A game intervention control system based on autism double-time scale reinforcement learning

Abstract

The application relates to the technical field of artificial intelligence education assistance, in particular to a game intervention control system based on autism double-time-scale reinforcement learning, which comprises a multi-modal acquisition preprocessing module, a modal quality evaluation and reliability gating module, a cross-modal interaction state estimation module, a double-time-scale reinforcement learning module, a safety constraint checking module, a game interaction execution module and a course atlas management module; the multi-modal acquisition preprocessing module is used for synchronously acquiring multi-source interaction sensing data and game running logs, completing time alignment, denoising slicing and standardization processing, and generating time sequence window data with missing marks; through the modal quality evaluation and reliability gating mechanism, the multi-source sensing data are subjected to quality grading and dynamic weight distribution, the grading processing of five kinds of modal running states is supported, the stability of state estimation can be maintained under an abnormal scene, and intervention decision errors caused by single modal failure can be avoided.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence educational assistance technology, specifically a game intervention and control system based on dual-timescale reinforcement learning for autism. Background Technology

[0002] In recent years, digital interventions for children with autism have evolved from simple multimedia courseware to intelligent intervention systems that combine serious games, AR / VR interaction, and multimodal perception. Existing solutions typically adjust the difficulty of levels based on performance indicators such as game accuracy and task completion time, or introduce multimodal data such as facial expressions and motion capture to assist in assessing the child's condition. Some studies have attempted to achieve adaptive adjustments to game content through reinforcement learning.

[0003] However, in real-world intervention scenarios, issues such as changes in children's posture, hand occlusion, environmental noise, and device misalignment can cause dynamic fluctuations in sensor data quality. Conventional multimodal fusion solutions assume all modalities are stable and effective, and abnormal data can easily lead to misjudgments of state. Furthermore, existing adaptive solutions often focus on optimizing immediate feedback from single interactions, lacking systematic planning for long-term learning paths, which can lead to maximizing short-term rewards while neglecting long-term ability generalization. In addition, most systems lack rigid safety constraints, making it difficult to proactively avoid risks such as stimulus overload and continuous frustration when state estimation is uncertain. Essentially, this stems from the lack of a complete closed-loop control mechanism encompassing sensor quality verification, reliable state estimation, dual-scale decision-making, and safety boundary management, failing to simultaneously ensure robustness in complex scenarios, adaptability of the intervention process, and scientific rigor for long-term development. Summary of the Invention

[0004] The purpose of this invention is to provide a game intervention control system based on dual-timescale reinforcement learning for autism, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A game-based intervention and control system for autism based on dual-temporal-scale reinforcement learning includes:

[0007] The multimodal acquisition and preprocessing module is used to simultaneously acquire multi-source interactive sensor data and game operation logs, complete time alignment, noise reduction slicing and standardization processing, and generate time-series window data with missing data markers.

[0008] The modal quality assessment and credibility gating module is used to calculate the quality vector for each modal data in each time window, generate modal credibility and gating weights, divide the modal running state according to the credibility and execute the corresponding fusion strategy.

[0009] The cross-modal interaction state estimation module is used to fuse features from various modalities with gated weights and output the children's game interaction state vector and the corresponding state estimation uncertainty.

[0010] The dual-timescale reinforcement learning module includes short-timescale policy units and long-timescale policy units. The short-timescale policy units generate real-time game control actions based on the interaction state vector and uncertainty, while the long-timescale policy units update the course graph and long-term intervention path based on the lesson summary data.

[0011] The safety constraint verification module is used to perform safety boundary verification and trimming on the game control actions output by the short-time scale strategy unit, and output compliant safety control actions.

[0012] The game interaction execution module is used to drive the game terminal to run according to security control actions, adjust the game difficulty, prompt mechanism, feedback form and virtual character behavior, and generate new game operation logs;

[0013] The course graph management module is used to store and update the course graph, which consists of capability nodes, task nodes, scenario nodes, and evidence nodes. It receives course actions from long-term strategy units and iterates individualized intervention paths.

