Method and apparatus for ai control through decision path comparative evaluation based on failure possibility determination

Abstract

An artificial intelligence control device according to one embodiment may perform the following operations: acquiring a logit vector output from a predetermined artificial intelligence model; using a plurality of logic drivers that determine at least one logic event among logic breakdown of the artificial intelligence model, policy-value mismatch, and instability of the prediction result, determining whether there is a possibility of failure in a decision path corresponding to an original judgment probability distribution generated for the logit vector according to a predefined criterion; if the possibility of failure is determined, generating a re-evaluation path from an upper node within the decision structure of the artificial intelligence model and deriving a re-evaluation probability distribution based on the re-evaluation path; comparing the original judgment probability distribution and the re-evaluation probability distribution to select a probability distribution corresponding to a result with high reliability based on a predetermined reliability index; and deriving a final output of the artificial intelligence model based on the selected probability distribution.

Description

Technology Field

[0001] The present invention relates to a control technology for an artificial intelligence model, and more specifically, to a technology for deriving a more reliable decision by comparing reliability indicators for an original decision path and an alternative path derived by an artificial intelligence model. Background Technology

[0002] Artificial intelligence models have been designed to perform predictions or judgments through a process of statistical generalization based on given data. In particular, deep learning-based neural network models have made rapid advancements in terms of pattern recognition and inference accuracy, and are widely utilized in various application areas. For instance, large-scale language models, game AI, robot control, and decision support systems are representative examples of systems that operate by relying on neural network-based inference results.

[0003] These AI models generate probabilistic responses to inputs and construct decision paths based on optimized model parameters. However, most current AI models are designed to minimize the loss function during the training phase, and since decisions are made during the execution phase based on fixed parameters and model structures, they structurally lack the capability to detect or correct errors or logical inconsistencies occurring during inference in real time. In other words, the system does not possess the flexible adjustment capability to independently re-examine the validity of a determined decision path, adjust distributions according to changed circumstances, or withdraw existing decisions to explore new paths when necessary.

[0004] As a result, when faced with unexpected situations or changes in context, artificial intelligence models tend to repeat incomplete or strategically inappropriate choices without recognizing errors, which acts as a major cause of undermining the reliability and stability of the system, particularly in fields requiring high-level judgment or long-term planning.

[0005] Consequently, even though artificial intelligence models operate relatively stably within the distribution of learned data, they reveal limitations in that they cannot prevent or respond in a timely manner to structural failures such as judgment errors, logic breakdowns, and policy-value mismatches in external environments.

[0006] Against this backdrop, there is a demand for a new control structure capable of logically tracking the system's decision flow during the execution phase of AI models, actively adjusting the decision distribution as needed, or retracing incorrect decision paths to evaluate alternative routes. This approach can complement the rigid reasoning structures of existing AI models and serve as a core foundation for enabling a more adaptive and explainable decision framework. Prior art literature

[0007] Republic of Korea Published Patent Application No. 10-2025-0085220 The problem to be solved

[0008] The problem that the present invention aims to solve is to provide a control structure capable of guaranteeing the consistency and adaptability of inference logic in real time during the execution phase of an artificial intelligence model. To this end, the present invention seeks to implement an execution environment centered on a Logic Driver Layer (LDL) that can be applied in a plug-in manner between the output of an artificial intelligence model and the final decision-making process. This environment is capable of assigning various logical interpretations to the output results of a neural network and adjusting or supplementing the decision path according to the system's purpose, context, and state changes.

[0009] In addition, the Logic-Weighted Distribution for Symbolic Signal Kernel (LWD-SSK) proposed by the present invention performs the role of logically weighting and adjusting the probability distribution of a neural network to induce the selection of a meaningful action path, and the Reevaluation Mechanism (Re:eval) aims to enable the system to autonomously repeat evaluations by returning to the existing decision and attempting an alternative path when structural errors or instability are detected in the decision path.

