AI-based scenario-based decision interface dynamic generation and continuous optimization method

Abstract

This invention relates to the field of artificial intelligence learning technology and discloses a method for dynamically generating and continuously optimizing a scenario-based decision-making interface based on AI. The method includes: acquiring a real-time user interaction behavior sequence, task context information, and an interface component library; fusing and encoding to generate a real-time task state vector; predicting the probability distribution of the interface components required for the next decision step using a sequence prediction model; dynamically combining and rendering the target interface according to the probability distribution and layout constraints; and calculating a reward value based on user feedback, and continuously optimizing the prediction model using reinforcement learning. This invention enables the interface to dynamically evolve and adaptively optimize according to user intent, improving interaction efficiency and user experience.

Description

Technical Field

[0001] This invention belongs to the field of machine learning, specifically relating to an AI-based method for dynamically generating and continuously optimizing scenario-based decision-making interfaces. Background Technology

[0002] As artificial intelligence technology penetrates deeper into complex decision-making scenarios such as enterprise operations, financial analysis, medical diagnosis, and intelligent customer service, users are placing higher demands on the intelligence and context-awareness of interactive interfaces.

[0003] Traditional decision support systems generally adopt static interface design. Their layout, controls, and information presentation methods are fixed after the system is deployed and cannot be dynamically adjusted according to the user's real-time decision status, task context, or historical behavior patterns.

[0004] Such static interfaces assume that the user's operation path is linear and predictable. However, the actual decision-making process is often highly nonlinear and exploratory, involving multi-step reasoning, information backtracking, and path switching.

[0005] When a user makes a decision, the system cannot proactively predict the user's next possible action, forcing the user to repeatedly switch between multiple independent pages, tabs, or functional modules. This not only prolongs the operation path but also disrupts the continuity of the decision-making process, reduces decision-making efficiency, and increases cognitive load.

[0006] AI-based dynamic generation technology for contextualized decision-making interfaces aims to adaptively reconstruct interface elements to match users' potential operational needs by sensing their current decision-making stage in real time and combining this with behavioral patterns from similar historical cases. The core of this technology lies in transforming the interface from a passively responsive tool into a proactive, collaborative intelligent agent, enabling it to proactively provide the most relevant function entry points, data views, or operational suggestions, thereby shortening the interaction steps required for users to achieve their goals.

[0007] In existing technologies, some systems attempt to introduce user behavior logs or clickstream data for interface optimization, but these are mostly limited to post-event analysis or static redesign driven by A / B testing, lacking the ability to model the causal logic of the decision-making process. Other methods, although using machine learning to predict the user's next action, generally rely on correlation statistics rather than causal inference, making it difficult to distinguish between accidental clicks and true intentions, and also unable to effectively handle counterfactual situations and intervention effects in the decision-making path.

[0008] Furthermore, existing solutions typically neglect the dynamic evolution of decision states, failing to deeply align the current context with historical cases at the semantic and structural levels. This results in weak generalization ability and poor scenario adaptability of prediction results. In high-complexity, high-risk decision-making scenarios, such deficiencies can easily lead to inaccurate interface recommendations, omissions of key operations, or information overload, severely restricting the fluency and reliability of human-machine collaborative decision-making. Therefore, there is an urgent need for an intelligent decision support method that can be based on causal inference mechanisms, deeply integrate current states and historical experience, and achieve dynamic interface generation and continuous optimization. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method for dynamic generation and continuous optimization of scenario-based decision-making interface based on artificial intelligence, which aims to overcome the defects of the existing technology, such as fixed user interface structure, inability to adapt to the dynamic decision-making process of users, resulting in long operation path, high cognitive load and low interaction efficiency.

[0010] To address the problems existing in the aforementioned background technology, this invention provides a method for dynamically generating and continuously optimizing a scenario-based decision-making interface based on artificial intelligence. This method includes the following steps:

[0011] The system acquires the user's real-time interaction behavior sequence in the decision-making interface, the current task context information, and the decision-making interface component library. The real-time interaction behavior sequence includes the user's click, hover, input, and scroll operations, as well as the corresponding timestamps and interface element identifiers. The current task context information includes the task type, task objective, and user role permissions. The decision-making interface component library pre-stores multiple atomic interface components and their associated metadata, including component function definitions, data input / output interface specifications, and layout constraint rules.

