Human-machine collaborative causal revision method for adversarial decision

Abstract

The application relates to the field of complex system performance evaluation and intelligent game, and particularly discloses a man-machine cooperative cause-effect correction method for antagonistic decision, which comprises the following steps: generating a cause-effect draft atlas with a cause-effect direction, a time delay range and a confidence through data preprocessing; revising the cause-effect atlas by experts, and performing consistency checking and conflict solving on the revised cause-effect atlas; depositing expert knowledge into computable structure prior and discovery template; realizing atlas incremental updating based on a sliding window, triggering an alarm and a relearning process when detecting that data distribution or a cause-effect structure drifts, and finally storing a checked cause-effect atlas version to serve an upper game situation prediction and strategy deduction system. The method can continuously maintain a time-delay cause-effect atlas which is interpretable, traceable and self-adaptive.

Description

Technical Field

[0001] This invention relates to the field of performance evaluation of complex systems and intelligent game theory, and specifically provides a human-machine collaborative causal correction method for adversarial decision-making. Background Technology

[0002] The effectiveness evaluation of adversarial decision-making is a core component of modern intelligent decision-making systems. Its core task is to accurately identify "which factors act on key performance indicators in what order and with what time delay," and to use the analysis results for game situation prediction, adversarial strategy deduction, and system diagnosis. This process highly relies on multimodal interactive data from environmental detectors, optical sensors, cooperative communication networks, and decision control systems. These data are characterized by inconsistent sampling frequencies, clock inconsistencies, and significant impact from adversarial interference, resulting in asynchronous and significant time delays in the effectiveness of causal information channels.

[0003] Currently, various technical approaches exist in the field of game theory performance analysis. Among them, prediction methods based on correlation analysis or black-box models focus on the statistical associations between variables, but are weak in revealing causal mechanisms and have extremely coarse expressions of time delays; methods based on dynamic graphs or graph networks improve model accuracy, but have poor interpretability, and their causal structures are prone to drift in different game environments, resulting in insufficient stability; traditional causal discovery algorithms often assume that variables have contemporaneous causal relationships or simple fixed lags, making it difficult to incorporate the prior knowledge of domain experts and the hard constraints at the physical and game strategy levels; while knowledge engineering-based methods are interpretable, their construction and maintenance costs are high, updates are slow, and they are difficult to adapt to the fast-paced changes of highly dynamic adversarial environments.

[0004] In summary, existing technologies suffer from attribution bias due to neglecting causal delays, and the constructed models exhibit poor cross-environment stability, lack online updates and expert knowledge injection capabilities, and also lack logical consistency verification and version management of the graphs, resulting in insufficient credibility of their analysis results.

[0005] Therefore, proposing a causal correction method that can form a closed loop of temporal causal discovery, edge-level human-machine collaborative correction, multi-constraint consistency verification, and online incremental update has become an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a human-machine collaborative causal correction method for adversarial decision-making, so as to solve the problems existing in the prior art.

[0007] To achieve the above objectives, this invention provides a human-machine collaborative causal correction method for adversarial decision-making, comprising:

[0008] S1: Establish a unified time reference for multi-source data, configure the sliding window, step size, and time delay search range for causal analysis, and configure a clock synchronization scheme;

[0009] S2: Access multimodal simulation data from sensing units, collaborative communication networks, and decision control systems; perform data cleaning, frequency unification, and clock synchronization; and generate analysis samples with environmental labels using a sliding window.

[0010] S3: Based on rules and prior constraints, the causal discovery algorithm is used to process the samples to generate a preliminary causal map with causal direction, time delay range and confidence level;

[0011] S4: Receive expert revisions to the initial causal diagram;

[0012] S5: Perform a consistency check on the revised causal graph to check whether it satisfies the constraints of directed acyclicity, time realizability, and physical / game strategy. If a conflict is found, generate and apply minimal modifications to restore consistency.

[0013] S6: Transform the experts' revisions and rationale into computable structural priors and discovery templates;

[0014] S7: Based on the sliding window, incremental causal discovery is performed on the continuously input simulation data. When a shift in data distribution or causal structure is detected, an alarm and relearning process is triggered.

[0015] Preferably, in S3, the causal discovery algorithm adopts a hybrid framework based on constraints and scores, including: determining the undirected causal skeleton and the corresponding time delay based on time-series constraints; and determining the causal direction of the undirected causal skeleton based on scores.

[0016] Further optimization, in S3, the undirected causal skeleton and its corresponding time delay are determined based on time constraints as follows: spurious associations are eliminated through conditional independence tests, and statistical judgments are made to preliminarily determine the undirected causal skeleton and its corresponding time delay.