[0010] The present invention aims to provide a structural foundation that integrates these components in a modular form, enabling universal application without being dependent on a specific artificial intelligence model, and simultaneously realizing improved inference quality, enhanced self-correction capabilities of the system, and flexible control tailored to the user's purpose.

[0011] Meanwhile, the technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by a person skilled in the art from the description below. means of solving the problem

[0012] A method performed by an artificial intelligence control device operated by a processor according to one embodiment may include: acquiring a signal generated by a predetermined artificial intelligence model during a decision-making process; determining whether at least one logical event among logical collapse, policy-value mismatch, and instability of a prediction result has occurred for the signal by using a logic driver configured in a plug-in form to detect the occurrence of a predefined logical event with respect to the output of the artificial intelligence model; dynamically adjusting the value of the signal generated during the execution process of the artificial intelligence model or the decision path of the artificial intelligence model based on whether the logical event has occurred and the type thereof; and deriving a final decision output of the artificial intelligence model based on the adjustment.

[0013] In addition, the above signal may include internal signals or external signals generated during the process in which the artificial intelligence model performs decision-making.

[0014] Additionally, the internal signal may include at least one of a logit vector, a probability distribution, and a confidence score for each judgment candidate generated in the artificial intelligence model, and the external signal may include at least one of a normative behavioral rule, a user preference condition, and a safety constraint condition required for the artificial intelligence model to follow.

[0015] In addition, the operation of determining whether the above logic event has occurred may include an operation of determining whether a logic collapse event has occurred based on the judgment criteria of a CollapseDriver, which is a logic driver that uses a judgment criterion of whether a rapid decrease in the winning rate of the output result of the artificial intelligence model or the occurrence of a structural contradiction is a logic driver.

[0016] In addition, the operation of determining whether the above logic event has occurred may include determining whether an instability event of the prediction result has occurred according to the judgment criteria of a logic driver, SurpriseDriver, which uses the occurrence of a result observed with a probability below a predetermined threshold based on previously stored statistics regarding the output of the artificial intelligence model.

[0017] In addition, the operation of determining whether the above logic event has occurred may include determining whether a policy-value mismatch event has occurred according to the judgment criteria of a ControlDriver, which is a logic driver that uses the judgment criteria of whether the output context of the artificial intelligence model matches a predefined pattern or policy condition.

[0018] In addition, the above-mentioned adjusting operation may include adjusting the selection probability weights of the logit vector corresponding to the output value of the artificial intelligence model, or generating alternative paths separate from the decision path determined by the artificial intelligence model and evaluating each path.

[0019] Additionally, the dynamically adjusting operation may include, when at least one logical event occurs, calculating the logical confidence of a judgment candidate corresponding to each logit value constituting the logit vector output by the artificial intelligence model, correcting the logit vector by applying differential weights to each logit value according to the magnitude of the logical confidence calculated for each logit value, and generating a re-adjusted distribution vector based on the corrected logit vector.

[0020] Additionally, the dynamically adjusting operation may include, when at least one logical event occurs, generating a re-evaluation path from an upper node within the decision structure of the artificial intelligence model, separately from the decision path corresponding to the original judgment probability distribution generated based on the logit vector output by the artificial intelligence model, deriving a re-evaluation probability distribution based on the re-evaluation path, comparing the original judgment probability distribution with the re-evaluation probability distribution, and selecting a decision corresponding to a result with high reliability based on a predetermined reliability index.

[0021] A method performed by an artificial intelligence control device operated by a processor according to one embodiment may include: obtaining a logit vector output from a predetermined artificial intelligence model; calculating the logical reliability of a judgment candidate corresponding to each logit value constituting the logit vector by using a plurality of logic drivers that determine at least one logical event among logical collapse of the artificial intelligence model, policy-value mismatch, and instability of the prediction result; correcting the logit vector by applying differential weights to each logit value according to the magnitude of the logical reliability calculated for each logit value, and generating a rebalanced distribution vector based on the corrected logit vector; and deriving a final output of the artificial intelligence model based on the rebalanced distribution vector.