[0012] The real-time interaction behavior sequence and the current task context information are fused and encoded in a multimodal manner to generate a real-time task state vector that represents the user's current decision intention and task progress status.

[0013] Based on the real-time task state vector, the probability distribution of one or more interface components that the user will most need to use in the next decision step is predicted by a pre-trained sequence prediction model.

[0014] Based on the probability distribution, one or more high-probability interface components are selected from the decision interface component library, and the high-probability interface components are dynamically combined and rendered according to the layout constraint rules defined in the metadata, thereby generating a target decision interface for the next decision step.

[0015] The target decision interface is presented to the user, and the user's subsequent interaction with the target decision interface is continuously acquired, and the subsequent interaction is used as a feedback signal.

[0016] Based on the feedback signal, a reward value is calculated to evaluate the generation effect of the target decision interface. A reinforcement learning algorithm is then used to iteratively update the parameters of the sequence prediction model using the reward value, so as to continuously optimize the generation strategy of subsequent target decision interfaces.

[0017] As one embodiment of the present invention, obtaining the real-time interaction sequence of the user in the decision-making interface specifically includes:

[0018] By deploying an event listening script in the front-end application of the decision interface, the script captures all user interaction events in the document object model in real time. The data structure of the interaction event includes the event type, event timestamp, target interface element path identifier, mouse pointer coordinates, and keyboard input content.

[0019] The captured interactive event data is encapsulated into standardized log messages, which are key-value pair data structures. The log messages are then sent asynchronously to the backend data acquisition server for storage and processing via a secure Hypertext Transfer Protocol interface.

[0020] As one embodiment of the present invention, the step of performing multimodal fusion encoding on the real-time interaction behavior sequence and the current task context information to generate a real-time task state vector specifically includes:

[0021] First, the real-time interactive behavior sequence is preprocessed to convert discrete interactive events into a fixed-length action identifier sequence;

[0022] Next, the action identifier sequence is input into the first encoder of the long short-term memory network, and the network extracts the temporal dependency features in the user behavior sequence to generate a behavior feature vector.

[0023] Meanwhile, the task type, task objective, and user role permissions in the current task context information are vectorized through independent embedding layers, and then these vectors are concatenated to generate a context feature vector.

[0024] Finally, the behavioral feature vector is concatenated with the context feature vector to form a unified, high-dimensional real-time task state vector, which comprehensively represents the user's immediate decision-making state in a specific task context.

[0025] As one embodiment of the present invention, the step of predicting the probability distribution of one or more interface components that the user will most need to use in the next decision step based on the real-time task state vector and through a pre-trained sequence prediction model specifically includes:

[0026] The real-time task state vector is input into the encoder part of the transformer model based on the self-attention mechanism. The complex correlation between the features of each dimension inside the task state vector is captured by the multi-head self-attention layer to generate a deep context-aware state representation.

[0027] The state representation is input to the decoder part of the transformer model. The output layer of the decoder is a fully connected network with a normalized exponential function as its activation function. The number of output nodes of the network is equal to the total number of atomic interface components in the decision interface component library.

[0028] The decoder ultimately outputs a probability vector with a dimension equal to the total number of components. Each element in the vector represents the probability that the corresponding interface component will be needed by the user in the next decision step.

[0029] As one embodiment of the present invention, the step of dynamically combining and rendering the high-probability interface components according to the probability distribution specifically includes:

[0030] Set a preset probability threshold, and filter out all interface components whose probability values ​​are greater than the threshold from the probability vector to form a candidate component set;

[0031] Traverse the candidate component set and query the decision interface component library to obtain the metadata of each candidate component. The layout constraint rules defined in the metadata include the relative position relationship of the components, the size adaptation rules, and the explicit and implicit dependencies.

[0032] Start the constraint solving engine, take the layout constraint rules of all candidate components as input, and the constraint solving engine uses a backtracking search algorithm to calculate a conflict-free interface layout scheme that satisfies all constraints.

[0033] The layout scheme is defined in a structured description language format, which describes the precise position, size, and hierarchical relationship of each component in the interface grid system;

[0034] After receiving the layout scheme in the structured description language format, the front-end application calls the interface rendering engine, loads and instantiates the corresponding components from the component library according to the scheme, and finally presents the target decision interface on the user interface.

