A cognitive training task and ancillary content self-adaptive distribution method and system

Abstract

The application discloses a cognitive training task and auxiliary content self-adaptive distribution method and system, comprising the following steps: step S1, establishing a cross-modal constraint associated index for a cognitive training task; step S2, receiving a cognitive training task completion event and performing candidate retrieval of a corresponding content unit identification list; step S3, updating a current task continuous difficulty parameter by adopting a dead zone proportional control law; step S4, writing the task into a double-track event table and updating a state according to an EWMA recursive formula; step S5, mapping a continuous difficulty output by the control law into a discrete gear to show the training terminal; and step S6, selecting at least one content unit according to a preset selection strategy and outputting to the training terminal. The application adopts four processes of distribution-control-feedback-interface, and forms a self-consistent feedback closed loop on a data stream.

Description

Technical Field

[0001] This invention relates to the field of medical information, and in particular to a method, system, and device for adaptive distribution of cognitive training tasks and auxiliary content. Background Technology

[0002] Existing technologies, in providing structured task-content association storage and event-driven distribution, real-time difficulty adaptive control, time-weighted estimation of reach and execution feedback, and continuous control and discrete display coordination for large-scale users, optimize the four functional modules of task and auxiliary content distribution, task difficulty adjustment, reach / compliance feedback collection, and continuous control-discrete display interface independently, lacking closed-loop coordination. This mainly presents the following four problems:

[0003] (i) In terms of the distribution layer, existing computerized training tools and content distribution tools generally treat them as two independent modules, with no data linkage between task execution and content distribution; even if there is related data, they usually use unconstrained binary relation tables, lacking cross-modal coverage integrity constraints, resulting in related data possibly containing only a single modality, unable to be incrementally expanded at runtime, and having no usable feedback baseline when new users are cold-started. Some existing technologies use runtime semantic matching based on graph neural networks or pre-trained models (e.g., traversing the knowledge graph through meta-path rules, dynamically calculating cross-modal similarity through attention mechanisms). The retrieval time complexity of such schemes increases to the level of O(logn) to O(n) as the size of the knowledge graph grows, which cannot meet the requirement of O(1) deterministic table lookup, and semantic matching depends on runtime model reasoning, and the storage structure itself fails to provide hard constraints on cross-modal integrity.

[0004] (II) Regarding the control layer, existing training tools mostly use preset discrete difficulty levels (such as L1-L5), which are manually selected by users or pre-configured by administrators. They fail to provide an adaptive adjustment mechanism based on task performance data, making it difficult to maintain the user's continuous training state within the flow range (accuracy approximately 70-85%). Some adaptive methods (such as arithmetic step size, reinforcement learning policy gradient, window counting heuristic, and biosignal threshold triggering) suffer from problems such as arithmetic step size easily causing sudden shifts in perceived difficulty, reinforcement learning having a large sample size and low interpretability, and heuristic methods lagging in responding to boundary conditions. Even with closed-loop control with a dead zone, the dead zone width in existing schemes is usually a global constant, failing to consider the historical variability between users and tasks. Stable users are obscured at higher resolutions, while fluctuating users oscillate repeatedly at the dead zone boundary. Furthermore, each task controller is independent, failing to provide a cross-task capability transfer mechanism within the same cognitive domain, requiring each task within the same domain to cold-start from the baseline. In the field of industrial automatic control, dead zone functions (such as valve or motor dead zone compensation) are usually C¹ smooth approximations (such as tanh inverse models), which are different from the piecewise linear form of this invention. Moreover, their application scenario is to eliminate the perception blind zone caused by physical nonlinearity (passive compensation), which is opposite to the purpose of this invention to actively introduce dead zone to suppress oscillations.

[0005] (III) Regarding the feedback layer, existing solutions typically merge "content has been distributed to the user" and "the user actually executed the content" into a single "completion status" Boolean flag (with no more than 1 bit of effective information). This makes it impossible to analyze the two different types of signals, namely, reach and execution, independently, resulting in low resolution of the feedback signal. If cumulative statistics are used, the complete event history must be stored to calculate the time-weighted compliance rate, and the space overhead increases linearly with the number of events. Even if time-weighted estimation is used, if it is only updated when a compliance event is triggered, the estimated value for discontinued users does not decay and cannot reflect the time sensitivity of the compliance status. Exponentially weighted moving average (EWMA), as a classic technique in the field of statistical process control, has been known to be implemented in applications such as quality control charts, abnormal flow detection, battery temperature monitoring, and EEG signal sample screening. However, none of these applications directly use the EWMA output to drive the error signal smoothing of the feedback control law.

[0006] (iv) Regarding the interface layer, the control layer needs to run the control law (minimum step size) in a continuous parameter space (d ∈ [0, 1]) to ensure control accuracy. The step size is approximately 0.03, while the display to end users requires discrete gears (L ∈ {1, ..., M}, where M is typically 5, corresponding to a gear interval ΔL = 0.25) to avoid cognitive burden caused by numerical readings. If the control law is run directly in the discrete gear space, the step size of 0.03 is easily swallowed up by the gear quantization loss of 0.25, and the control cannot respond precisely to the accuracy deviation. If the control is continuous and the original d value is displayed, the gear will repeatedly switch near the boundary (d = 0.25 ± 0.02), causing visual jumps. Some existing technologies provide quantization mapping or threshold switching (e.g., Schmitt Trigger), but these are single-layer structures and do not form a collaborative architecture of "continuous control layer + discrete display layer + bidirectional homomorphic positive and negative quantizers + bidirectional hysteresis of gear hysteresis". Summary of the Invention

[0007] The purpose of this invention is to provide an adaptive distribution method and computer system for cognitive training tasks and auxiliary content, which solves the above four interrelated technical problems in an integrated manner through closed-loop coordination of distribution, control, feedback and interface during task execution.

[0008] The above four types of technical problems must be solved jointly within a self-consistent technical solution; solving any one category alone cannot eliminate the overall technical defects. The following are examples of problems: Lack of distribution layer integrity: No available candidate pool exists during cold starts for new users; regardless of the quality of control / feedback / interface, the distribution decision lacks a baseline. Lack of control layer refinement: The flow band does not converge; regardless of the quality of distribution / feedback / interface, the difficulty-capability match fails. Lack of independent dual-track feedback layer: The feedback signal does not exceed 1 bit or the spatial overhead increases linearly; regardless of the quality of distribution / control / interface, the control law lacks high-resolution input. Lack of interface layer quantization hysteresis: Continuous control is swallowed up by gear quantization or oscillates at gear boundaries; regardless of the quality of distribution / control / feedback, the coordination between the control layer and the presentation layer fails.

[0009] To achieve the above-mentioned objectives, this invention provides a method for adaptive distribution of cognitive training tasks and auxiliary content, comprising the following steps:

[0010] Step S1: Establish a cross-modal constraint association index for the cognitive training task;

[0011] Step S2: Receive the cognitive training task completion event and perform a candidate retrieval of the corresponding content unit identifier list;

[0012] Step S3: Update the current task's continuous difficulty parameters using a dead-zone proportional control law;

[0013] Step S4: Write the task into the dual-track event table and update the status according to the EWMA recursive formula. The dual-track event table consists of an reach event table and a compliance event table. The two tables are not merged into a single completion status field.