[0022] In addition, the logit vector output from the artificial intelligence model is composed of logit values ​​corresponding to each of a plurality of judgment candidates, and the logit value corresponding to each judgment candidate includes a score value indicating the probability that each judgment candidate will be selected, and the logical reliability includes a value that quantitatively evaluates whether each logit value is logically reliable, and each judgment candidate may include a candidate class of the result that the artificial intelligence model intends to classify or predict with respect to the input information.

[0023] Additionally, the operation of calculating the logic reliability may include evaluating whether the logic of the logit value corresponding to the judgment candidate has collapsed according to the judgment criteria of CollapseDriver, which is a logic driver that uses the relative logit value deviation or structural distribution distortion between judgment candidates in the distribution of the logit vector as a judgment criterion, and calculating the logic reliability of the judgment candidate based on the evaluation.

[0024] Additionally, the operation of calculating the logic reliability may include evaluating whether there is prediction instability corresponding to the judgment candidate based on the judgment criteria of a SurpriseDriver, which is a logic driver that uses whether the logit value of the judgment candidate deviates from a predefined range of occurrence probability based on pre-stored output statistical information for each judgment candidate, and calculating the logic reliability of the judgment candidate based on the evaluation.

[0025] Additionally, the operation of calculating the logical reliability may include evaluating whether there is a policy-value mismatch corresponding to the judgment candidate according to the judgment criteria of a ControlDriver, which is a logic driver that uses the output of a prediction result conforming to a policy or objective function as a judgment criterion, and calculating the logical reliability of the judgment candidate based on the evaluation.

[0026] Additionally, the operation of generating the recalibrated distribution vector may include an operation of generating a corrected logit vector by assigning weights to each logit value included in the logit vector by multiplying a logit value corresponding to a judgment candidate whose calculated logical reliability is above a predetermined threshold by an amplification factor, and multiplying a logit value corresponding to a judgment candidate whose calculated logical reliability is below a predetermined threshold by an attenuation factor.

[0027] Additionally, the operation of generating the rebalanced distribution vector may include the operation of generating the rebalanced distribution vector by applying a normalization function to the corrected logit vector to convert it into a probability distribution for each judgment candidate.

[0028] In addition, the operation of generating the above-mentioned rebalanced distribution vector may include the operation of generating the above-mentioned rebalanced distribution vector based on the following mathematical formula 1, which applies a normalization function to the above-mentioned corrected logit vector to convert it into a probability distribution for each judgment candidate.

[0029] [Mathematical Formula 1]

[0030]

[0031] ( is the rebalanced distribution vector, is the corrected logit vector, is the logit vector output from the artificial intelligence model, is the logical reliability calculated by a single logical driver, is a weight that is differentially set according to the magnitude of the confidence calculated by a single logic driver,)

[0032] In addition, the operation of obtaining a logit vector output from the artificial intelligence model may include, when the artificial intelligence model is composed of an ensemble structure including a plurality of neural networks, an operation of constructing an integrated logit vector output from the artificial intelligence model based on the output logit vector of each neural network and weight coefficients according to the input according to the following mathematical formula 2.

[0033] [Mathematical Formula 2]

[0034]

[0035] ( is the combined logit vector output from multiple neural networks, m is the total number of multiple neural networks, is the logit vector output by the j-th neural network model, is the weight coefficient assigned to the j-th neural network for input x)

[0036] A method performed by an artificial intelligence control device operated by a processor according to one embodiment may include: obtaining a logit vector output from a predetermined artificial intelligence model; using a plurality of logic drivers that determine at least one logic event among logic breakdown of the artificial intelligence model, policy-value mismatch, and instability of the prediction result, determining whether there is a possibility of failure in a decision path corresponding to an original judgment probability distribution generated for the logit vector according to a predefined criterion; if the possibility of failure is determined, generating a re-evaluation path from an upper node within the decision structure of the artificial intelligence model and deriving a re-evaluation probability distribution based on the re-evaluation path; comparing the original judgment probability distribution and the re-evaluation probability distribution to select a probability distribution corresponding to a result with high reliability based on a predetermined reliability index; and deriving a final output of the artificial intelligence model based on the selected probability distribution.