[0014] Furthermore, the working process of the multimodal acquisition and preprocessing module is as follows:

[0015] Integrate at least two types of sensor data sources, including facial video, voice and audio, eye tracking, limb skeleton, physiological signals, and game logs, and mark each data source with a unified timestamp;

[0016] Perform sampling rate alignment, noise filtering, and outlier removal on each data stream, and mark missing data, transmission delay, and abnormal acquisition locations;

[0017] The synchronized data is sliced ​​according to a preset sliding window length, and feature standardization is performed on the data in each window to generate time-series window data with missing and anomaly markers.

[0018] Furthermore, the working process of the modal quality evaluation and reliability gating module is as follows:

[0019] For single-modal data within each time window, calculate a quality vector that includes data completeness, signal-to-noise ratio, relative latency, missing rate, artifact metrics, and task effectiveness;

[0020] By combining the quality vector, historical stability, and task context, the credibility of a single mode is calculated, and the gating weights of each mode are generated through normalization.

[0021] Based on the credibility threshold, the single modality is divided into five operating states: normal participation, deweighted participation, frozen and maintained, alternative compensation, and manual confirmation, and corresponding fusion strategies are executed for each.

[0022] Furthermore, the working process of the cross-modal interaction state estimation module is as follows:

[0023] Each modal data within each time window is feature-encoded to obtain the feature representation of the corresponding modality.

[0024] We use gating weights to weight and fuse the feature representations of each modality to generate a unified cross-modality feature.

[0025] Based on cross-modal unified feature mapping, an interactive state vector containing participation, task load, failure risk, and social response level is obtained, while the uncertainty of the current window's state estimate is also output.

[0026] Furthermore, in the dual-timescale reinforcement learning module, the working process of the short-timescale policy unit is as follows:

[0027] Taking the interaction state vector, state estimation uncertainty and task context as input, the game control actions are output through a short-timescale reinforcement learning strategy. The action types include difficulty adjustment, prompt level, feedback form, interaction rhythm and scene parameters.

[0028] A short-term reward function is constructed, with participation and task progress as positive rewards and frustration risk, overload, and cue consumption as negative constraints, guiding the strategy to maintain children's participation while avoiding interaction risks.

[0029] Furthermore, in the dual-timescale reinforcement learning module, the working process of the long-timescale policy unit is as follows:

[0030] After a single intervention session, the status summary, task completion data, and safety event records for the entire session are aggregated to generate session summary data.

[0031] Using lesson summary data, current course map, stage intervention goals and historical learning trajectory as inputs, the course actions are output through long-term reinforcement learning strategies, including the selection of ability nodes, the arrangement of scenario transfer, the setting of review ratio and the adjustment of difficulty span;

[0032] Construct a long-term reward function, with positive rewards such as ability improvement, scenario generalization, and intervention stability, and negative constraints such as exit risk and safety events, to guide the curriculum path to continuously adapt to children's long-term development.

[0033] Furthermore, the working process of the security constraint verification module is as follows:

[0034] Preset safety boundary rules include difficulty limit, stimulus intensity, number of consecutive failures, and duration of a single session;

[0035] Receive game control actions output by short-timescale policies, and verify whether the actions conform to the safety boundaries by combining the current interaction state, uncertainty and session history;

[0036] Actions that exceed safety boundaries will be pruned, downgraded, or blocked; when the state uncertainty exceeds the limit or multiple modes are abnormal at the same time, a conservative mode is triggered, allowing only actions that reduce difficulty, reduce stimulation, pause, or request manual confirmation.

[0037] Furthermore, the working process of the course graph management module is as follows:

[0038] Construct a curriculum map with a four-level node structure: ability nodes correspond to intervention target dimensions, task nodes correspond to specific interactive tasks in the game, scenario nodes correspond to different application scenarios, and evidence nodes correspond to the behavior and status records during task execution.