[0017] Further optimization, in S3, determining the causal direction of the undirected causal skeleton based on the score specifically includes: introducing the Bayesian information criterion as the scoring function; using a greedy equivalence search algorithm, with the score as the optimization objective, performing edge addition, deletion, and reversal operations on the undirected causal skeleton until the score no longer increases, thereby obtaining the causal direction of the undirected causal skeleton.

[0018] In a further preferred embodiment, in S4, the revision includes determining, reversing, deleting, and limiting the time window.

[0019] Further optimization, in S4, there is also an operation to recommend high uncertainty causal edges to experts based on historical templates, so as to prompt experts to review them first. The method for determining high uncertainty causal edges is as follows: calculate the confidence variance of each causal edge or the posterior probability entropy of the causal edge direction, and determine the edge with the highest entropy value or the largest variance as the highest uncertainty causal edge.

[0020] Further optimization, in S5, a minimum modification set is generated by calculating the graph edit distance, where the graph edit distance is defined as the sum of the costs of the minimum number of editing operations required to transform the current graph into a target graph that satisfies all constraints; a search algorithm is used to find the modification sequence that satisfies all consistency constraints and has the minimum graph edit distance from the minimum modification set.

[0021] Further optimization reveals the following incremental causal discovery process: When a new time window of data arrives, the new window data is compared with the data from the previous several windows; for variables and their associated edges that have undergone significant statistical changes, based on the previous version of the causal graph, the causal discovery algorithm is re-run only on these highly variable local subgraphs to perform conditional independence testing and score optimization, in order to update the causal structure of the local region; finally, the updated local subgraphs are merged with the global graph, and the logical correctness of the integrated graph is ensured through global consistency verification.

[0022] Further optimization involves S7, where data distribution drift is monitored by calculating the KS statistic of data distribution between adjacent windows; and causal structure drift is monitored by calculating the Jaccard index of structural similarity between causal graphs of adjacent windows.

[0023] Furthermore, the human-machine collaborative causal correction method for adversarial decision-making further includes the following steps:

[0024] S8: The causal graph that has passed the consistency check will be stored in a versioned manner, the audit log will be recorded, and the causal graph data will be generated.

[0025] The beneficial effects of the human-machine collaborative causal correction method for adversarial decision-making provided by this invention are as follows:

[0026] (1) Through the causal expression of "direction + time delay range + confidence level", an interpretable and quantifiable time delay causal map is formed, which can clearly answer the game strategy questions of "when it takes effect, how fast it is, and how reliable it is".

[0027] (2) A closed loop of human-machine collaboration at the edge level was introduced, which structurated and precipitated expert knowledge into priors and templates, significantly reducing false edges and misjudgments of direction in causal discovery, and achieving continuous optimization that becomes more accurate with use;

[0028] (3) Before release, multi-dimensional consistency verification is performed and conflicts are automatically repaired with minimal modifications. Combined with version auditing, the logical rationality of the causal graph and the reliable traceability of the evolution process are ensured.

[0029] (4) By using incremental discovery and drift detection based on sliding window, the online correction and adaptive update of the causal graph are realized, enabling it to continuously evolve with the evolution of combat data and environment, effectively improving the cross-environment stability and long-term practical value of the model. Attached Figure Description

[0030] Figure 1 A flowchart of a human-machine collaborative causal correction method for adversarial decision-making provided in an embodiment of the present invention;

[0031] Figure 2 Flowchart for generating the initial causal graph;

[0032] Figure 3 Here is a flowchart of the causal discovery algorithm;

[0033] Figure 4 Flowchart for consistency verification and conflict resolution;

[0034] Figure 5 This is a schematic diagram of incremental updates and drift detection;

[0035] Figure 6 This is an architecture diagram for versioned storage and audit logs. Detailed Implementation

[0036] The present invention will be further explained below with reference to specific embodiments, but it is not limited to the present invention.

[0037] To address the problems of attribution bias caused by ignoring time delays, poor cross-environment stability due to static and rigid models, and insufficient model interpretability due to the lack of effective human-computer interaction mechanisms, such as... Figure 1 As shown, this invention provides a human-machine collaborative causal correction method for adversarial decision-making, comprising the following steps:

[0038] S1: Establish a unified time reference for multi-source data, configure the sliding window, step size, and time delay search range for causal analysis, and configure a clock synchronization scheme;

[0039] This step is the data foundation stage of the entire method, and its core task is to solve the problems of "clock synchronization" and "analysis scale" of multi-source data.

[0040] The time base is used to uniformly correct the timestamps of all access data to the same time base, eliminating timing errors caused by independent clocks of devices.