[0037] In addition, the operation of determining whether the above-mentioned possibility of failure exists may include the operation of identifying a judgment candidate whose probability value among the original judgment probability distributions is greater than or equal to a predetermined threshold, and extracting the decision path by tracing the internal layers of the artificial intelligence model until the identified judgment candidate is derived and configuring the path of the main node reaching the identified judgment candidate.

[0038] Additionally, the operation of determining whether the possibility of failure exists may include evaluating whether the logic of the decision path corresponding to the original decision probability distribution collapses according to the judgment criteria of CollapseDriver, a logic driver that takes the relative logit value deviation or structural distribution distortion between judgment candidates in the distribution of the logit vector as a judgment criterion, and determining the possibility of failure of the decision path based on the evaluation.

[0039] Additionally, the operation of determining whether the above-mentioned possibility of failure exists may include evaluating the instability of the prediction result corresponding to the original judgment probability distribution based on the judgment criteria of a SurpriseDriver, which is a logic driver that uses whether the logit value of a judgment candidate within the original judgment probability distribution deviates from a predefined range of occurrence probability based on the output statistical information for each judgment candidate stored in advance, and determining the possibility of failure of the above-mentioned decision path based on the evaluation.

[0040] Additionally, the operation of determining whether the above-mentioned possibility of failure exists may include evaluating whether there is a policy-value mismatch of the prediction result corresponding to the above-mentioned original judgment probability distribution according to the judgment criteria of a ControlDriver, which is a logic driver that uses the output of a prediction result conforming to a policy or objective function as a judgment criterion, and determining the possibility of failure of the above-mentioned decision path based on the evaluation.

[0041] In addition, the operation of deriving the above re-evaluation probability distribution may include the operation of identifying the major decision points that have had an influence among the upper nodes within the decision structure of the artificial intelligence model.

[0042] Additionally, the operation of deriving the re-evaluation probability distribution may include tracking the key decision points through backpropagation-based contribution analysis or attention weight analysis to establish a re-evaluation path, and deriving a re-evaluation probability distribution based on the re-evaluation path.

[0043] In addition, the operation of deriving the above re-evaluation probability distribution may include the operation of setting the depth of the re-evaluation path based on the following mathematical formula 1.

[0044] [Mathematical Formula 1]

[0045]

[0046] ( is the depth of the reevaluation path that determines how many steps prior to reevaluation in the decision path, is the magnitude of win rate collapse, indicating the degree to which the win rate drops sharply after the AI ​​model's decision, is the number of times the Principal Variation Rank changed while traversing the decision path, is the sum of the number of error detections detected by the logical driver, is the weighting coefficient for each term)

[0047] Additionally, the operation of selecting the above decision may include an operation of calculating a reliability index based on the value of Shannon entropy or KL Divergence for each of the above original judgment distribution and the above re-evaluation distribution. Effects of the invention

[0048] The present invention can be extended beyond the existing method that simply relies on the statistical output of a neural network to an active decision system capable of structurally adjusting and supplementing the execution decision flow through a Logic Driver Layer (LDL) applicable in a plug-in manner.

[0049] In particular, the present invention can perform the role of guiding a judgment direction appropriate to the situation in real time by operating as a plugin between the output of an artificial intelligence model and the final decision-making process. Through this, the artificial intelligence model to which the present invention is applied can implement high-level cognitive functions at the execution stage, such as dynamic changes in judgment strategies based on context, goal-based behavior guidance, and pre-detection and response to potential failures.