[0039] After each task is completed, update the success rate, hint dependencies, and reaction time data of the corresponding evidence node and synchronize them to the corresponding capability node.

[0040] Receive course actions output from long-term strategy units, adjust the weights of ability nodes, the order of task nodes, and the ratio of scenario nodes, and generate an individualized intervention path for the next stage.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention uses modal quality assessment and credibility gating mechanisms to perform quality classification and dynamic weight allocation on multi-source sensor data, supports hierarchical processing of five modal operating states, and can maintain the stability of state estimation under common abnormal scenarios such as facial occlusion, voice noise, and device drift, avoiding intervention decision errors caused by single modal failure.

[0043] This invention employs a dual-timescale reinforcement learning architecture. The short-timescale strategy adjusts game interaction parameters in real time at the second / task level to ensure children's participation and smooth interaction. The long-timescale strategy iterates the curriculum map and intervention path at the lesson / stage level, taking into account both short-term interactive experience and long-term ability development, and avoiding training bias caused by simply pursuing immediate accuracy.

[0044] This invention features an independent safety constraint verification module that performs rigid boundary verification on all control actions. It supports the automatic triggering of a conservative mode when uncertainty exceeds the limit, effectively avoiding risks such as stimulus overload, continuous frustration, and emotional overload, and ensuring that the intervention process is safe and controllable.

[0045] This invention constructs a curriculum graph based on four-level nodes, which structurally links intervention goals, game tasks, application scenarios and process evidence. The intervention path is interpretable, traceable and can be manually intervened, adapting to the deployment needs of multiple scenarios such as institutions, families and assessment centers. At the same time, it supports local de-identification and federated update mechanisms, which meet the requirements for children's data privacy protection. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall structure of a game intervention control system based on dual-timescale reinforcement learning for autism according to the present invention;

[0047] Figure 2 This is a schematic diagram of the credibility gating and dual-timescale reinforcement learning process in this invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1-2 As shown, the present invention is a game intervention control system based on dual-timescale reinforcement learning for autism, including a multimodal acquisition and preprocessing module, a modality quality evaluation and credibility gating module, a cross-modal interaction state estimation module, a dual-timescale reinforcement learning module, a security constraint verification module, a game interaction execution module, and a curriculum graph management module.

[0050] The multimodal acquisition and preprocessing module is used to simultaneously acquire multi-source interactive sensor data and game operation logs, complete time alignment, noise reduction slicing and standardization processing, and generate time-series window data with missing data markers.

[0051] The specific workflow is as follows: The system integrates at least two types of sensor data sources, including facial video, audio, eye tracking, limb skeleton, physiological signals, and game logs. The sensor combinations can be flexibly configured according to the deployment scenario; for example, an institution can be configured with a complete multimodal acquisition suite, while a home system can retain only a tablet camera, microphone, and touch logs. The system marks each data stream with a timestamp based on a unified clock, performs sampling rate alignment, noise filtering, and outlier removal on each data stream, and marks the locations of missing data, transmission delays, and acquisition anomalies. Subsequently, the synchronized data is sliced ​​according to a preset sliding window length, which can be set from 0.5 seconds to 10 seconds depending on the task type, preferably 1 to 3 seconds. Feature standardization is performed on the data within each window, ultimately generating time-series window data with missing and outlier markers, which is then input into the subsequent modal quality assessment stage.

[0052] The modal quality assessment and credibility gating module calculates a quality vector for each modal data within each time window, generates modal credibility and gating weights, classifies modal operation states based on credibility, and executes corresponding fusion strategies. Specifically, for each single-modal data within a time window, it calculates a quality vector including data completeness, signal-to-noise ratio, relative latency, missing rate, artifact metrics, and task effectiveness. The quality vector is expressed as follows:

[0053]

[0054] in, Indicates the first Each modality within the time window The mass vector within; Indicates data integrity. Indicates signal-to-noise ratio or signal stability. Indicates relative delay, Indicates the missing rate. This indicates an artifact or anomaly. This indicates the task validity of the modality in the current game task. Task validity is determined in conjunction with the attributes of the current game task. For example, in a silent waiting task, the lack of voice is not considered a sensor malfunction, so the task validity score does not decrease.