[0041] The sliding window defines the duration of data examined in a single analysis, the step size determines the interval of window sliding, thus affecting the analysis density, and the time delay search range limits the maximum and minimum boundaries of the time difference between the cause and the result explored by the algorithm, providing key constraints for subsequent causal discovery;

[0042] S2: Access multimodal simulation data from sensing units, collaborative communication networks, and decision control systems; perform data cleaning, frequency unification, and clock synchronization; and generate analysis samples with environmental labels using a sliding window.

[0043] This step is the core of data processing and sample preparation. After the system receives real-time data streams from different sources, it first performs data cleaning to repair or remove outliers and missing segments to ensure data quality. Then, it uses resampling technology to unify the frequency of all data to the same standard and completes precise clock synchronization so that all data points are aligned to a unified time axis. Finally, according to the sliding window and step size set in step S1, the continuous data stream is cut into continuous and equal-length data segments, and loop labels are attached to each segment to generate a standardized analysis sample set that can be directly analyzed by the algorithm.

[0044] S3: Based on rules and prior constraints, the sample is processed using a causal discovery algorithm to generate a preliminary causal map with causal direction, time delay range and confidence level, which serves as the causal basis for the performance analysis of the dynamic game system.

[0045] This step employs a hybrid causal discovery algorithm based on constraints and scores. Its core objective is to automatically generate a preliminary causal graph with direction, time delay, and confidence level based on time-series data. The overall process for generating the preliminary causal graph is as follows: Figure 2 As shown, its core is divided into two key stages:

[0046] Phase 1: Determine the undirected causal framework and corresponding time delay based on time sequence constraints;

[0047] This step aims to discover potential causal relationships between variables and estimate time delays, using conditional independence tests to eliminate spurious associations. The test is based on the following statistics:

[0048] ;

[0049] in, It is a past value of variable X. With the current value of variable Y Given a set of conditions Partial correlation coefficient (including other variables and historical values) The sample size; the p-value obtained from this test allows for statistical judgment and preliminary determination of the undirected causal framework and its corresponding time delay τ; the significance level threshold for the conditional independence test. Set to 0.05; if the p-value is greater than... Then remove the edges between variables; time delay search range The process iterates through data sampling intervals in increments.

[0050] Phase 2: Determine the causal direction of the undirected causal skeleton based on fractions to achieve graph structure optimization;

[0051] This stage refines the undirected causal framework to determine its final causal direction; specifically, it introduces the Bayesian information criterion as a scoring function for optimization.

[0052] ;

[0053] The core advantage of this function lies in its trade-off capability: the first term on the right-hand side of the equation measures the goodness of fit of the causal graph to the data, while the second term penalizes the complexity of the model (k is the total number of parameters). By optimizing this BIC score, the algorithm can automatically select a causal graph that fits the data well and has a simple structure, effectively preventing overfitting and ensuring the robustness of the results. This score-based optimization method uses a greedy equivalence search algorithm, with the BIC score as the optimization objective, to perform edge addition, deletion, and reversal operations on the undirected causal skeleton until the BIC score no longer increases.

[0054] Ultimately, through such Figure 3 The complete causal discovery algorithm flow shown generates a preliminary causal graph with causal direction, time delay range, and confidence level, providing a high-quality, interpretable, and automated foundation for subsequent human-machine collaborative correction in S4.

[0055] S4: Provide a human-computer interaction interface to receive revisions from experts on the initial causal graph, wherein the revisions include confirmation, reversal, deletion, and limiting the time window;

[0056] In this step, experts can receive revisions through a visual human-computer interaction interface designed specifically for causal analysis. This interface can be as follows: The central main area clearly presents the initial causal graph in a graph structure, where nodes represent key state variables, directed edges represent causal relationships, and different colors or line types are used to visually indicate the causal direction, time delay range, and algorithm confidence. An operation panel is located on one side of the interface, allowing experts to select individual causal edges in the graph and perform revision operations such as "confirm," "reverse," "delete," or "limit time window." Furthermore, this interface can integrate an intelligent recommendation function based on historical templates, highlighting causal edges with high uncertainty and prompting experts to review them first, thereby significantly improving the efficiency and accuracy of human-computer collaboration. The method for determining high-uncertainty causal edges is as follows: calculate the confidence variance or the posterior probability entropy of the causal edge direction for each causal edge, and determine the edge with the highest entropy value or the largest variance as the highest uncertainty causal edge.

[0057] S5: Perform a consistency check on the revised causal graph to check whether it satisfies the constraints of directed acyclicity, time realizability, and physical / game strategy. If a conflict is found, generate and apply minimal modifications to restore consistency.