[0014] Step S5: Map the continuous difficulty of the control law output to discrete levels for display to the training terminal;

[0015] Step S6: Select at least one content unit according to the preset selection strategy and output it to the training terminal; if the content unit constitutes a new touch event, write it back to the touch event table as the state variable of the next round of feedback control closed loop.

[0016] Furthermore, in step S1, the associated index adopts a hash access structure, with the key being the task identifier and the value being a list of content unit identifiers. The hash access structure consists of two associated tables in a relational database. A hash cache with the task identifier as the key is established in memory to support runtime O(1) retrieval. The two associated tables are the task table and the task content unit associated table. Alternatively, an independent key-value data structure (such as an in-memory hash table, Redis Hash, or embedded KV storage) can be used to achieve equivalent hash access.

[0017] Furthermore, in step S2, the completion event carries at least a user identifier, a task identifier, an accuracy rate (expressed as a percentage), and a completion indicator. The at least one processor normalizes the accuracy rate (mapping the percentage [0, 100] to [0, 1]). If the accuracy rate is empty (e.g., the user leaves early or does not complete any valid trials), the completion event does not trigger a control law update, and the current difficulty parameter remains unchanged to avoid invalid samples contaminating the control state. The candidate retrieval performs hash addressing in the associated index P using the task identifier as the key, and retrieves the corresponding content unit identifier list in O(1) time complexity as candidates for subsequent output selection. The retrieval is based on key-based hash addressing, which does not rely on probability prediction of user behavior or runtime classification of training questions.

[0018] Furthermore, in step S3, the dead-zone proportional control law takes the accuracy deviation e as input. It does not adjust when the deviation falls within the dead-zone half-width B. When the deviation exceeds the dead zone, the adjustment amount is calculated according to the linear step size coefficient k and output after step size upper and lower limit saturation and continuous interval projection. The formula is expressed as follows:

[0019] ,

[0020] ,

[0021] Where T∈(0,1) is the flow target, B>0 is the dead zone half-width, and k>0 is the linear step size coefficient. ≤ Here, represents the upper and lower limits of the step size, Proj represents the projection over a continuous interval, sat is the saturation operator, and sgn is the sign function. Let [i] be the continuous difficulty parameter for task i before the t-th update, with a value range of [0,1]. Let be the normalized accuracy corresponding to the completion event of task i in the t-th task, with a value in the range [0,1]. The input is a feedback control function with dead zone, which is used to output the amount of continuous difficulty adjustment for this operation. For the error signal, when its input is the above error signal hour, Output the amount of difficulty adjustment for this consecutive iteration.

[0022] Furthermore, the dead zone half-width B is upgraded from a global constant to a user-task history variability function. It is calculated in the following form. The formula is expressed as follows:

[0023] ,

[0024] ,

[0025] in > 0 is the lower limit of the dead zone. > 0 represents the fluctuation sensitivity coefficient, [ , ] represents the feasible interval of the dead zone, σ is the rolling sample standard deviation of the accuracy of the most recent N games, and N is the length of the rolling window. This represents the mean of the normalized accuracy within the scrolling window. The rolling sample standard deviation of the normalized accuracy of user u in task c over the most recent N rounds, calculated based on an N-1 correction, with typical values ​​as follows: =0.05, =0.5, N=5, [ , = [0.05, 0.15].

[0026] Furthermore, a co-domain coupling term is superimposed on the basic step size to achieve mutual calibration of task capabilities within the same cognitive domain and strict isolation across domains. The formula is as follows:

[0027] ,

[0028] in Let λ be the set of tasks in the cognitive domain where task i is located, and λ ∈ [0, ..., ... ] represents the coupling coefficient. ≥0 represents the coupling weight and =1; when j does not belong to hour = 0, meaning cross-domain tasks are strictly isolated.

[0029] Furthermore, in step S4, the reach event table records the system output time, and the compliance event table records the user execution receipt; for each user-content unit pair (u, c), the EWMA (Exponentially Weighted Moving Average) time-weighted compliance rate is calculated using the following recursive formula:

[0030] ,

[0031] in ∈ (0, 1) is the attenuation coefficient. ∈ {0, 1} indicates whether the corresponding compliance event occurs within the preset time window for the t-th touch event. EWMA estimates that each (u, c) pair only occupies O(1) space and does not require storing the complete event history.

[0032] Furthermore, in step S4, a batch processing job is executed according to a preset scan cycle (typically 1 day). For expired notifications that have exceeded the preset time window but still have no corresponding compliance event, an EWMA decay update with x = 0 is triggered. The purpose of this formula is to automatically decay the compliance rate of discontinued users at a geometric rate—by triggering an EWMA update with x = 0 for each non-compliance notification event in each scan cycle.

[0033] ,

[0034] Persistent decay and deduplication logs ensure that each expired message triggers decay at most once (idempotency). The compliance rate of users who have stopped training converges geometrically, strictly guaranteeing time sensitivity.

[0035] Furthermore, in step S5, the control law operates on the continuous difficulty space d ∈ [0, 1]. The presentation layer maps the continuous difficulty to discrete gears L ∈ {1, ..., M} through positive vectorization, and uses reverse quantization to deduce continuous values ​​from the gears for initialization, as shown in the following formula:

[0036] ,

[0037] ,

[0038] Where d is the continuous difficulty parameter of the control law output, with a value range of [0,1]; M is the total number of discrete gears and M>=2; Q(d) is the positive vectorization function that maps the continuous difficulty parameter d to discrete gears L; L is the discrete gears of the display layer, with a value of {1,...,M}; This is the inverse quantization function that derives the continuous difficulty representative value from the discrete gear L; This represents the continuous difficulty value derived from gear L.

[0039] Furthermore, maintain the user's current display level. A two-way hysteresis zone is introduced near the gear boundary to suppress frequent gear jumps caused by small fluctuations in continuous difficulty.

[0040]

[0041] in Let i be the midpoint boundary of the continuous values ​​between adjacent gear positions i and i+1. >0 is the hysteresis threshold. This refers to the gear level ultimately displayed to the training terminal. The control law reads and writes continuously in the range of difficulty 'd'. The hysteresis mechanism only affects the display direction and does not provide feedback to influence the control law. Typical value: M=5. =0.05. When M=5, b1=0.125, b2=0.375, b3=0.625, b4=0.875.

[0042] Another aspect of the present invention provides a cognitive training task and auxiliary content adaptive distribution system, including an association index management module, a receiving module, a dead-zone control module, a compliance rate EWMA module, a quantization mapping module, and an output module, wherein:

[0043] The association index management module is used to build association indexes for cross-modal constraints for cognitive training tasks;

[0044] The receiving module is used to receive cognitive training task completion events and perform candidate retrieval of the corresponding content unit identifier list;

[0045] The dead-zone control module is used to update the continuous difficulty parameters of the current task using a dead-zone proportional control law;

[0046] The compliance rate EWMA module is used to write tasks into a dual-track event table and update the status according to the EWMA recursive formula. The dual-track event table consists of an reach event table and a compliance event table. The two tables are not merged into a single completion status field.