[0055] The specific workflow is as follows: For each time-series window containing single-modal data, a quality vector is calculated, including data completeness, signal-to-noise ratio, relative latency, missing rate, artifact metrics, and task validity. Task validity is determined by considering the attributes of the current game task; for example, in a silent waiting task, missing voice is not considered a sensor malfunction, therefore the task validity score is not reduced. Combining the quality vector, the modal stability of historical windows, and the current task context, single-modal credibility is calculated using a mapping function, and then normalized to generate gating weights for each modality. The specific calculation process is as follows:

[0056]

[0057]

[0058] in, For the first The credibility of each modality This refers to the historical stability characteristics of this mode in the previous window. For task context features, , , For trainable weight matrix, For bias terms; For the normalized gating weights, Temperature is a parameter used to adjust the smoothness of the weight distribution. When the values ​​are larger, the modal weights are smoother. When the weight is smaller, the weight of the high-confidence mode is more prominent.

[0059] Based on a credibility threshold, single modalities are divided into five operational states: Normal Participation (corresponding to satisfactory credibility, with full weighted participation in fusion); Reduced Weighted Participation (corresponding to minor anomalies, retaining participation after reducing the modality's weight); Frozen / Hold State (corresponding to transient anomalies, using feature values ​​from the previous stable window to prevent sudden state changes); Replacement / Compensation State (corresponding to modality failure, using other related modalities or game logs for feature replacement); and Manual Confirmation State (corresponding to simultaneous anomalies in multiple modalities under critical tasks, suspending high-risk actions and prompting operators for confirmation). This five-state hierarchical processing ensures continuous and stable system operation in complex interaction scenarios, avoiding intervention interruptions or decision-making errors due to fluctuations in a single modality.

[0060] The cross-modal interaction state estimation module is used to fuse features from various modalities with gating weights and output the children's game interaction state vector and the corresponding state estimation uncertainty.

[0061] The specific workflow is as follows: Feature encoding is performed on each modality of data within each time window. Facial videos are processed using convolutional networks or visual Transformers to extract expression and posture features. Speech data is processed using Mel spectrum and acoustic models to extract emotion and vocalization features. Limb skeletons are processed using time-series graph convolution to extract motion features. Physiological signals are processed to extract time-frequency domain features. Game logs are processed to extract event sequence features. The feature representations of each modality are then weighted and fused using gating weights to generate a unified cross-modal feature. The fusion process is expressed as follows:

[0062]

[0063] in, For the encoding features of the corresponding modality, For the unified cross-modal features after fusion, The total number of modalities participating in the fusion. Feature fusion operators can be implemented using concatenation fusion, weighted summation, or attention fusion. Based on a cross-modal unified feature map, an interaction state vector is obtained, containing four core dimensions: engagement, task load, frustration risk, and social response level. Simultaneously, the uncertainty of the current window's state estimate is output. The state vector expression is as follows:

[0064]

[0065] in, Indicates participation level. Indicates cognitive or task load. This indicates a risk of frustration or overload. Indicates the level of social responsiveness. This represents the uncertainty in state estimation. All state dimensions are defined as behavioral and state indicators during the educational interaction process, without involving medical diagnostic conclusions. Engagement reflects the child's level of commitment to the current task; task load reflects the level of cognitive and operational stress; frustration risk reflects the probability of emotional breakdown and abandonment; and social response reflects the degree of feedback to social interaction cues. State estimation uncertainty is used to characterize the reliability of the current outcome, providing a basis for subsequent safety constraints and strategy decisions.

