible set does not exceed the total complexity budget. effective complexity cost The computation of the effective complexity cost Ceff follows the following rules: wherein, Cbase is the base complexity cost, Cmotion is a motion impact factor generated from quantification of the device physical motion data, Clight is a light impact factor generated from quantification of the ambient light intensity, Cmotion and are the sensitivity coefficients of the interface element to motion and light respectively stored in the user device. 2.The method of claim 1, wherein, The steps for calculating the behavior efficiency index based on the purification operation trajectory data are as follows: obtain two adjacent interaction operations in the purification operation trajectory data; measure the time interval between the two adjacent interaction operations; determine whether there is an invalid exploration behavior in the purification operation trajectory data, such as the user opening the menu but not selecting any option in the menu and then closing the menu; and generate the behavior efficiency index based on the time interval and the occurrence of invalid exploration behavior. 3.The method of claim 1, wherein, The steps for updating the total complexity budget based on the behavioral efficiency index are as follows: when the value of the behavioral efficiency index is consistently higher than the positive efficiency threshold, the value of the total complexity budget is increased; when the value of the behavioral efficiency index is consistently lower than the negative efficiency threshold, and the information entropy value is not lower than the confidence threshold, the value of the total complexity budget is decreased. 4.The method of claim 1, wherein, The steps for calculating an information entropy value based on the expected reward distribution of the reinforcement learning model are as follows: after the reinforcement learning model calculates the expected reward value for each of the candidate interface element visibility adjustment actions; the set of all expected reward values ​​forms a probability distribution; and the information entropy value is calculated based on this probability distribution.

5. The method of claim 1, wherein, When the information entropy value is lower than the confidence threshold, the method triggers a learning incentive state. In the learning incentive state, the execution of operations that reduce the total complexity budget due to the value of the behavioral efficiency index being lower than the negative efficiency threshold is suspended.

6. The method of claim 1, wherein, The steps for filtering user operation trajectory data based on the frequency characteristics of device physical motion data are as follows: performing a fast Fourier transform on the device physical motion data to determine a jitter master frequency; constructing a digital band-stop filter, the center frequency of which is set according to the jitter master frequency; and using the digital band-stop filter to process the user operation trajectory data.

7. The method of claim 1, wherein, The method also includes: triggering an implicit cognitive load detection step when no user operation trajectory data is obtained within a continuous time window; the implicit cognitive load detection step includes generating a visual stimulus signal of subthreshold for the human eye through the interactive interface, and capturing the user's pupil micro-saccade signal through an image sensor; calculating the mutual information value between the pupil micro-saccade signal and the subthreshold visual stimulus signal; and providing a negative reward to the reinforcement learning model when the mutual information value exceeds the cognitive overload threshold.

8. The personalized interactive interface adaptive method based on reinforcement learning according to claim 1, characterized in that, All steps of the method, including acquisition, computation, judgment, selection, and adjustment, are executed locally on the user's device and do not rely on data communication with cloud servers to complete the decision-making of the reinforcement learning model.

Images