[0035] As one embodiment of the present invention, the step of calculating a reward value for evaluating the generation effect of the target decision interface based on the feedback signal, and iteratively updating the parameters of the sequence prediction model using a reinforcement learning algorithm specifically includes:

[0036] Define a quantitative reward function to evaluate the effectiveness of the target decision interface. The reward function is calculated as follows: the reward value equals the task completion efficiency weight multiplied by the reciprocal of the task completion time, plus the operation smoothness weight multiplied by the number of effective operations, minus the rework penalty weight multiplied by the number of navigation return operations.

[0037] The weighting coefficients for task completion efficiency, operational smoothness, and rework penalty are preset hyperparameters.

[0038] A reinforcement learning environment is constructed, wherein the real-time task state vector is defined as the state, the combination of interface components generated by the sequence prediction model is defined as the action, and the reward value calculated by the reward function is defined as the reward.

[0039] A proximal policy optimization algorithm is adopted to collect the state, action and reward sequence of the entire task process after the user completes a complete decision task;

[0040] Using the collected sequence data, the advantage function is calculated and a pruning objective function is constructed. The network weights of the sequence prediction model, i.e. the transformer model, are updated through gradient ascent, so that the actions generated by the model can maximize the cumulative expected return, thereby driving the interface generation strategy to continuously evolve towards a more efficient and user-intuitive direction.

[0041] This invention also provides a system for dynamically generating and continuously optimizing a scenario-based decision-making interface based on artificial intelligence, the system comprising:

[0042] The data acquisition module is used to acquire the user's real-time interactive behavior sequence, current task context information, and decision interface component library in the decision interface; the real-time interactive behavior sequence includes the user's click, hover, input, and scroll operations, as well as the corresponding timestamps and interface element identifiers; the current task context information includes the task type, task objective, and user role permissions; the decision interface component library pre-stores multiple atomic interface components and their associated metadata, the metadata including component function definitions, data input / output interface specifications, and layout constraint rules;

[0043] The task state vector generation module, connected to the data acquisition module, is used to perform multimodal fusion encoding on the real-time interactive behavior sequence and the current task context information to generate a real-time task state vector that represents the user's current decision intention and task progress status.

[0044] The interface component prediction module is connected to the task state vector generation module. It has a built-in pre-trained sequence prediction model, which is used to predict the probability distribution of one or more interface components that the user will most need to use in the next decision step based on the real-time task state vector.

[0045] The dynamic interface generation module, connected to the interface component prediction module and the data acquisition module, is used to select one or more high-probability interface components from the decision interface component library according to the probability distribution, and dynamically combine and render the high-probability interface components according to the layout constraint rules defined in the metadata, thereby generating a target decision interface for the next decision step.

[0046] The interface presentation and feedback module is connected to the dynamic interface generation module and is used to present the target decision interface to the user and continuously acquire the user's subsequent interaction behavior with the target decision interface, and use the subsequent interaction behavior as a feedback signal.

[0047] The model optimization module, connected to the interface presentation and feedback module and the interface component prediction module, is used to calculate a reward value for evaluating the generation effect of the target decision interface based on the feedback signal, and to use a reinforcement learning algorithm to iteratively update the parameters of the sequence prediction model using the reward value to continuously optimize the generation strategy of subsequent target decision interfaces.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. By capturing user interaction behavior sequences in real time and encoding them in conjunction with the task context, a task state vector that can accurately represent the user's dynamic intentions is generated. This makes the basis for interface generation decisions no longer a static user profile or preset process, but the user's current real and immediate decision needs, thus realizing the interface's deep perception of user intentions.

[0050] 2. By adopting a sequence prediction model based on self-attention mechanism, it can effectively capture the complex dependencies in the user's long-term operation sequence and accurately predict the functional components that the user is most likely to need in the next decision step. This realizes the transformation of the interaction mode from "people looking for functions" to "functions waiting for people", which greatly shortens the user's operation path.

[0051] 3. A dynamic interface generation mechanism was created, which combines and renders the interface in real time based on the prediction results and preset layout constraints. This breaks the shackles of traditional fixed layouts, allowing the interface structure to evolve in real time with the user's decision-making process. This provides users with a highly coherent and seamless operating environment, reducing the user's cognitive load and context switching costs.