[0047] The quantization mapping module is used to map the continuous difficulty of the control law output into discrete levels for display to the training terminal;

[0048] The output module is used to select at least one content unit according to a preset selection strategy and output it to the training terminal.

[0049] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the method described above.

[0050] Compared with existing technologies, this system and method have the following advantages:

[0051] This invention establishes an association index by verifying cross-modal coverage constraints before deployment, and locates candidates in O(1) time complexity at runtime; new user cold starts do not rely on feedback data. Cross-modal integrity constraints ensure that each association covers at least two modalities, eliminating the loss of comprehensive support value from single-modal associations.

[0052] Under steady-state conditions with zero mean deviation and bounded variance, this invention employs a dead-zone proportional control law to ensure that continuous difficulty parameters tend to and remain within the difficulty zone corresponding to the flow interval. Monte Carlo simulations, based on 5 independent random seeds and 30 simulated user samples for each seed, verify that this invention improves the flow zone convergence rate by approximately 43.4% compared to ML decision-making methods, with a robustness range of 27%-58% across the 5 random seeds.

[0053] In this invention, the reach and compliance events are maintained using independent data tables, and the resolution of the feedback signal is improved from a single Boolean bit to a continuous EWMA value (3-4 bits of valid information). The continuous control layer and the discrete presentation layer are decoupled through bidirectional homomorphic mapping of positive and negative quantization. The control law operates in continuous space to ensure accuracy, and the discretization of the presentation layer avoids cognitive burden.

[0054] This invention employs a four-step process—distribution, control, feedback, and interface—to form a self-consistent feedback loop in the data flow: the distribution layer outputs tasks / content to the user terminal; after user execution, two types of objective events are generated: reach and compliance; the feedback layer recursively updates the EWMA time-weighted compliance rate on a dual-track event table; the control layer runs a proportional control law with a dead zone in the continuous difficulty space; and the interface layer maps continuous difficulty to discrete levels visible to the user through continuous-discrete homomorphic quantization and level hysteresis, sending these levels back to the next round of distribution selection. This constitutes a closed feedback control structure, achieving good results in task and auxiliary content association distribution, difficulty adaptive control, reach / compliance feedback collection, and continuous control-discrete display interface collaboration for cognitive impairment assisted training scenarios. Attached Figure Description

[0055] Figure 1 This is a flowchart of an adaptive distribution method for cognitive training tasks and auxiliary content.

[0056] Figure 2The diagram shows the hash storage structure of the associated index, where (32) is the associated index P data, (70) is the associated index management module, (72) is the cross-modal integrity pre-checker, and (75) is the pre-check rejection write output.

[0057] Figure 3 The diagram shows the closed-loop relationship of the data flow, where (34) is the event table, (36) is the compliance event table, (70) is the associated index management module, (80) is the dead zone ratio control module, (90) is the compliance rate EWMA module, (92) is the periodic decay sub-module, (100) is the quantization mapping module, and (106) is the content unit of the comprehensive selection output.

[0058] Figure 4 The diagram shows the input-output partitioning of a proportional control law with dead zone, where (74) is the input terminal of the deviation signal e, (76) is the output terminal of the control law, (80) is the dead zone half-width threshold B, (81) is the negative out-of-band region, (82) is the dead zone, (83) is the positive out-of-band region, and (86) is the interval projection.

[0059] Figure 5 This is a schematic diagram of state-related dead zone calculation, where (80) is the dead zone half-width threshold interface with dead zone ratio control module, (82) is the dead zone sub-interval, (82a) is the rolling sample standard deviation calculation end, and (82b) is the dead zone half-width output end.

[0060] Figure 6 The diagram shows the task topology of intra-domain coupling and cross-domain isolation, where (60) is the set of tasks in the attention cognitive domain, (62) is the set of tasks in the memory cognitive domain, (64) is the set of tasks in the executive function cognitive domain, (66) is cross-domain strict isolation, (68) is the coupling coefficient and normalization constraint, and (84) is the intra-domain coupling weight.

[0061] Figure 7 The diagram shows the time sequence of the EWMA recursive estimation of compliance rate with event number, where (34) is the reach event table, (36) is the compliance event table, and (90) is the compliance rate EWMA module.

[0062] Figure 8 The diagram shows the dual-track event table storage structure, where (34) is the reach event table, (36) is the compliance event table, (38) is the decay deduplication log, (90) is the compliance rate EWMA module, and (92) is the periodic decay submodule.

[0063] Figure 9This is a schematic diagram of the non-compliance path periodic decay scan operation process, where (34) is the reach event table, (36) is the compliance event table, (38) is the decay deduplication log, (90) is the compliance rate EWMA module, and (92) is the periodic decay sub-module.

[0064] Figure 10 This is a schematic diagram of continuous-discrete double-layer homomorphic quantization and gear hysteresis characteristics, where (76) is the continuous difficulty input, (77) is the equivalent continuous difficulty, (78) is the display gear output, (100) is the quantization mapping module, (102) is the gear hysteresis sub-module, (103) is the boundary midpoint, and (104) is the hysteresis region.

[0065] Figure 11 The diagram shows the gear lag determination process, where (76) is the continuous difficulty input, (78) is the gear output display, (102) is the gear lag submodule, (103) is the boundary midpoint, and (104) is the lag zone determination branch.

[0066] Figure 12 This is a schematic diagram of the overall architecture of a cognitive training task and auxiliary content adaptive distribution system, where (10) is the system as a whole, (20) is the processor, (30) is the memory, (40) is the network interface, (50) is the training terminal, (70) is the associated index management module, (80) is the dead zone ratio control module, (90) is the compliance rate EWMA module, and (100) is the quantization mapping module.

[0067] Figure 13 The diagram shows the multi-terminal access topology, where (10) is the overall system, (40) is the network interface, (42) is the semantic field whitelist filter, (50) is the training terminal, (52) is the management terminal, and (54) is the related person terminal. Detailed Implementation

[0068] 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.

[0069] The cognitive training tasks in this invention are structured interactive activities presented by a computer terminal, typically including tasks such as attention selection, short-term memory, executive function, spatial orientation, and digit processing. During task execution, the system can collect execution data, including at least completion rate and accuracy, as input for subsequent decision-making. The auxiliary content units consist of various types of information items related to the user's daily life, typically including categories such as physical exercise, nutritional guidance, social activities, and health knowledge.

[0070] like Figure 1 The diagram shown is a flowchart of the method of the present invention. An embodiment of the present invention provides a method for adaptive distribution of cognitive training tasks and auxiliary content, the specific steps of which are as follows:

[0071] Step S1: Establish a cross-modal constraint association index P for the cognitive training task. For example... Figure 2 The associated index P shown adopts a hash access structure, with the key being the task identifier (taskId) and the value being the content unit identifier list (itemIdList). For any given taskId, the retrieval time complexity is O(1). Each itemIdList must satisfy the cross-modal coverage integrity constraint, that is, the cardinality of the modal type set corresponding to the list is not less than 2. Entries that do not satisfy this constraint are rejected during the write phase. The associated index P is established before system deployment through cross-modal integrity pre-verification and supports runtime incremental updates.