[0050] Specifically, the present invention adjusts the search probability by applying logical weights to the logits generated by the neural network through the Logic-Weighted Distribution for Symbolic Signal Kernel (LWD-SSK), thereby enabling a more explainable and user-intended action selection.

[0051] In addition, if structural errors or logical breakdowns are detected during the judgment process through the Reevaluation Mechanism (Re:eval), an alternative path different from the existing path is additionally generated and compared with the other; by reflecting the path that yielded the superior result, a highly reliable decision can be derived.

[0052] As such, the method proposed by the present invention can be easily applied without changing the structure of existing neural network models, and thus can provide the effect of significantly improving the consistency and stability of judgment quality, as well as the autonomy and resilience of the system, in various application environments.

[0053] Meanwhile, the effects of the present invention are not limited to those mentioned above, and other unmentioned technical effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0054] FIG. 1 is a configuration diagram showing a computing environment in which an artificial intelligence control device according to one embodiment operates. FIG. 2 is a configuration diagram of an artificial intelligence control device according to one embodiment. FIG. 3 is a flowchart showing the steps of an operation performed by an artificial intelligence control device according to one embodiment. FIG. 4 is a conceptual diagram showing an overall architecture including a plurality of layers controlled by an artificial intelligence control device according to one embodiment in accordance with the operation of FIG. 3, and a flow of data processed in accordance with the operation of FIG. 3. Figure 5 is an example of a logic driver configured in the form of a plugin to detect the occurrence of a predefined logic event on the output of an artificial intelligence model. FIG. 6 is an example of a user interface showing a readjusted probability distribution of a judgment candidate group based on the detection of a logical event of a logical driver according to one embodiment. FIG. 7 is a flowchart showing the steps of the operation performed by the logit correction layer of an artificial intelligence control device according to one embodiment. FIG. 8 is a flowchart showing the steps of an operation performed by a decision coordination layer of an artificial intelligence control device according to one embodiment. Specific details for implementing the invention

[0055] Detailed information regarding the purpose, technical configuration, and resulting effects of the present invention will be more clearly understood through the following detailed description based on the drawings attached to the specification of the present invention. An embodiment according to the present invention will be described in detail with reference to the attached drawings.

[0056] The embodiments disclosed herein should not be interpreted or used to limit the scope of the invention. It is obvious to those skilled in the art that the description including the embodiments herein has various applications. Accordingly, any embodiments described in the detailed description of the invention are illustrative for better explaining the invention and are not intended to limit the scope of the invention to the embodiments.

[0057] The functional blocks shown in the drawings and described below are merely examples of possible implementations. In other implementations, other functional blocks may be used without departing from the spirit and scope of the detailed description. Additionally, while one or more functional blocks of the present invention are shown as individual blocks, one or more of the functional blocks of the present invention may be a combination of various hardware and software configurations that perform the same function.

[0058] Furthermore, the expression that it includes certain components is an “open-ended” expression that merely refers to the existence of such components and should not be understood as excluding additional components.

[0059] Furthermore, when it is stated that one component is “connected” or “joined” to another component, it should be understood that while it may be directly connected or joined to that other component, there may also be other components present in between.

[0060] Hereinafter, various embodiments of the present invention are described with reference to the accompanying drawings. However, this is not intended to limit the present invention to specific embodiments and should be understood to include various modifications, equivalents, and / or alternatives of the embodiments of the present invention.

[0061] The present invention proposes an artificial intelligence control device (100) that implements a technology for interpreting a logical signal regarding the inference result of an artificial intelligence model and thereby dynamically adjusting the decision-making of the artificial intelligence model.

[0062] Accordingly, the description of the present specification explains the configuration of a computing environment in which an artificial intelligence control device (100) operates together with FIG. 1, and then sequentially explains the components of the artificial intelligence control device (100) and the operation process thereof with reference to the drawings that follow.

[0063] FIG. 1 is a configuration diagram showing a computing environment in which an artificial intelligence control device (100) according to one embodiment operates.