[0052] 4. A closed-loop optimization mechanism based on reinforcement learning is introduced, which quantifies the interaction between the user and the dynamically generated interface into reward signals. The prediction model is continuously and automatically iterated and optimized, enabling the system's interface generation strategy to have the ability to learn adaptively and evolve continuously. This allows it to continuously improve the accuracy of its predictions and the effectiveness of the generated interface, thereby maintaining the best user experience in the long term. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall technical solution architecture of the AI-based scenario-based decision-making interface dynamic generation and continuous optimization method proposed in this invention.

[0054] Figure 2 This is a schematic diagram of the core principle framework of the sequence prediction model based on the self-attention mechanism in this invention;

[0055] Figure 3 This is a flowchart illustrating the multimodal fusion encoding logic of real-time interactive behavior sequences and task context information in this invention.

[0056] Figure 4 This is a logical flow diagram of the dynamic combination and constraint-driven rendering of high-probability interface components in this invention.

[0057] Figure 5 This is a schematic diagram of the closed-loop feedback and reinforcement learning optimization mechanism between the user and the dynamically generated interface in this invention;

[0058] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow of the front-end event listening, back-end data collection and model service working together in this invention. Detailed Implementation

[0059] Please refer to Figures 1 to 6 This invention provides a method for dynamically generating and continuously optimizing a scenario-based decision-making interface based on artificial intelligence, aiming to solve the technical problems of long operation paths, high cognitive load, and low interaction efficiency caused by the rigid structure of traditional user interfaces and their inability to adapt to the dynamic decision-making process of users.

[0060] This method dynamically predicts and generates the interface components required for the next decision step by sensing the user's interactive behavior and task context in real time, and introduces a reinforcement learning mechanism to continuously optimize the prediction model, thereby realizing an intelligent interactive experience in which the interface structure evolves in real time with the user's decision-making process.

[0061] The method includes the following steps: S1, obtaining the user's real-time interactive behavior sequence in the decision interface, the current task context information, and the decision interface component library;

[0062] S2, perform multimodal fusion encoding on the real-time interaction behavior sequence and the current task context information to generate a real-time task state vector representing the user's current decision intention and task progress status;

[0063] S3, based on the real-time task state vector, predict the probability distribution of one or more interface components that the user will most need to use in the next decision step through a pre-trained sequence prediction model.

[0064] S4. Based on the probability distribution, select one or more high-probability interface components from the decision interface component library, and dynamically combine and render the high-probability interface components according to the layout constraint rules defined in the metadata, thereby generating a target decision interface for the next decision step.

[0065] S5, present the target decision interface to the user, and continuously acquire the user's subsequent interaction behavior with the target decision interface, and use the subsequent interaction behavior as a feedback signal;

[0066] S6. Based on the feedback signal, calculate the reward value for evaluating the generation effect of the target decision interface, and use the reinforcement learning algorithm to iteratively update the parameters of the sequence prediction model using the reward value to continuously optimize the generation strategy of the subsequent target decision interface.

[0067] In step S1, the real-time interaction sequence of the user in the decision-making interface, the current task context information, and the decision-making interface component library are obtained. The real-time interaction sequence includes the user's clicks, hovers, inputs, scrolling operations, and the corresponding timestamps and interface element identifiers.

[0068] The current task context information includes the task type, task objective, and user role permissions.

[0069] The decision interface component library pre-stores multiple atomic interface components and their associated metadata, which includes component function definitions, data input / output interface specifications, and layout constraint rules.

[0070] Specifically, an event listener script is deployed in the front-end application of the decision interface, which captures all user interaction events in the document object model in real time.

[0071] The data structure of the interactive event includes the event type, event timestamp, target UI element path identifier, mouse pointer coordinates, and keyboard input. Event types include, but are not limited to, clicks, double clicks, right-click menu triggers, text input, dropdown selections, slider drags, and page scrolling.

[0072] The target interface element path identifier uses a hierarchical path expression method, such as "main panel / sidebar / filter / date range selector", to ensure that each interactive event can be uniquely mapped to a specific interface component instance.