[0072] This embodiment uses two relational database tables (task table + task-content unit relational table) to implement the hash access structure, and establishes a hash cache in memory with the task identifier as the key to support runtime O(1) retrieval; alternatively, an independent key-value data structure (such as memory hash table, Redis Hash, embedded KV storage) can be used to implement equivalent hash access.

[0073] In this embodiment, the interface path is POST / api / admin / catalog / init, which reads a preset set of associated items (41 items in this embodiment), each item being in the form of {taskId: "A1", itemIdList: ["ex-01", "nu-03", "so-05"]}. For each item, the cross-modal integrity pre-validator (72) is called: it queries the modal type tag of each content unit in itemIdList and determines the cardinality of its modal set; if the cardinality of the modal set is less than 2, the item is rejected from being written and a rejection log is recorded, such as... Figure 2 Label (75); if it is not less than 2, the entry is written to the associated index P and its hash cache.

[0074] Example: If an itemIdList contains only the body movement modality and the cardinality of the modality set is 1, the item is rejected; if the itemIdList contains three modalities—body movement, nutrition, and social—and the cardinality is 3 ≥ 2, the item is written. During runtime, hash addressing is performed in the associated index P using the task identifier as the key, and the list of content unit identifiers is retrieved in O(1) time complexity; it also supports incremental addition of items through the management interface. Newly added items are also pre-verified for cross-modal integrity before being written. After successful writing, the hash cache is incrementally updated without restarting the service process.

[0075] Step S2: Receive the cognitive training task completion event and perform a candidate retrieval of the corresponding content unit identifier list.

[0076] like Figure 3 As shown, at the end of each task session, the training terminal (50) submits a task completion event to the processor (20) via the network interface (40). The completion event carries at least the user identifier, task identifier, accuracy (expressed as a percentage), and completion indicator. At least one processor normalizes the accuracy (mapping the percentage [0, 100] to [0, 1]); if the accuracy is empty (e.g., the user leaves early or does not complete any valid trials), the completion event does not trigger a control law update, and the current difficulty parameter remains unchanged to avoid invalid samples contaminating the control state. At least one processor performs hash addressing in the associated index P with the task identifier as the key, retrieving the corresponding list of content unit identifiers in O(1) time complexity as candidates for subsequent output selection. The retrieval is based on key-based hash addressing, which does not rely on probability prediction of user behavior or runtime classification of training questions.

[0077] Step S3: Update the current task's continuous difficulty parameters using a dead-zone proportional control law.

[0078] like Figure 4 As shown, the determination process for updating the continuous difficulty parameters of the current task using the dead zone proportional control law is as follows: if the absolute value of the deviation is less than half the width of the dead zone, the task falls into the dead zone. Figure 4 Label (82)), return to the original difficulty unchanged; if the absolute value of the deviation is not less than the dead zone half width, out-of-band trigger ( Figure 4 Label (81) negative out-of-band region or (83) positive out-of-band region), the calculation step size is the product of the sign function and the saturation operator, and the difficulty is updated after continuous interval projection output.

[0079] Control law based on accuracy deviation As input (T is the flow target), no adjustment is made when the deviation falls within the dead zone half-width B. When the deviation crosses the dead zone, the adjustment amount is calculated according to the linear step size coefficient k and output after step size upper and lower limit saturation and continuous interval projection. When the deviation between the user's current accuracy and the flow target is large, the difficulty is adjusted downward or upward proportionally. Small deviations falling into the dead zone are not adjusted, and the single step size has upper and lower limit protection, as shown in the following formula:

[0080]

[0081] Define a "feedback control law with dead zone" using a piecewise function - output 0 (no adjustment) when the deviation is less than half the dead zone width, otherwise output the correction amount in the form of linear step size + upper and lower limit saturation.

[0082]

[0083] Where T∈(0,1) is the flow target, B>0 is the dead zone half-width, and k>0 is the linear step size coefficient. ≤ Here, represents the upper and lower limits of the step size, Proj represents the projection over a continuous interval, sat is the saturation operator, and sgn is the sign function. Let [i] be the continuous difficulty parameter for task i before the t-th update, with a value range of [0,1]. Let be the normalized accuracy corresponding to the completion event of task i in the t-th task, with a value in the range [0,1]. The input is a feedback control function with dead zone, which is used to output the amount of continuous difficulty adjustment for this operation. For the error signal, when its input is the above error signal hour, Output the amount of difficulty adjustment for this consecutive iteration.

[0084] The test scenario was a new user task A1, with an initial difficulty of 0.300. The accuracy rate was 90% (0.90 after normalization). Parameters: T = 0.775, B = 0.075, k = 0.15. = 0.03、 = 0.08. Calculate the deviation e = 0.90 − 0.775 = 0.125; |e| = 0.125 > B = 0.075, out-of-band trigger; naive step size k · |e| = 0.15 × 0.125 = 0.01875; after truncation by the saturation operator, the step size is 0.030 (lower limit effective); since the deviation is positive, the difficulty needs to be increased, so the step size is +0.030; the updated difficulty is equal to the projection of Proj onto 0.300 + 0.030 in the interval [0, 1], resulting in 0.330, the theoretical value for byte-level matching. The updated difficulty is persisted to the difficulty state table, and the previous difficulty is saved as the prev_d field to support subsequent presentation layer processing.

[0085] In a more preferred embodiment, the dead zone half-width B is upgraded from a global constant to a user-task history volatility function B(u, c, t), i.e., state-dependent dead zone hardening, calculated in the following form:

[0086] ,

[0087] The above formula upgrades the dead zone half-width from a global constant to an adaptive value based on the user's recent performance fluctuations—stable users receive fine-tuning with a small dead zone, while volatile users receive oscillation suppression with a large dead zone. Then, the sample standard deviation of the user's accuracy over the most recent N games (adjusted based on N-1) is calculated and used as the volatility measure in the previous formula: The formula is as follows:

[0088]

[0089] in > 0 is the lower limit of the dead zone. > 0 represents the fluctuation sensitivity coefficient, [ , ] represents the feasible interval of the dead zone, σ is the rolling sample standard deviation of the accuracy of the most recent N games, and N is the length of the rolling window. This represents the mean of the normalized accuracy within the scrolling window. The rolling sample standard deviation of the normalized accuracy of user u in task c over the most recent N rounds is calculated based on an N-1 correction.

[0090] like Figure 5 As shown, this embodiment enables the function via the system switch `algo.v2.enabled = true`. Before step S4, a dynamic calculation of the dead zone half-width is performed: the user's most recent N-round accuracy records for this task are queried (N = 5 in this embodiment), the sample standard deviation is calculated using the formula above, and then the dead zone half-width used for this update is determined by the truncation function. One specific value in this embodiment is... = 0.05, β = 0.5, [ , = [0.05, 0.15]. In the case of insufficient samples (N), the default fluctuation value configured by the system (0.05 in this embodiment) is used as a fallback. In the case of excessively large sample standard deviation, the fluctuation saturation upper limit (0.2 in this embodiment) is used as a cutoff value to avoid excessively wide dead zones caused by extreme variances, which would prevent the control law from responding.