[0058] The specific procedures for consistency verification and conflict resolution are as follows: Figure 4 As shown; the consistency check includes checks in three dimensions: Directed acyclicity is verified using a topological sorting algorithm to ensure that there are no directed cycles in the graph; Time realizability is checked by traversing all causal edges and verifying whether the timestamp of the cause node plus the delay is less than or equal to the timestamp of the result node; Physical / game theory constraints are checked by querying a predefined declarative rule base (e.g., rules encoded using first-order logic statements). For example, the rule base contains the following first-order logic rules: The rule states that when a critical strike action (such as the release of high-energy effects or braking interception) is the cause, its effect cannot be instantaneous (immediate-delayed). (Must be greater than 0), which excludes physically impossible causal relationships; when a conflict is detected, the system generates the minimum set of modifications by calculating the graph edit distance: the graph edit distance is defined as the minimum set of modifications that the current graph must modify. Transform into a target graph that satisfies all constraints The total cost of the minimum number of edit operations required (including adding, deleting, and reversing edges), i.e. ,in, The cost of each operation; minimized by constructing a constraint-satisfying problem model. With the objective of satisfying all consistency constraints, a search algorithm (such as A* or greedy search) is used to find the sequence of specific modification operations with the minimum cost and automatically applied to restore the consistency of the graph; specifically, the system models conflict resolution as an optimization problem: in the graph Find a modified diagram from all possible modifications. , making Satisfies all consistency constraints, and the graph edit distance Minimum; where the graph edit distance is defined as:

[0059] ;

[0060] in, For the first The system can use a breadth-first search algorithm with pruning to find the optimal modification sequence for the cost of each edit operation (add, delete, reverse edge).

[0061] S6: Transform the experts' revisions and rationale into computable structural priors and discovery templates;

[0062] This step is crucial for the system's continuous self-optimization. The system records all revisions made by experts in S4 and the reasons behind them, and "translates" this human knowledge into a machine-understandable form. For example, frequently deleted causal edges will be transformed into "prohibited connections" structural priors; expert-preferred causal patterns can be abstracted into "discovery templates." This accumulated knowledge will be prioritized in subsequent automated analysis, thereby reducing similar errors and improving discovery efficiency and accuracy.

[0063] S7: Based on the sliding window, incremental causal discovery is performed on the continuously input simulation data. When a shift in data distribution or causal structure is detected, an alarm and relearning process is triggered.

[0064] This step enables the method to adapt to dynamically changing environments, and its core mechanism is illustrated in the diagram below. Figure 5As shown, the system uses a sliding window as the basic unit to perform incremental causal discovery on continuously input new data. Its core process is as follows: When a new time window of data arrives, the system does not recalculate the entire causal graph from scratch. Instead, it first compares the new window data with the data from the previous few windows. Next, the system focuses on variables and their associated edges that have undergone significant statistical changes. Based on the previous version of the causal graph, it re-runs the causal discovery algorithm only on these highly variable local subgraphs, performing conditional independence tests and score optimization to update the causal structure of that local region. Finally, the updated local subgraphs are merged with the global graph, and the logical correctness of the integrated graph is ensured through a global consistency check (as described in S5). Simultaneously, the system continuously monitors the overall changes in data distribution and existing association structures. Once a significant "causal drift" is detected by calculating the structural similarity between adjacent window maps or abrupt changes in the model fit residuals, it indicates that the underlying mechanism may have changed. The underlying mechanism refers to the fundamental principles or dynamic processes that support the causal relationships between data; it is the law that determines how data influence and interact with each other, such as the impact of game strategy decisions on hit rate, or the impact of environmental factors on device performance. These mechanisms adjust as the external environment or data changes. If such a change in mechanism is detected, the system will immediately trigger an alarm and initiate the model relearning process.

[0065] The data distribution drift monitoring method is as follows: data distribution drift is monitored by calculating the Kolmogorov-Smirnov (KS) statistic of the data distribution between adjacent windows; for any variable Its KS statistic ,in, and Let be the cumulative distribution function of adjacent windows; if If the threshold is exceeded, a significant drift is determined to have occurred.

[0066] The method for monitoring the stability of causal structures is as follows: The stability of causal structures is monitored by calculating the Jaccard index, which represents the structural similarity between causal graphs of adjacent windows; the Jaccard index is defined as:

[0067] ;

[0068] in, and Let be the set of causal edges between adjacent windows, if If the value is below a preset threshold, it is determined that the causal structure has changed significantly.