[0066] The dual-timescale reinforcement learning module includes short-timescale policy units and long-timescale policy units, which are responsible for real-time interactive control and long-term course evolution, respectively.

[0067] The short-timescale policy unit operates as follows: It takes the interaction state vector, state estimation uncertainty, and task context as input, and outputs game control actions through a short-timescale reinforcement learning policy. Action types include five categories: difficulty adjustment, prompt level, feedback format, interaction rhythm, and scene parameters. All actions are selected from a predefined action set, ensuring interpretable and traceable decisions. The output of the short-timescale policy is expressed as follows:

[0068]

[0069] in, For short-term control actions, For short-timescale policy networks, For short-time policy network parameters, This refers to the context features of the current task.

[0070] A short-term reward function is constructed, with increased participation and progress as positive rewards, and increased risk of frustration, overload, and resource consumption as negative constraints. This guides the strategy to maintain children's enthusiasm for participation while proactively mitigating interaction risks and preventing frustration caused by excessively rapid increases in difficulty. The short-term reward function is expressed as follows:

[0071]

[0072] in, As an engagement metric, For task progress indicators To defeat risk indicators, For overload indicators, To indicate the cost incurred; to These are the weighting coefficients for each sub-item, which can be configured according to the intervention objectives.

[0073] The long-term strategy unit works as follows: After each intervention lesson, it aggregates the average status of the entire lesson, task completion data, safety event records, and teacher evaluations to generate structured lesson summary data. Using the lesson summary data, the current course graph, the stage intervention goals, and historical learning trajectories as input, it outputs course actions through long-term reinforcement learning strategies, specifically including competency node selection, scenario transfer arrangements, review ratio settings, and difficulty range adjustments. The output of the long-term strategy is expressed as follows:

[0074]

[0075] in, For the first The long-duration course activities of the stage, For long-term policy networks, These are the parameters for the long-term policy network. Indicates the first Lesson or the first Summary data of lesson hours for each stage. This represents the current course map. Indicates the stage intervention goals. It represents the historical learning trajectory.

[0076] A long-term reward function is constructed, with positive rewards based on the magnitude of ability improvement, cross-scenario generalization effect, and stability of the intervention process, and negative constraints based on the risk of child withdrawal and the occurrence of safety incidents. This guides the curriculum path to continuously adapt to the child's long-term developmental pace, preventing short-term training goals from crowding out long-term ability generalization. The long-term reward function is expressed as follows:

[0077]

[0078] in, As a capability improvement indicator, As a generalized metric for specific scenarios, To intervene in stability indicators, To exit risk indicators, As an indicator of security incidents; to These are the weighting coefficients for each sub-item, which can be configured according to long-term intervention goals.

[0079] The security constraint verification module is used to perform security boundary verification and trimming on the game control actions output by the short-time-scale strategy unit, and output compliant security control actions.

[0080] The specific workflow is as follows: Pre-set safety boundary rules, covering dimensions such as upper limit of task difficulty, intensity of sensory stimulation, number of consecutive failures, maximum duration of a single session, and upper limit of prompt frequency. These rules can be configured according to individual child circumstances and institutional guidelines. After receiving the game control actions output by the short-timescale strategy, the system combines the current interaction state, state estimation uncertainty, and session history to verify whether each action conforms to the safety boundary. Finally, the system outputs the final safe control action through the safety constraint projection operator, expressed as:

[0081]

[0082] in, This represents the safety-constrained projection operator. It is a set of actionable actions consisting of preset rules, institutional security policies, and individualized restrictions. For the current session history, Preset security boundary rules.

[0083] Actions exceeding safety boundaries are pruned, downgraded, or blocked. For example, when consecutive failures reach a threshold, actions that increase difficulty are prohibited, and actions that reduce difficulty or add prompts are forcibly triggered. When the uncertainty of state estimation exceeds the limit, or multiple core modes simultaneously enter an abnormal state, the system automatically triggers a conservative mode, allowing only actions that reduce difficulty, decrease stimulation, pause the task, or request manual confirmation, prohibiting any operations that increase challenge or enhance stimulation, thus ensuring the safety of the intervention process from a mechanism perspective.