[0091] Among them, the accuracy of stable users in the last 5 games is {0.80, 0.85, 0.82, 0.83, 0.81}, with β = 0.5, Substituting 0.05, the window mean is 0.822; the sum of squared biases is 0.00148; the sample variance is 0.00148 / 4 = 0.00037; the sample standard deviation is approximately 0.0192; the dead zone half-width is equal to the clamp's truncation of 0.05 + 0.5 × 0.0192 in the interval [0.05, 0.15], resulting in approximately 0.060. If the input accuracy is 0.85 (bias 0.075), then |e| = 0.075 > B = 0.060, triggering out-of-band, and the difficulty updates from 0.300 to 0.270; if the unhardened fixed dead zone B = 0.075 is used, then there is no in-band adjustment, demonstrating that hardening provides finer resolution for stable users (the dead zone is compressed from 0.075 to 0.060, equivalent to an approximately 1.25x increase in resolution). At a degenerate value of β = 0, the dead zone half-width degenerates to a fixed value. It is consistent with the basic version at the byte level.

[0092] In a preferred embodiment, a coupling term between tasks within the same cognitive domain—the mutual calibration capability level of tasks within the same domain—is superimposed on the basic control law to achieve strict isolation, with a cross-domain weight of 0. The formula is as follows:

[0093] ,

[0094] in Let λ be the set of tasks in the cognitive domain where task i is located, and λ ∈ [0, ..., ... ] represents the coupling coefficient. ≥0 represents the coupling weight and =1; when j does not belong to hour = 0, meaning cross-domain tasks are strictly isolated.

[0095] like Figure 6As shown, when co-domain coupling reinforcement is enabled in this embodiment, after calculating the basic step size in step S4 and before projecting the continuous interval, the co-domain coupling term calculation is performed first. The system calculates the co-domain coupling term based on the membership relationship of the cognitive domain in which the task is located. Figure 6 The task set (60), which is the attention cognitive domain, the memory cognitive domain, and the executive function cognitive domain, is used to find other task identifier sets within the cognitive domain to which the task belongs. In this embodiment, the cognitive domain membership is derived by mapping the first letter of the task identifier (for example, the first letter A is mapped to the attention cognitive domain, the first letter M is mapped to the memory cognitive domain, and the first letter C is mapped to the computational cognitive domain). The coupling term is calculated according to the above formula. In this embodiment, a specific value is λ = 0.03, which is a uniform co-domain coupling weight. = 1 / |siblings|; Cross-domain tasks are strictly isolated, i.e., the cross-domain weight is 0. The coupling item only participates in the difficulty update once when the task itself is completed, and does not update the difficulty of other tasks in the same domain synchronously to avoid event cascading.

[0096] The steady-state difficulty of task A1 in the attentional cognitive domain is 0.700, the initial difficulty of task A2 in the same cognitive domain is 0.300, and the initial difficulty of tasks A3 / A4 is 0.300; the siblings of A2 are {A1, A3, A4}, with a cardinality of 3; substituting λ = 0.03, the coupling terms... The base step size of the superposition assumption is 0.030. After the A2 update, the difficulty = 0.300 + 0.030 + 0.00400 = 0.33400, which is the lower bound of the byte-level matching theory. Cross-cognitive domain isolation verification: Under the steady-state A1 of 0.700, the coupling term for computing the cognitive domain task C1 is 0 (because C1's siblings do not include A1, the cross-domain weight is 0). Under the degenerate value of λ = 0, the coupling term is 0, consistent with the basic byte-level version.

[0097] Step S4: Write the task into the dual-track event table and update the status according to the EWMA recursive formula. The dual-track event table consists of an reach event table and a compliance event table. The two tables are not merged into a single completion status field.

[0098] like Figure 7 , Figure 8As shown, each record in the reach event table (named reach_events in this embodiment) contains at least a user identifier, a content unit identifier, and an output timestamp; the table uses (user identifier + associated identifier) ​​as a unique constraint. Each record in the conformity event table contains at least a user identifier, a content unit identifier, an execution timestamp, and an execution result; the table uses (user identifier + associated identifier + calendar day) as a unique constraint. When the system outputs a task / content unit to the training terminal, the system synchronously appends a reach record to the reach event table; when the user completes the content unit via the training terminal and returns the execution result, the system appends a conformity record to the conformity event table.

[0099] For each user-content unit pair, the system maintains a floating-point compliance rate. Whenever the pair triggers a "compliance event occurs within a preset time window" determination (corresponding to x = 1), the EWMA time-weighted compliance rate is calculated using the following recursive formula. The compliance rate of each user-content unit pair is recursively estimated using a first-order exponentially weighted moving average (EWMA), occupying only constant space and not requiring storage of the complete event history. The formula is as follows:

[0100] ,

[0101] in ∈ (0, 1) is the attenuation coefficient. ∈ {0, 1} indicates whether the corresponding compliance event occurs within the preset time window for the t-th touch event. EWMA estimates that each (u, c) pair only occupies O(1) space and does not require storing the complete event history.

[0102] In this embodiment, a specific value is set as the attenuation coefficient α = 0.7 and the time window is 1 day. Each user-content unit pair occupies only O(1) space (a single floating-point state + a single last update timestamp), and does not need to store the complete event history. The initial state compliance rate is 0, and after 10 consecutive compliance events, the theoretical parsing value is... 0.9717524751, the actual compliance rate of the system is 0.9718 (rounded to four decimal places), which is the theoretical value of byte-level matching.

[0103] In a more preferred embodiment, such as Figure 9 As shown, batch processing jobs are executed according to a preset scanning cycle (typically 1 day). For expired events that have exceeded the preset time window but still have no corresponding compliance event, an EWMA decay update with x = 0 is triggered.

[0104] The compliance rate of users who have stopped training automatically decays at a geometric rate—an EWMA update of x=0 is triggered once for each non-compliance outreach event during each scan cycle, as shown in the following formula:

[0105] ,

[0106] Persistent decay and deduplication logs ensure that each expired message triggers decay at most once (idempotency). The compliance rate of users who have stopped training converges geometrically, strictly guaranteeing time sensitivity.

[0107] When periodic decay hardening is enabled in this embodiment, the system runs a periodic batch job to trigger decay updates for expired, non-compliant events. In this embodiment, one specific value is once daily, executed at 00:05 the following day; the cron expression `0 5 0 * * ?` is an example engineering configuration, indicating that a periodic decay scan job is triggered daily at 00:05. This job processes the target date d*=today-1, i.e., yesterday; its purpose is to scan records that were reached but not compliant yesterday after the complete preset compliance judgment time window tau_window=1 day has passed, avoiding misjudgment of non-compliance before the window closes. The processing flow is as follows: query the set of (user ID + task ID) that were reached within the target date but have no corresponding compliance record within that date; for each pair in the set, read its EWMA status and last_decay_date; if it is equal to the target date, skip (ensuring idempotency); otherwise, trigger an EWMA decay update at x = 0 and write it to the decay deduplication log (see...). Figure 8 , Figure 9 Label (38)).