[0064] Referring to FIG. 1, the configuration proposed in the present invention consists of a Logic Driver Layer (LDL) (20), a Logit Adjustment Layer (LWD-SSK) (30), and a Decision Adjustment Layer (Re:eval) (40), which can be implemented in a structure that can be inserted as a plug-in at the end of neural network signal processing.

[0065] Here, a 'layer' refers to a unit of computation or processing module designed to operate individually at a stage after the neural network output to perform a specific function, and functions as an external control component that operates separately from the neural network body.

[0066] This structure can be applied without changing the architecture of existing neural network models, thus providing high versatility and scalability. Each module operates in an independent yet interconnected manner, allowing them to be called individually or interact depending on the situation. In other words, LWD-SSK and Re:eval operate as lower layers of LDL and can also operate independently.

[0067] Specifically, the computing environment of the present invention may include a logic driver layer (20) that receives the output of a neural network signal (10) and performs subsequent control, a logit correction layer (30) and a decision adjustment layer (40) that perform correction or evaluation work according to the result of the judgment.

[0068] At this time, the neural network signal (10) may include internal or external signals related to the neural network model, and the logic driver layer (20) determines a structural logic signal through a plurality of logic drivers, such as CollapseDriver, SurpriseDriver, and ControlDriver, which will be described later, based on the signal, and performs control accordingly.

[0069] The logic driver layer (20) outputs these control results through the user interface (21) and can optionally call at least one subsequent module among the logit correction layer (30) or the decision adjustment layer (40).

[0070] The logit correction layer (30) corrects the output distribution of the neural network by adjusting the logit value according to the judgment signal transmitted from the logic driver layer (20), and the decision adjustment layer (40) performs a re-evaluation of the control judgment of the logic driver layer (20) to search for alternative paths or adjust the decision reliability.

[0071] These logic driver layer (20), logit correction layer (30), and decision coordination layer (40) can be implemented in the form of submodules that are integrated as plug-ins into a processor-based computing device. Specifically, the computing environment illustrated in FIG. 1 can be implemented through the configuration of FIG. 2.

[0072] FIG. 2 is a configuration diagram of an artificial intelligence control device (100) (hereinafter referred to as 'device (100)') according to one embodiment.

[0073] Referring to FIG. 2, a device (100) according to one embodiment may each include a memory (110), a processor (120), an input / output interface (130), and a communication interface (140).

[0074] The memory (110) can store data obtained from an external device or data generated by itself. The memory (110) can store instructions that can perform operations of the processor (120). For example, the memory (110) can store program code for implementing the logic driver layer (20), the logit correction layer (30), and the decision adjustment layer (40), as well as data such as parameters, model information, and control conditions required for their execution.

[0075] The processor (120) is a computational device that controls the overall operation. The processor (120) can execute instructions stored in memory (110). The operation of the device (100) according to the embodiment of the present document can be understood as an operation performed by the processor (120).

[0076] The input / output interface (130) may include a hardware interface or a software interface for inputting or outputting information.

[0077] The communication interface (140) enables the transmission and reception of information through a communication network. To this end, the communication interface (140) may include a wireless communication module or a wired communication module.

[0078] The device (100) can be implemented in various forms of devices capable of performing calculations through a processor (120) and transmitting and receiving information through a network. For example, it can be implemented in the form of a server, a computer device, a portable communication device, a smartphone, a portable multimedia device, a laptop, a tablet PC, etc., but is not limited to these examples.

[0079] In the following description, the configuration and judgment flow of the logic driver layer (20) will be described with reference to FIGS. 3 to 5, followed by the logit-based probability distribution correction method of the logit correction layer (30) with reference to FIG. 7, and the path evaluation and recombination mechanism performed by the decision adjustment layer (40) will be described in detail with reference to FIG. 8.