[0073] Mouse pointer coordinates are recorded in pixels relative to the current viewport to help determine the user's focus area. Keyboard input, excluding sensitive information, is recorded as a semantic summary, such as "a query containing the keyword 'budget' was entered."

[0074] The captured interaction event data is encapsulated into standardized log messages.

[0075] The log message is a key-value pair data structure, whose fields include event identifier, session identifier, user identifier, device type, browser environment, network latency index, and the specific content of the aforementioned interaction event.

[0076] The log message is sent asynchronously to the backend data acquisition server in real time via a secure Hypertext Transfer Protocol interface.

[0077] The data acquisition server performs integrity verification, deduplication, and time-series alignment on the received log messages, and finally persists them in a distributed time-series database for subsequent steps to access.

[0078] Meanwhile, the current task context information is generated and maintained by the task management system when the user starts the decision task.

[0079] The task type identifies the business area to which the current task belongs, such as "financial approval", "supply chain scheduling" or "customer risk assessment".

[0080] The task objective describes the specific results expected to be achieved in this task, such as "completing the quarterly budget allocation plan" or "identifying a list of high-risk clients." User role permissions are provided by a unified identity authentication system, which clarifies the scope of operations that users can perform and the data dimensions they can access in the current task, such as "department managers can only view data within their own department" or "auditors have read-only permissions."

[0081] The above three pieces of information constitute a complete task context, serving as a key background for understanding the user's behavioral intent.

[0082] The decision-making interface component library is stored in the central component registry and uses a version-based management mechanism.

[0083] Each atomic UI component is an independently encapsulated front-end module with a clear input / output contract.

[0084] The component function definition describes its core business logic. For example, the bar chart visualization component is used to display the comparative relationship of multidimensional classification data.

[0085] The data input / output interface specification clearly defines the data structure, format, and triggering conditions required by the component, such as receiving a two-dimensional array containing timestamps and values, and automatically redrawing when the data is updated.

[0086] Layout constraints define the spatial behavior of components in the interface, including minimum / maximum width and height, whether they are collapsible, their relative positional dependence on other components (e.g., component A must be to the right of component B), and visibility conditions (e.g., component C is only displayed when the user's role is administrator).

[0087] All metadata is stored in a structured configuration file, supporting hot updates and dynamic loading.

[0088] In step S2, the real-time interactive behavior sequence and the current task context information are multimodal fusion encoded to generate a real-time task state vector.

[0089] First, the real-time interactive behavior sequence is preprocessed. The original interactive event stream is sorted by timestamp, and the most recent event is extracted. Each event constitutes a behavior window. The default fixed length is typically between 50 and 200. Each event is mapped to a unique action identifier based on its type and target element path identifier, forming a discrete sequence of action identifiers. This sequence is standardized to a fixed length through padding or truncation operations, with missing parts filled with special placeholders.

[0090] Next, the action identifier sequence is input into the first encoder of the Long Short-Term Memory (LSTM) network. This LSM network contains two layers of hidden units, each with 256 units. The network effectively mitigates the vanishing gradient problem through a gating mechanism, enabling it to capture long-term dependencies in user behavior sequences. For example, a user clicking the "Export" button and then selecting PDF format in a pop-up dialog box can be accurately modeled as a cross-step operation. The encoder ultimately outputs a 128-dimensional behavioral feature vector, which encapsulates the user's recent operational history and potential intentions.

[0091] Simultaneously, the task type, task objective, and user role permissions in the current task context information are vectorized through independent embedding layers. The task type embedding layer maps predefined task categories to 60-dimensional dense vectors; the task objective, after extracting keywords from the natural language processing module, is generated into an 80-dimensional vector through word embedding average pooling; the user role permissions are converted into 32-dimensional binary vectors through permission bitmasks and then compressed into 40-dimensional real-number vectors through a fully connected layer. The above three vectors are concatenated to form a 180-dimensional context feature vector.

[0092] Finally, the 128-dimensional behavioral feature vector is concatenated with the 180-dimensional context feature vector to form a unified 308-dimensional real-time task state vector. This vector comprehensively represents the user's real-time decision-making state in a specific task context, including historical behavioral patterns and incorporating task objectives and permission constraints, providing high-information-density input for subsequent component prediction.