[0108] After a user stops training for k scan cycles, the compliance rate converges geometrically. For example, with α = 0.7, after stopping training for 5 scan cycles, the compliance rate converges from 0.9718 to 0.9718 × 0.7^5 ≈ 0.163. Even if a scan job restarts due to a single node failure, at most one decay will be performed on the same (user ID + task ID) pair on the same target date. Under the degradation value where the scan cycle approaches infinity (i.e., no scan job is scheduled), this scheme does not trigger any x = 0 updates, consistent with the basic version which only updates the EWMA at the byte level when a compliance event is triggered.

[0109] Step S5: Map the continuous difficulty of the control law output to discrete levels for display to the training terminal.

[0110] like Figure 10As shown, the control law operates in the continuous difficulty space d ∈ [0, 1]. The presentation layer maps the continuous difficulty to discrete gears L ∈ {1, ..., M} through positive vectorization, and uses reverse quantization to deduce continuous values ​​from the gears for initialization. The continuous difficulty output by the control law is mapped to user-visible discrete gears (typically 5 gears), ensuring correct integer boundaries for the gears, as shown in the following formula:

[0111] ,

[0112] The formula for deriving continuous difficulty from discrete gear levels, used in initialization or gear reversal scenarios, is as follows:

[0113] ,

[0114] Where d is the continuous difficulty parameter of the control law output, with a value range of [0,1]; M is the total number of discrete gears and M>=2; Q(d) is the positive vectorization function that maps the continuous difficulty parameter d to discrete gears L; L is the discrete gears of the display layer, with a value of {1,...,M}; This is the inverse quantization function that derives the continuous difficulty representative value from the discrete gear L; This represents the continuous difficulty value derived from gear L.

[0115] In this embodiment, the number of difficulty levels M = 5. The control law operates in continuous space, and quantization is only applied to the presentation layer. Byte-level verification of the 5 representative values: Q(0.00) = 1, Q(0.25) = 2, Q(0.50) = 3, Q(0.75) = 4, Q(1.00) = 5, all five levels match at the byte level. Example of intermediate value verification: Q(0.700) = round(4 × 0.700) + 1 = round(2.8) + 1 = 3 + 1 = 4. Reverse quantization is used for: regressing the initial continuous difficulty value from the initial difficulty level configured in the management terminal when the task is first enabled; and exporting representative values ​​of the difficulty levels for use by external data interfaces.

[0116] In a preferred embodiment, gear shift hysteresis stabilization is implemented. The currently displayed gear is maintained for the user. A two-way hysteresis zone is introduced near the gear boundary to suppress frequent gear jumps caused by small fluctuations in continuous difficulty. A hysteresis band is set near the gear boundary, and gear switching only occurs when the continuous difficulty exceeds the boundary plus the hysteresis threshold, avoiding visual jumps near the boundary. The formula is as follows:

[0117] ,

[0118] in Let i be the midpoint boundary of the continuous values ​​between adjacent gear positions i and i+1. > 0 represents the hysteresis threshold. The control law reads and writes continuously at difficulty d, and the hysteresis mechanism only affects the display direction and does not provide feedback to the control law.

[0119] like Figure 11 As shown, in this embodiment, the system maintains the user's current display level in addition to quantization mapping. And its corresponding prev_d (saves synchronously each time the control law updates the difficulty). Gear boundaries are calculated based on the midpoint of consecutive values ​​of adjacent gears: , i ∈ {1, 2, ..., M − 1}; the boundary values ​​are {0.125, 0.375, 0.625, 0.875} when M = 5. Hysteresis determination process: For continuous input difficulty d, first calculate the naive level without hysteresis. = Q(d); Query the prev_d field of this user-task pair. If no historical value is found, return the result directly. ;calculate = Q(prev_d) if = Return directly Otherwise, identify and Calculate the distance between the gear boundaries b. If Δ < hysteresis threshold, return. (Maintain within the hysteresis zone); otherwise return. (Switch).

[0120] The hysteresis threshold value is determined based on the following: In this embodiment, a specific value of 0.05 is used; this value is determined because it is less than the half-width of the gear position. = 0.125, ensuring that the hysteresis effective interval is in the local neighborhood near the gear boundary and is greater than the minimum step size of 0.03, ensuring that the hysteresis mechanism is not swallowed up by the single-step update of the control law.

[0121] The control law consistently reads and writes the continuous difficulty level d. The hysteresis mechanism only affects the display direction and does not provide feedback to the control law. Even if the display level remains unchanged due to hysteresis, the underlying continuous difficulty level continues to update according to the control law. Once the continuous difficulty level crosses the hysteresis zone, the display level switches to a new naive level. Under the degenerate value where the hysteresis threshold is 0, the distance of any level change is greater than 0, and the hysteresis condition is never met, consistent with the byte-level quantization of the basic version.

[0122] Step S6: Select at least one content unit according to a preset selection strategy and output it to the training terminal. If the content unit constitutes a new touch event, write it back to the touch event table as a state variable for the next round of feedback control closed loop. See also Figure 3Data flow closed loop and Figure 9 The path-cycle decay scanning process is not followed. At least one content unit is selected from the list of retrieved candidate content units according to the preset selection strategy and output to the training terminal via the network interface (40). The output constitutes a new touch event and is written back to the touch event table as the state variable of the next round of feedback control closed loop.

[0123] Typical values ​​and effective ranges of parameters in this embodiment are as follows: Target flow T: typical 0.775, effective range [0.70, 0.85]; Lower limit of dynamic dead zone. Typical value: 0.05; Effective range: [0.03, 0.06]; Fluctuation sensitivity coefficient Typical value 0.5, effective range [0.3, 0.8]; Scrolling window sample size N: Typical value 5, effective range [3, 10]; Linear step size coefficient k: Typical value 0.15, effective range [0.10, 0.25]; Step size lower limit Typical value: 0.03; Effective range: [0.02, 0.05]; Maximum step size. Typical value: 0.08, effective range: [0.06, 0.12]; EWMA attenuation coefficient α: typical value: 0.7, effective range: [0.5, 0.9]; compliance decision time window: typical value: 1 day, effective range: [1 hour, 7 days]; periodic attenuation scan period: typical value: 1 day, must not exceed the compliance decision time window; co-domain coupling coefficient λ: typical value: 0.03, effective range: [0, 0.1]; number of gears M: typical value: 5, must not be less than 2; gear hysteresis threshold. Typical value is 0.05, effective range is [0.03, 0.08]; the default starting point for difficulty is 0.3, effective range is [0.1, 0.5].

[0124] All four hardening implementation schemes described above underwent degradation equivalence verification: the state-dependent dead zone degenerates to a fixed dead zone basic control law when β = 0; co-domain coupling degenerates to independent single-input single-output control when λ = 0; period decay degenerates to event-triggered EWMA (no x = 0 decay) when the scan period tends to infinity; and gear hysteresis degenerates to pure quantization output when the hysteresis threshold is 0. Under the default configuration (all hardening items disabled), the system behavior is consistent with the base version before hardening at the byte level, with no regression.