[0084] The game interaction execution module is used to drive the game terminal to run according to security control actions, adjust the game difficulty, prompt mechanism, feedback form and virtual character behavior, and generate new game operation logs.

[0085] The specific working process is as follows: The game terminal supports multiple forms such as AR devices, VR devices, tablets, and screen projection. After receiving security control action commands, it adjusts the difficulty parameters such as the number of targets, the complexity of steps, and the intensity of interference in the game. It also adjusts the level and frequency of visual cues, voice cues, and demonstration cues, switches feedback forms such as reward animations, encouraging voices, and background music, and controls the rhythm of the virtual character's movements, the intensity of facial expressions, and the interaction distance. Each time an action is executed or a task is completed, the system synchronously records a game event log, including information such as timestamp, action type, task result, and reaction time, and feeds it back to the multimodal acquisition and preprocessing module to form a real-time closed loop of perception-decision-execution.

[0086] The course graph management module is used to store and update the course graph, which consists of capability nodes, task nodes, scenario nodes, and evidence nodes. It receives course actions from long-term strategy units and iterates individualized intervention paths.

[0087] The specific work process is as follows: Construct a curriculum graph with a four-level node structure. The ability nodes correspond to intervention target dimensions such as eye contact, taking turns, emotion recognition, and rule understanding; the task nodes correspond to specific interactive tasks in the game, forming a many-to-many mapping with the ability nodes; the scenario nodes correspond to different application scenarios such as classrooms, homes, supermarkets, and hospitals, used to train scenario generalization ability; and the evidence nodes correspond to behavioral data and status records in each task execution, serving as an objective basis for ability assessment.

[0088] Each time a child completes a game task, the system automatically updates data such as the success rate, cue dependence, and average reaction time of the corresponding evidence node, and synchronously links this data to the upstream ability nodes to update the ability mastery assessment. After receiving the course actions output by the long-term strategy unit, the system adjusts the training weights of ability nodes, the order and difficulty gradient of task nodes, and the proportion of scenario nodes introduced to generate an individualized intervention path for the next stage. The course map is fully traceable and manually modifiable, allowing teachers or guardians to intervene at any time to adjust intervention goals and progress.

[0089] Example 1: Application of AR Interactive Classroom in Institutions

[0090] In AR interactive classroom scenarios at rehabilitation facilities or special education schools, the system is deployed on an edge computing host, equipped with cameras, depth cameras, microphone arrays, AR display devices, and a teacher control terminal. Optional accessories include eye trackers and physiological data acquisition devices. After a child enters the classroom, the system first performs short-term baseline data acquisition to calibrate the current device status and the child's baseline status. Then, it initiates social situational games, such as taking turns throwing a ball, facial expression recognition, and joint attention object finding training tasks.

[0091] During the interaction, if a child's hand obscures their face, causing a decrease in video modality quality, the modality quality evaluation and credibility gating module switches the video modality to a reduced-weight participation state, increasing the weight of the body movement and game log modalities to maintain stable state estimation. If excessive environmental noise causes the speech modality to fail, it switches to an alternative compensation state, replacing speech features with action and log features. The short-term strategy dynamically adjusts the cue density, task difficulty, and virtual character feedback based on the real-time interaction state. When the child's risk of frustration increases, the safety constraint verification module automatically blocks actions that increase difficulty and switches to a low-stimulation feedback mode. After a single lesson, the long-term strategy updates the course graph based on the lesson summary, determining the goal consolidation, scene transfer, or difficulty adjustment plan for the next lesson. Teachers can view and manually confirm modifications on the control panel.