[0093] In step S3, based on the real-time task state vector, a pre-trained sequence prediction model is used to predict the probability distribution of one or more interface components that the user will most need to use in the next decision step. The sequence prediction model is a transformer model based on a self-attention mechanism. The encoder part of this model contains a structure of alternating stacks of 6 multi-head self-attention layers and feedforward neural networks. Each layer has 8 self-attention heads, and the hidden layer dimension is 512.

[0094] The 308-dimensional real-time task state vector is input into the encoder part of the transformer model. The encoder uses a multi-head self-attention mechanism to compute the correlation weights between the features of each dimension within the state vector in parallel. For example, when the task type is financial analysis and the user has just performed the operation of filtering data for this quarter, the model will automatically enhance the correlation weights between the year-on-year and month-on-month calculation components and the trend prediction chart components. After 6 layers of encoding, a 512-dimensional deep context-aware state representation is output.

[0095] This state representation is fed into the decoder part of the transformer model. The decoder also contains a 6-layer structure, with its self-attention mechanism focusing on the predicted component sequence (initially empty), while the cross-attention mechanism focuses on the state representation output by the encoder. The output layer of the decoder is a fully connected network, and the number of its output nodes equals the total number of atomic interface components in the decision interface component library. The activation function of this fully connected network is a normalized exponential function, ensuring that all elements of the output vector are non-negative and sum to 1.

[0096] Decoder final output dimensional probability vector ,in Indicates the first The probability distribution of a UI component represents the probability that it will be needed by the user in the next decision-making step. This probability distribution reflects the model's quantitative judgment of the user's intention for the next action and serves as the direct basis for dynamic UI generation.

[0097] In step S4, based on the probability distribution, one or more high-probability interface components are selected from the decision interface component library. These high-probability interface components are then dynamically combined and rendered according to the layout constraints defined in the metadata, thereby generating the target decision interface. A preset probability threshold is set. Typically, the value ranges from 0.15 to 0.3. From the probability vector... Filter out all that meet the criteria The component indexes form a candidate component set. .

[0098] Traverse the candidate component collection Query the decision interface component library to obtain each candidate component. The complete metadata is extracted. The focus is on extracting layout constraints, including component size ranges, anchor point positions, spatial dependencies with other components, and explicit / implicit conditions. All constraints are formalized into first-order logical predicates, such as... or .

[0099] The constraint solving engine is activated, employing a backtracking search algorithm combined with a local consistency pruning strategy. The engine takes the layout constraints of all candidate components as input and attempts to assign position and size to each component within a predefined interface grid system (e.g., a 12-column grid). During the search, if any assignment is found to violate any constraint, the engine immediately backtracks and tries other combinations. The search terminates when the first conflict-free layout solution that satisfies all constraints is found.

[0100] The layout scheme is defined in a structured description language format, which uses a nested object structure. Each component node contains its unique identifier, grid starting column, number of columns spanned, hierarchical index, and dynamic attribute binding expression.

[0101] After receiving the structured layout scheme via a message queue, the front-end application invokes the UI rendering engine. The rendering engine parses each component node in the scheme, loads the corresponding atomic component module from a local or remote component library, instantiates the component, and passes in the bound data source and configuration parameters. All components are arranged according to... The specified hierarchical order is used to overlay and draw the final target decision interface, which is presented on the user interface with a compact structure and focused functions.

[0102] In step S5, the target decision interface is presented to the user, and subsequent user interactions with the target decision interface are continuously acquired and used as feedback signals. All user operations on the new interface—including using, ignoring, modifying, or returning to the previous interface—are fully recorded by the event listening script and used as the original basis for evaluating the effect of this interface generation.

[0103] In step S6, based on the feedback signal, a reward value is calculated to evaluate the generation effect of the target decision interface, and a reinforcement learning algorithm is used to iteratively update the parameters of the sequence prediction model. A quantized reward function is defined. The calculation formula is as follows:

[0104] ;

[0105] The time taken for the user to complete the current subtask from the moment the interface appears, in seconds; The number of valid actions performed by the user on the target interface is defined as either direct use of the prediction component or subsequent actions based on its output. The number of times a user performs a navigation back action reflects the degree to which the interface does not meet expectations; , , These are the task completion efficiency weight, operation smoothness weight, and rework penalty weight, respectively. They are all preset positive real number hyperparameters, with typical values ​​of 1.0, 0.5, and 2.0, respectively.