[0125] Another embodiment of the present invention is a cognitive training task and auxiliary content adaptive distribution system, including an association index management module, a receiving module, a dead-zone control module, a compliance rate EWMA module, a quantization mapping module, and an output module, wherein:

[0126] The association index management module is used to build association indexes for cross-modal constraints for cognitive training tasks;

[0127] The receiving module is used to receive cognitive training task completion events and perform candidate retrieval of the corresponding content unit identifier list;

[0128] The dead-zone control module is used to update the continuous difficulty parameters of the current task using a dead-zone proportional control law;

[0129] The compliance rate EWMA module is used to write tasks into the dual-track event table and update the status according to the EWMA recursive formula;

[0130] The quantization mapping module is used to map the continuous difficulty of the control law output into discrete levels for display to the training terminal;

[0131] The output module is used to select at least one content unit according to a preset selection strategy and output it to the training terminal.

[0132] In this embodiment, the cognitive training task and auxiliary content adaptive distribution system backend service is implemented as a software component executed by a processor and memory, with a relational database as persistent storage. The deployment scale is 18 cognitive training tasks, 5 cognitive domains, 30 content units, and 41 associated index entries. The system has completed end-to-end unit testing and integration testing for coverage control law, compliance rate estimation, periodic decay scanning, and quantization and hysteresis switching (all 105 tests passed, with no regression).

[0133] like Figure 12 As shown, the system of the present invention includes at least one processor (20), a memory (30), and a network interface (40), which together constitute the system as a whole (10). During the deployment phase, the processor (20) in the system as a whole (10) establishes an associated index P in the memory (30). During system operation, the training terminal (50) submits a task completion event to the processor (20) via the network interface (40) at the end of each task session. At least one processor calls the dead-zone proportional control module (80) to update the continuous difficulty parameter of the current task according to the dead-zone proportional control law. The system maps the continuous difficulty output by the control law to discrete levels to display to the training terminal (50). At least one content unit is selected from the list of retrieved candidate content units according to the preset selection strategy and output to the training terminal (50) via the network interface (40). For each user-content unit pair, the system maintains a floating-point state compliance rate. Whenever the pair triggers a judgment that "compliance event occurs within a preset time window" (corresponding to x = 1), the state is updated according to the EWMA recursive formula. If the output constitutes a new event, it is written back to the event table and used as the state variable for the next round of feedback control closed loop.

[0134] This embodiment can be deployed on a single machine. The structure includes at least one processor, memory, and network interface, forming a single server node. The processor is a general-purpose CPU (64-bit x86-64 multi-core processor), the memory is a combination of persistent storage and running memory (a combination of SSD hard disk + DDR4 RAM, the persistent layer uses a relational database or key-value store to carry the associated index P, the reach event table, the compliance event table, the decay deduplication log, the difficulty status table, and the EWMA status table), and the network interface is a network transceiver that supports HTTP / HTTPS protocols, communicating with training terminals, management terminals, and related party terminals via the Internet or internal network.

[0135] This embodiment can also be a cluster deployment, where at least one processor is a collection of distributed computing nodes, specifically a deployment of multiple stateless application nodes plus a shared database, scheduled by a load balancer via a network interface; the storage is a persistent storage cluster (relational database master-slave, key-value storage cluster, object storage, etc.); at least one processor performs control law calculation, EWMA recursive update, periodic decay scan job, etc., distributed on different nodes, wherein the periodic decay scan job is guaranteed to be executed at most once for the same target date by a distributed lock in the distributed deployment.

[0136] like Figure 13 The network interface (40) communicates with the training terminal (50), management terminal (52), and related party terminal (54) through a predetermined application layer interface (RESTful HTTP interface or gRPC interface). The network interface (40) contains a semantic field whitelist filter (42) to perform semantic filtering on the returned fields of the related party terminal (positive fields are retained: cumulative number of completions, positive compliance rate, stage participation identifier, cumulative number of contacts, overall progress percentage, etc.; negative fields are excluded: error rate, list of unfinished tasks, original difficulty value, negative compliance rate, training interruption record, etc.).

[0137] The end-to-end latency of the algorithm layer in this system is P95 = 5.571 milliseconds under 50 concurrent users (measured by JMH micro-benchmark); in a synthetic production environment (algorithm layer + 2 MySQL queries), P95 ranges from 6.57 to 9.57 milliseconds; and in a single-user scenario, P95 = 75.78 microseconds. The latency is O(1) related to the user scale and the historical event scale (only a constant 2 queries to the memory are required to complete each event processing, which is independent of the number of historical events).

[0138] In terms of the distribution layer, the association index P is established before deployment by cross-modal coverage constraint verification, and candidates are located in O(1) time complexity at runtime; the cold start of new users does not rely on feedback data (the first-round matching rate of collaborative filtering methods is typically 15-35%, but this limitation does not apply to this invention). The cross-modal integrity constraint ensures that each association covers at least two modalities, eliminating the loss of comprehensive support value of single-modal associations.

[0139] Under steady-state conditions with zero mean deviation and bounded variance, this invention employs a dead-zone proportional control law to ensure that continuous difficulty parameters tend to and remain within the difficulty zone corresponding to the flow interval. Monte Carlo simulations, based on 5 independent random seeds and 30 simulated user samples for each seed, verify that this invention improves the flow zone convergence rate by approximately 43.4% compared to ML decision-making methods, with a robustness range of 27%-58% across the 5 random seeds.

[0140] The state-related dead zone is adaptively adjusted based on the historical volatility of the user-task relationship. Stable users enjoy fine-tuning, while volatile users experience oscillation suppression. Same-domain coupling terms enable cross-task capability transfer within the same cognitive domain. In a real system, the measured same-domain coupling coefficient λ = 0.03, and the cross-domain coupling weight is 0, ensuring strict isolation (measured coverage of 18 tasks × 5 domains).

[0141] In terms of the feedback layer, this invention maintains reach and compliance events in independent data tables, and improves the feedback signal resolution from a single Boolean bit to a continuous EWMA value (3-4 bits of valid information). Each (u, c) pair of EWMA incremental updates occupies only O(1) space and does not increase linearly with the number of events. The measured EWMA attenuation coefficient α in the real system is 0.7 (steady-state value after 10 consecutive full compliance events). (Byte-level matching theoretical value). The periodic decay of non-compliance paths ensures that the compliance rate of users who have stopped training increases geometrically. Convergence is achieved with strict time sensitivity; decaying deduplication logs ensure idempotency, and repeated scans do not result in repeated decay.

[0142] In terms of the interface layer, this invention decouples continuous control and discrete display through bidirectional homomorphic mapping using positive and negative quantization. The control law operates in continuous space to ensure accuracy, while the discretization of the display layer avoids cognitive burden. The system's measured number of gears is M = 5, and the gear hysteresis threshold is... = 0.05. JMH micro-benchmark test results (Apple M2 Pro / JDK 20 / JMH 1.37, 1 / 5 / 50 concurrent threads × over 100,000 samples per method, total sample size over 280 million data points): End-to-end closed-loop P95 latency is 75.78 microseconds in a single-user scenario, 280.58 microseconds for 5 concurrent users, and 5.571 milliseconds for 50 concurrent users (algorithm layer, Mockito isolates database I / O); in a synthetic production environment (algorithm layer + 2 MySQL primary key / index queries, 1-4 milliseconds), P95 is 6.57 to 9.57 milliseconds, far below the target upper limit of 10 milliseconds. After introducing state-related dead zones, co-domain coupling, and gear hysteresis, the hardened control law only increases absolute overhead by 21.7 microseconds compared to the basic version.