[0080] At this time, the logic driver layer (20) performs the role of a higher-level control system that analyzes the neural network signal received from the artificial intelligence model and generates a structured logic signal to trigger subsequent correction or evaluation operations.

[0081] Depending on the judgment of the logic driver layer (20), the logit correction layer (30) and the decision adjustment layer (40) may be called to perform subsequent operations, and at the same time, they may operate individually or in parallel according to independent control flows.

[0082] However, the operation of the logit correction layer (30) and the decision adjustment layer (40) can be activated independently according to specific conditions or predefined policies, regardless of the control flow of the logic driver layer (20), and can function as an independent module that performs correction of the neural network output or reliability adjustment of the decision itself.

[0083] FIG. 3 is a flowchart showing the steps of an operation in which a device (100) according to one embodiment performs an operation of a logic driver layer (20), and FIG. 4 is a conceptual diagram showing an overall architecture including a plurality of layers controlled by a device (100) according to one embodiment according to the operation of FIG. 3, and a flow of data processed according to the operation of FIG. 3. The operation of the device (100) according to the embodiments of FIG. 3 and FIG. 4 can be understood as an operation performed by a processor (120).

[0084] Each step disclosed in FIGS. 3 and 4 is merely a preferred embodiment for achieving the purpose of the present invention, and some steps may be added or deleted as needed, and any one step may be included in another step. The order of each operation disclosed in FIGS. 3 and 4 is arranged for convenience of understanding only, and such order is not limited to a chronological order, and the order may be changed and operated differently according to the designer's choice.

[0085] Referring to FIGS. 3 and FIGS. 4, in step S1010, the device (100) can acquire a signal generated by a predetermined artificial intelligence model in a decision-making process.

[0086] Here, an artificial intelligence model refers to a neural network or machine learning-based model trained to perform inference, such as prediction, classification, and judgment, based on input data, and may include, for example, large language models (LLMs) such as Transformer, CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), BERT (Bidirectional Encoder Representations from Transformers), and GPT (Generative Pretrained Transformer). However, the artificial intelligence model applied to the embodiments of the present invention is not limited to a specific structure and can be applied to any form of inference model capable of generating an output signal.

[0087] In addition, the above signal may include internal signals or external signals generated during the process of the artificial intelligence model making a decision.

[0088] For example, internal signals may include at least one of a logit vector, a probability distribution, and a confidence score for each judgment candidate generated by the artificial intelligence model, and external signals may include at least one of a normative behavioral rule, a user preference condition, and a safety constraint condition required for the artificial intelligence model to follow.

[0089] In step S1020, the device (100) can determine whether at least one of the logic events—logic breakdown of a signal, policy-value mismatch, and instability of a prediction result—occurs by using a logic driver configured in a plug-in form to detect the occurrence of a predefined logic event on the output of an artificial intelligence model.

[0090] Figure 5 is an example of a logic driver configured in the form of a plugin to detect the occurrence of a predefined logic event on the output of an artificial intelligence model.

[0091] Referring to FIG. 5, the logic driver may be composed of CollapseDriver, SurpriseDriver, and ControlDriver, each of which may be designed to respond to logic events occurring under specific conditions.

[0092] A logic driver corresponds to a subunit or functional plugin on a logic driver layer (20) that is implemented in a modularized logic tree-based structure. Such a logic driver performs the function of automatically detecting logical features, such as abnormalities, structural contradictions, or stable patterns inherent in the inference results of an artificial intelligence model, and generating a trigger signal for subsequent correction or path evaluation operations.

[0093] Logical events refer to structural anomalies, reliability degradation, or abnormal judgment flows detected in the output or inference path of an artificial intelligence model, and may include situations where the model's decision-making involves signals that contradict the overall system's policies, values, or prediction structures, going beyond mere probabilistic errors.