[0106] Construct a reinforcement learning environment, in which states The real-time task state vector generated in step S2, action 'a' being the component combination corresponding to the probability distribution output in step S3 (after threshold filtering), and reward 'r' being the scalar value calculated by the reward function R mentioned above. A proximate policy optimization algorithm is used as the learning strategy. After the user completes a full decision-making task, the system collects the state-action-reward sequence throughout the entire task process. .

[0107] Using this sequence, we can calculate the generalized advantage estimate and construct a sheared policy gradient objective function:

[0108] ;

[0109] This represents the probability ratio between the old and new strategies. This is the estimated value of the dominance function. The clipping range hyperparameter is typically 0.2. This objective function is maximized using stochastic gradient ascent, updating all network weights of the transformer model. This process makes the model more inclined to generate component combinations that bring high rewards in subsequent predictions, thereby driving the interface generation strategy to continuously evolve towards efficiency and intuitiveness.

[0110] The system includes a data acquisition module, a task state vector generation module, an interface component prediction module, a dynamic interface generation module, an interface presentation and feedback module, and a model optimization module.

[0111] The data acquisition module is responsible for real-time collection of user interaction behavior and task context, and maintaining the component library. The task state vector generation module implements multimodal fusion encoding. The interface component prediction module incorporates the aforementioned transformer model. The dynamic interface generation module integrates a constraint solving engine and a rendering scheduler. The interface presentation and feedback module manages front-end display and behavior feedback. The model optimization module executes reinforcement learning training loops.

[0112] Each module works in concert with shared memory via a message bus to form a closed-loop intelligent interface generation and optimization system.

[0113] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish an entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based dynamic generation and continuous optimization method for a scenario-based decision interface, characterized in that, The method comprises the following steps: acquiring a real-time interaction behavior sequence of a user in a decision interface, current task context information, and a decision interface component library; the real-time interaction behavior sequence includes the user's click, hover, input, scroll operation, and corresponding timestamp and interface element identifier; the current task context information includes task type, task target, and user role permission; the decision interface component library pre-stores a plurality of atomized interface components and their associated metadata, the metadata including component function definition, data input / output interface specification, and layout constraint rules; multi-modal fusion coding is performed on the real-time interaction behavior sequence and the current task context information to generate a real-time task state vector representing the user's current decision intention and task progress state; based on the real-time task state vector, a pre-trained sequence prediction model is used to predict the probability distribution of one or more interface components most needed by the user in the next decision step; according to the probability distribution, one or more high-probability interface components are selected from the decision interface component library, and the high-probability interface components are dynamically combined and rendered according to the layout constraint rules defined in the metadata, thereby generating a target decision interface for the next decision step, including: setting a preset probability threshold to filter out all interface components with a probability value greater than the threshold from the probability distribution vector to form a candidate component set; traverse the candidate component set, query the decision interface component library to obtain the metadata of each candidate component, and the layout constraint rules defined in the metadata include the relative position relationship, size adaptation rules, and visibility dependency relationship of the components; start a constraint solving engine, input all the layout constraint rules of the candidate components, and the constraint solving engine uses a backtracking search algorithm to calculate a conflict-free interface layout scheme that satisfies all the constraint conditions; the layout scheme is defined in a structured description language format, which describes the exact position, size, and hierarchical relationship of each component in the interface grid system; after the front-end application program receives the layout scheme in the structured description language format, it calls the interface rendering engine to load and instantiate the corresponding components from the component library according to the scheme, and finally presents the target decision interface on the user interface; present the target decision interface to the user and continuously acquire the subsequent interaction behavior of the user with the target decision interface, and use the subsequent interaction behavior as a feedback signal; based on the feedback signal, calculate a reward value for evaluating the generation effect of the target decision interface, and use a reinforcement learning algorithm to iteratively update the parameters of the sequence prediction model using the reward value to continuously optimize the generation strategy of the subsequent target decision interface; the pre-processing of the real-time interaction behavior sequence converts discrete interaction events into fixed-length action identifier sequences, including: The original interaction event stream is sorted by timestamp and truncated to the most recent events to form a behavior window, of a preset fixed length; each event is mapped to a unique action identifier according to its type and target element path identifier to form a discrete action identifier sequence; perform padding or truncation operation on the action identifier sequence to make it uniform in length, and fill the missing part with a special placeholder.