[0143] Regarding the universality of parameters across tasks, the system's actual measured single parameter set (T = 0.775) shows... = 0.05, k = 0.15, = 0.03, = 0.08, α = 0.7, M = 5, β = 0.5, λ = 0.03, = 0.05) can achieve flow zone convergence in all 18 cognitive training tasks × 5 cognitive domains, proving the universality of parameter selection for task types.

[0144] Another embodiment of the present invention is a computer-readable storage medium, which includes, but is not limited to: magnetic storage media (specifically, in this embodiment, a mechanical hard disk or magnetic tape), optical storage media (specifically, a CD-ROM, DVD-ROM, or Blu-ray), semiconductor storage media (specifically, a ROM, PROM, EPROM, EEPROM, flash memory, or solid-state drive), and any combination thereof; a computer program is stored on the medium. When the computer program is executed by at least one processor, the at least one processor performs all the steps of the aforementioned method, specifically including establishing an association index P of a hash structure and performing cross-modal integrity pre-verification on it, receiving task completion events and retrieving candidates through hash addressing, executing a dead-zone proportional control law to update the continuous difficulty, independently maintaining an access event table and a compliance event table and performing EWMA recursive updates, performing continuous-discrete quantization mapping and providing reverse quantization, selecting content units according to a preset strategy and outputting them to a training terminal via a network interface to form a feedback control closed loop. The specific form of a computer program may be compiled bytecode or natively executable binary, or interpreted script, or runtime packaged in the form of a container image, or any combination thereof; this invention does not limit the specific form of the program.

[0145] 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 method for adaptive distribution of cognitive training tasks and auxiliary content, characterized in that, Includes the following steps: Step S1: Establish a cross-modal constraint association index for the cognitive training task; Step S2: Receive the cognitive training task completion event and perform a candidate retrieval of the corresponding content unit identifier list; Step S3: Update the current task's continuous difficulty parameters using a dead-zone proportional control law; Step S4: Write the task into the dual-track event table and update the status according to the EWMA recursive formula. The dual-track event table consists of an reach event table and a compliance event table. The two tables are not merged into a single completion status field. Step S5: Map the continuous difficulty of the control law output to discrete levels for display to the training terminal; Step S6: Select at least one content unit according to the preset selection strategy and output it to the training terminal.

2. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 1, characterized in that, In step S1, the associated index adopts a hash access structure, with the key being the task identifier and the value being a list of content unit identifiers. The hash access structure consists of two associated tables in a relational database. The hash access structure is implemented, and a hash cache with the task identifier as the key is established in memory to support runtime O(1) retrieval. The two associated tables are the task table and the task content unit associated table.

3. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 1, characterized in that, In step S2, the completion event carries at least a user identifier, a task identifier, an accuracy rate and a completion degree indicator, and at least one processor normalizes the accuracy rate; if the accuracy rate is empty, the completion event does not trigger a control law update and returns the current difficulty parameter unchanged in order to avoid invalid samples polluting the control state; the candidate retrieval uses the task identifier as the key to perform hash addressing in the associated index and retrieves the corresponding content unit identifier list in O(1) time complexity as a candidate for subsequent selection output.

4. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 1, characterized in that, In step S3, the proportional control law with dead zone takes the accuracy deviation e as input. No adjustment is made when the deviation falls within the dead zone half-width B. When the deviation exceeds the dead zone, the adjustment amount is calculated according to the linear step size coefficient k and output after step size upper and lower limit saturation and continuous interval projection. The formula is expressed as follows: ， ， Where T∈(0,1) is the flow target, B>0 is the dead zone half-width, and k>0 is the linear step size coefficient. ≤ Here, represents the upper and lower limits of the step size, Proj represents the projection over a continuous interval, sat is the saturation operator, and sgn is the sign function. Let [i] be the continuous difficulty parameter for task i before the t-th update, with a value range of [0,1]. Let be the normalized accuracy corresponding to the completion event of task i in the t-th task, with a value in the range [0,1]. The input is a feedback control function with dead zone, which is used to output the amount of continuous difficulty adjustment for this operation. This is the error signal.

5. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 4, characterized in that, Dead zone half-width B is a user-task history variability function. Calculate using the following formula: ， ， in > 0 is the lower limit of the dead zone. > 0 represents the fluctuation sensitivity coefficient, [ , ] represents the feasible interval of the dead zone, σ is the rolling sample standard deviation of the accuracy of the most recent N games, and N is the length of the rolling window. This represents the mean of the normalized accuracy within the scrolling window. The rolling sample standard deviation of the normalized accuracy of user u in task c over the most recent N rounds is calculated based on an N-1 correction.

6. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 4, characterized in that, By superimposing a co-domain coupling term on top of the base step size, mutual calibration of task capabilities within the same cognitive domain and strict isolation across domains are achieved. The formula is as follows: ， in Let λ be the set of tasks in the cognitive domain where task i is located, and λ ∈ [0, ..., ... ] represents the coupling coefficient. ≥ 0 represents the coupling weight and =1; when j does not belong to hour = 0, strictly isolates cross-domain tasks.

7. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 1, characterized in that, In step S4, the reach event table records the system output time, and the compliance event table records the user execution receipt; for each user-content unit pair (u, c), the EWMA time-weighted compliance rate is calculated using the following recursive formula: ， in ∈ (0, 1) is the attenuation coefficient. ∈ {0, 1} indicates whether the corresponding compliance event occurs within the preset time window for the t-th touch event.

8. The adaptive distribution method for cognitive training tasks and auxiliary content according to claim 7, characterized in that, In step S4, a batch processing job is executed according to a preset scanning cycle. For expired events that have exceeded the preset time window but still have no corresponding compliance event, an EWMA decay update with x = 0 is triggered.

9. A cognitive training task and auxiliary content adaptive distribution system, characterized in that, It includes a relational index management module, a receiving module, a dead-zone control module, a compliance rate EWMA module, a quantization mapping module, and an output module, among which: The association index management module is used to build association indexes for cross-modal constraints for cognitive training tasks; The receiving module is used to receive cognitive training task completion events and perform candidate retrieval of the corresponding content unit identifier list; The dead-zone control module is used to update the continuous difficulty parameters of the current task using a dead-zone proportional control law; The compliance rate EWMA module is used to write tasks into a dual-track event table and update the status according to the EWMA recursive formula. The dual-track event table consists of an reach event table and a compliance event table. The two tables are not merged into a single completion status field. The quantization mapping module is used to map the continuous difficulty of the control law output into discrete levels for display to the training terminal; The output module is used to select at least one content unit according to a preset selection strategy and output it to the training terminal.

10. An electronic device, characterized in that, The electronic device includes: One or more processors; and A memory storing computer program instructions, which, when executed, cause the processor to perform the steps of the method as described in any one of claims 1 to 8.

Images