[0069] S8: The causal graphs that have passed consistency verification will be versioned and stored, audit logs will be recorded, and causal graph data will be generated that can be used by the upper-level situation prediction and strategy simulation system.

[0070] This step focuses on the management and output of results, and its storage and auditing architecture is as follows: Figure 6 As shown, all final causal graphs that pass the S5 consistency check will be assigned a version number and processed according to... Figure 6 The system stores data in a versioned storage structure and records complete audit logs to ensure traceability. The audit logs record the operator, timestamp, operation content, reason, and scope of impact, and support graph rollback based on version records. Finally, the system converts these causal graphs into standardized formats (such as JSON and XML) and outputs them to upper-level systems such as situation prediction and game strategy deduction, providing them with the core logical reasoning basis.

[0071] The method provided by this invention realizes a human-computer collaborative closed loop for causal discovery by constructing a complete closed loop that includes data preparation, human-computer interaction, consistency verification, knowledge accumulation, online updates and version management. It can continuously produce high-quality, interpretable causal graphs and provide reliable support for decision analysis.

[0072] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A human-machine collaborative causal correction method for adversarial decision-making, characterized in that, include: S1: Establish a unified time reference for multi-source data, configure the sliding window, step size, and time delay search range for causal analysis, and configure a clock synchronization scheme; S2: Input multimodal simulation data, perform data cleaning, frequency unification and clock synchronization, and generate analysis samples with environmental labels by sliding window; S3: Based on rules and prior constraints, the causal discovery algorithm is used to process the samples to generate a preliminary causal map with causal direction, time delay range and confidence level; S4: Receive expert revisions to the initial causal diagram; S5: Perform a consistency check on the revised causal graph to check whether it satisfies the constraints of directed acyclicity, time realizability, and physical / game strategy. If a conflict is found, generate and apply minimal modifications to restore consistency. S6: Transform the experts' revisions and rationale into computable structural priors and discovery templates; S7: Based on the sliding window, incremental causal discovery is performed on the continuously input simulation data. When a shift in data distribution or causal structure is detected, an alarm and relearning process is triggered.

2. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: In S3, the causal discovery algorithm adopts a hybrid framework based on constraints and scores, including: determining the undirected causal skeleton and the corresponding time delay based on time-series constraints; and determining the causal direction of the undirected causal skeleton based on scores.

3. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 2, characterized in that: In S3, the undirected causal skeleton and its corresponding time delay are determined based on time constraints as follows: spurious associations are eliminated through conditional independence tests, and statistical judgments are made to preliminarily determine the undirected causal skeleton and its corresponding time delay.

4. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 2, characterized in that: In S3, determining the causal direction of the undirected causal skeleton based on scores specifically includes: introducing the Bayesian information criterion as the scoring function; and using a greedy equivalence search algorithm with scores as the optimization objective to perform edge addition, deletion, and reversal operations on the undirected causal skeleton until the scores no longer increase, thereby obtaining the causal direction of the undirected causal skeleton.

5. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: In S4, the revision includes determining, reversing, deleting, and limiting the time window.

6. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: S4 also includes the operation of recommending high-uncertainty causal edges to experts based on historical templates, so as to prompt experts to review them first. The method for determining high-uncertainty causal edges is as follows: calculate the confidence variance of each causal edge or the posterior probability entropy of the causal edge direction, and determine the edge with the highest entropy value or the largest variance as the highest uncertainty causal edge.

7. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: In S5, a minimum modification set is generated by calculating the graph edit distance, where the graph edit distance is defined as the sum of the minimum cost of editing operations required to transform the current graph into a target graph that satisfies all constraints; a search algorithm is used to find the modification sequence that satisfies all consistency constraints and has the minimum graph edit distance from the minimum modification set.

8. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: The incremental causal discovery process is as follows: When a new time window of data arrives, the new window data is compared with the data from the previous several windows; for variables and their associated edges that have undergone significant statistical changes, based on the previous version of the causal graph, the causal discovery algorithm is re-run only on these highly variable local subgraphs to perform conditional independence testing and scoring optimization in order to update the causal structure of the local region; finally, the updated local subgraphs are merged with the global graph, and the logical correctness of the integrated graph is ensured through global consistency verification.

9. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: In S7, the drift of data distribution is monitored by calculating the KS statistic of data distribution between adjacent windows; the drift of causal structure is monitored by calculating the Jaccard index of structural similarity between causal graphs of adjacent windows.

10. The human-machine collaborative causal correction method for adversarial decision-making as described in claim 1, characterized in that: It also includes the following steps: S8: The causal graph that has passed the consistency check will be stored in a versioned manner, the audit log will be recorded, and the causal graph data will be generated.

Images