[0092] Example 2: Lightweight Applications on Home Tablets

[0093] In a home setting, the system uses a regular tablet as its platform, collecting data solely through a built-in camera, microphone, and touch interaction, forming a lightweight deployment solution. The system retains a complete quality evaluation and trustworthiness gating mechanism. When the camera angle is poor or the lighting is insufficient, the video modality weight is automatically reduced, and state estimation is mainly based on data such as touch logs, response time, and error type. When the ambient noise is high, the voice modality is only used as an auxiliary reference.

[0094] All raw audio and video data on the home end is cached locally for real-time calculations and is not stored long-term by default; only anonymized event logs and course summaries are saved. Guardians can export a one-page course progress report after authorization. Course maps and intervention strategies can be synchronized from the institution to the home end, ensuring consistency in intervention pathways between institutions and families, while protecting children's data privacy throughout the process.

[0095] Example 3: Application of Auxiliary Records in the Assessment Center

[0096] In hospital or assessment center settings, the system serves as an auxiliary recording tool for structured interactive assessments. Assessors select a standardized task template, and the system executes the interactive task according to a preset process, automatically recording the child's gaze trajectory, motor performance, verbal responses, interaction logs, and state estimation results throughout. After the assessment, the system outputs a structured process report, including process indicators such as task duration, frequency of prompt responses, number of proactive responses, and transition adaptation performance, providing objective reference data for assessors. This embodiment only performs process recording and auxiliary analysis functions and does not output any medical diagnostic conclusions; all data is stored locally with encryption in accordance with medical data standards.

[0097] 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. A game intervention control system based on dual-temporal-scale reinforcement learning for autism, characterized in that, include: The multimodal acquisition and preprocessing module is used to simultaneously acquire multi-source interactive sensor data and game operation logs, complete time alignment, noise reduction slicing and standardization processing, and generate time-series window data with missing data markers. The modal quality assessment and credibility gating module is used to calculate the quality vector for each modal data in each time window, generate modal credibility and gating weights, divide the modal running state according to the credibility and execute the corresponding fusion strategy. The cross-modal interaction state estimation module is used to fuse features from various modalities with gated weights and output the children's game interaction state vector and the corresponding state estimation uncertainty. The dual-timescale reinforcement learning module includes short-timescale policy units and long-timescale policy units. The short-timescale policy units generate real-time game control actions based on the interaction state vector and uncertainty, while the long-timescale policy units update the course graph and long-term intervention path based on the lesson summary data. The security constraint verification module is used to perform security boundary verification and trimming on the game control actions output by the short-time scale strategy unit, and output compliant security control actions. The game interaction execution module is used to drive the game terminal to run according to security control actions, adjust the game difficulty, prompt mechanism, feedback form and virtual character behavior, and generate new game operation logs; The course graph management module is used to store and update the course graph, which consists of capability nodes, task nodes, scenario nodes, and evidence nodes. It receives course actions from long-term strategy units and iterates individualized intervention paths.

2. The game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the multimodal acquisition and preprocessing module, the process of simultaneously acquiring multi-source interactive sensor data and game operation logs, completing time alignment, noise reduction slicing, and standardization, and generating time-series window data with missing data markers is as follows: Integrate at least two types of sensor data sources, including facial video, voice and audio, eye tracking, limb skeleton, physiological signals, and game logs, and mark each data source with a unified timestamp; Perform sampling rate alignment, noise filtering, and outlier removal on each data stream, and mark missing data, transmission delay, and abnormal acquisition locations; The synchronized data is sliced ​​according to a preset sliding window length, and feature standardization is performed on the data in each window to generate time-series window data with missing and anomaly markers.

3. The game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the modal quality evaluation and credibility gating module, the process of calculating the quality vector for each modal data within each time window, generating modal credibility and gating weights, dividing the modal operating state based on credibility, and executing the corresponding fusion strategy is as follows: For single-modal data within each time window, calculate a quality vector that includes data completeness, signal-to-noise ratio, relative latency, missing rate, artifact metrics, and task effectiveness. By combining the quality vector, historical stability, and task context, the credibility of a single mode is calculated, and the gating weights of each mode are generated through normalization. Based on the credibility threshold, the single modality is divided into five operating states: normal participation, deweighted participation, frozen and maintained, alternative compensation, and manual confirmation, and corresponding fusion strategies are executed for each.