2. The AI-based contextualized decision interface dynamic generation and continuous optimization method according to claim 1, characterized in that, The real-time interaction behavior sequence of the user in the decision interface is acquired, including: An event listening script is deployed in a front-end application of the decision interface, which captures all user interaction events in a document object model in real time, and a data structure of the interaction events includes an event type, an event occurrence timestamp, a target interface element path identifier, a mouse pointer coordinate, and keyboard input content; The captured interaction event data is encapsulated into a standardized log message, which is a key-value pair format data structure, and the log message is sent to a data collection server in the back end in real time in an asynchronous manner through a secure hypertext transfer protocol interface for storage and processing.

3. The AI-based contextualized decision interface dynamic generation and continuous optimization method of claim 2, wherein, The real-time interaction behavior sequence and the current task context information are multi-modal fusion coded to generate a real-time task state vector representing a current decision intention and a task progress state of the user, including: The real-time interaction behavior sequence is preprocessed to convert discrete interaction events into a fixed-length action identifier sequence; The action identifier sequence is input into a first encoder of a long short-term memory network to extract time sequence dependent features in the user behavior sequence through the network to generate a behavior feature vector; Task type, task target, and user role permission in the current task context information are respectively vectorized through independent embedding layers, and then the vectors are spliced to generate a context feature vector; The behavior feature vector and the context feature vector are spliced to form a unified high-dimensional real-time task state vector.

4. The AI-based contextualized decision interface dynamic generation and continuous optimization method of claim 3, wherein, Based on the real-time task state vector, a probability distribution of one or more interface components most needed by the user in the next decision step is predicted through a pre-trained sequence prediction model, including: The real-time task state vector is input into an encoder part of a transformer model based on a self-attention mechanism, and complex associations between features in each dimension of the task state vector are captured through a multi-head self-attention layer to generate a state representation with deep context awareness; The state representation is input into a decoder part of the transformer model, and an output layer of the decoder is a full connection network with a normalized exponential function as an activation function, and the number of output nodes of the network is equal to the total number of atomized interface components in the decision interface component library; The decoder finally outputs a probability vector with a dimension equal to the total number of components, and each element value in the vector represents a probability of a corresponding interface component being needed by the user in the next decision step.

5. The AI-based contextualized decision interface dynamic generation and continuous optimization method of claim 4, wherein, The action identifier sequence is input into a first encoder of a long short-term memory network, including: The long short-term memory network includes 2 layers of hidden units, and each layer has 256 units; The first encoder outputs a 128-dimensional behavior feature vector that condenses the historical trajectory and potential intention of the user's recent operation.

6. The AI-based contextualized decision interface dynamic generation and continuous optimization method of claim 5, wherein, The task type, task target, and user role permission in the current task context information are respectively vectorized through independent embedding layers, including: A task type embedding layer maps predefined task categories to 60-dimensional dense vectors; After the task goal is extracted by the natural language processing module, an 80-dimensional vector is generated by word embedding average pooling; User role permissions are converted into a 32-dimensional binary vector through permission bit masking and then compressed into a 40-dimensional real vector through a fully connected layer.

7. The AI-based contextualized decision interface dynamic generation and continuous optimization method of claim 6, wherein, The real-time task state vector is input into the encoder part of the transformer model based on the self-attention mechanism, including: The encoder part of the transformer model contains a structure of 6 layers of multi-head self-attention layers and feedforward neural networks stacked alternately; Each layer has 8 self-attention heads and a hidden layer dimension of 512; After 6 layers of encoding, a 512-dimensional deep context-aware state representation is output.

8. The AI-based contextualized decision interface dynamic generation and continuous optimization method of claim 7, wherein, The constraint solving engine is started, and the layout constraint rules of all candidate components are input, including: The layout constraint rules are formalized into first-order logic predicates; The constraint solving engine assigns a position and size to each component in the predefined interface grid system; The search terminates when the first conflict-free layout scheme that meets all the constraints is found.

Images