4. A game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the cross-modal interaction state estimation module, the process of fusing the features of each modality with gated weights and outputting the child's game interaction state vector and the corresponding state estimation uncertainty is as follows: Each modal data within each time window is feature-encoded to obtain the feature representation of the corresponding modality. We use gating weights to weight and fuse the feature representations of each modality to generate a unified cross-modality feature. Based on cross-modal unified feature mapping, an interaction state vector containing participation, task load, failure risk, and social response level is obtained, while the uncertainty of the current window's state estimate is also output.

5. A game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the dual-timescale reinforcement learning module, the process by which the short-timescale policy unit generates real-time game control actions based on the interaction state vector and uncertainty is as follows: Taking the interaction state vector, state estimation uncertainty and task context as input, the game control actions are output through a short-timescale reinforcement learning strategy. The action types include difficulty adjustment, prompt level, feedback form, interaction rhythm and scene parameters. A short-term reward function is constructed, with participation and task progress as positive rewards and frustration risk, overload, and cue consumption as negative constraints, guiding the strategy to maintain children's participation while avoiding interaction risks.

6. A game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the dual-timescale reinforcement learning module, the process by which the long-timescale policy unit updates the curriculum map and long-term intervention path based on lesson summary data is as follows: After a single intervention session, the status summary, task completion data, and safety event records for the entire session are aggregated to generate session summary data. Using lesson summary data, current course map, stage intervention goals and historical learning trajectory as inputs, the course actions are output through long-term reinforcement learning strategies, including the selection of ability nodes, the arrangement of scenario transfer, the setting of review ratio and the adjustment of difficulty span; Construct a long-term reward function, with positive rewards such as ability improvement, scenario generalization, and intervention stability, and negative constraints such as exit risk and safety events, to guide the curriculum path to continuously adapt to children's long-term development.

7. A game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the security constraint verification module, the process of performing security boundary verification and trimming on the game control actions output by the short-timescale strategy unit, and outputting compliant security control actions, is as follows: Preset safety boundary rules include difficulty limit, stimulus intensity, number of consecutive failures, and duration of a single session; Receive game control actions output by short-timescale policies, and verify whether the actions conform to the safety boundaries by combining the current interaction state, uncertainty and session history; Actions that exceed safety boundaries will be pruned, downgraded, or blocked; when the state uncertainty exceeds the limit or multiple modes are abnormal at the same time, a conservative mode is triggered, allowing only actions that reduce difficulty, reduce stimulation, pause, or request manual confirmation.

8. A game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, In the curriculum graph management module, the process of storing and updating the curriculum graph, which consists of capability nodes, task nodes, scenario nodes, and evidence nodes, and receiving curriculum actions from long-term strategy units to iterate individualized intervention paths, is as follows: Construct a curriculum map with a four-level node structure: ability nodes correspond to intervention target dimensions, task nodes correspond to specific interactive tasks in the game, scenario nodes correspond to different application scenarios, and evidence nodes correspond to the behavior and status records during task execution. After each task is completed, update the success rate, hint dependencies, and reaction time data of the corresponding evidence node and synchronize them to the corresponding capability node. Receive course actions output from long-term strategy units, adjust the weights of ability nodes, the order of task nodes, and the ratio of scenario nodes, and generate an individualized intervention path for the next stage.

9. A game intervention control system based on dual-temporal-scale reinforcement learning for autism according to claim 1, characterized in that, It also includes a privacy and federated update module, which performs desensitization processing on the raw sensor data collected locally, storing only event-level summaries and feature statistics; In cross-organizational collaboration scenarios, data is shared in the form of model parameters, gradients, or desensitized summaries, without transmitting the original audio, video, or identity information.

Images