[0094] In other words, logical events are distinct from general output errors or low accuracy rates, and focus on detecting structural decision-making problems or path selection errors that the AI ​​faces internally, such as prediction collapse, policy-value mismatch, and control failure. These logical events are identified based on trigger conditions defined in each logical driver (CollapseDriver, SurpriseDriver, ControlDriver) and can be utilized as reference signals for correction or re-evaluation by subsequent layers (LWD-SSK, Re:eval).

[0095] For example, a CollapseDriver can operate by lowering the confidence in the current decision path and increasing the rollback score in response to situations such as a sharp decrease in the winning rate of an AI model's output or structural contradictions. The CollapseDriver can be implemented by analyzing the AI ​​model's output and internal feedback information to detect decision failure patterns that exceed defined threshold conditions, and can generate logic collapse events based on monitoring threshold values ​​for specific signal sets (e.g., winrate curve, policy-value difference, etc.). Such a CollapseDriver is primarily suitable for deep reinforcement learning-based models or decision systems requiring multi-path evaluation, and functions as a core driver to ensure the structural stability of the decision flow.

[0096] Based on pre-stored statistics regarding the output of an AI model, the SurpriseDriver responds to abnormal events—such as incidents or exceptional situations observed with a probability below a predetermined threshold—and plays a role in increasing the propensity to explore and adjusting the output distribution to deviate from the base path. The SurpriseDriver can be implemented to identify events of instability in prediction results through anomaly detection or statistical deviation analysis techniques, based on criteria such as the frequency of a given signal, the degree of deviation in the distribution, or the rate of non-linear change relative to the previous decision flow. Such a SurpriseDriver is primarily useful in adaptive decision systems or reinforcement learning environments where discovering unexplored paths is critical; it functions as an exploration enhancement module that breaks away from the system's fixed decision paths and improves responsiveness to unstructured situations.

[0097] A ControlDriver can determine whether a policy-value mismatch event has occurred by analyzing, based on specific criteria, whether the output context of an AI model matches predefined patterns or policy conditions. The ControlDriver is activated when an input signal shows a high degree of agreement with an existing policy or a learned action value function, and can be implemented based on criteria such as pattern-matching triggers, stability indicators of decision paths, or the downward trend of the confusion index. Such a ControlDriver guides the system to avoid excessive exploration or changes in decisions in low-uncertainty environments and is utilized for purposes such as maintaining policy stability, ensuring system consistency, and optimizing execution efficiency.

[0098] As described above, the logic drivers generate structured logic signals based on neural network output signals, and each signal can be used as a trigger for correction or re-evaluation in a subsequent step.

[0099] In step S1030, the device (100) can dynamically adjust the value of a signal generated during the execution of the artificial intelligence model or the decision path of the artificial intelligence model based on whether a logic event occurs and the type of logic event (e.g., logic breakdown, policy-value mismatch, and instability of the prediction result). To do this, the device (100) may selectively call at least one of a logit correction layer (30) or a decision adjustment layer (40) designed to respond to the corresponding logic event.

[0100] The logit correction layer (30) can reconstruct the output distribution by adjusting the selection probability weights of the logit vectors corresponding to the output values ​​of the artificial intelligence model, and the decision adjustment layer (40) can re-evaluate alternative paths other than the current path or readjust the decision reliability to reset the direction of the decision flow.

[0101] For example, when at least one logical event occurs, the logit correction layer (30) can calculate the logical confidence of a judgment candidate corresponding to each logit value constituting the logit vector output by the artificial intelligence model, correct the logit vector by applying differential weights to each logit value according to the magnitude of the logical confidence calculated for each logit value, and generate a rebalanced distribution vector based on the corrected logit vector. The operation of the logit correction layer (30) will be described in more detail later together with FIG. 7.

[0102] For example, when at least one logical event occurs, the decision adjustment layer (40) may generate a re-evaluation path from a higher node within the decision structure of the artificial intelligence model, separate from the decision path corresponding to the original judgment probability distribution generated based on the logit vector output by the artificial intelligence model, derive a re-evaluation probability distribution based on the re-evaluation path, and compare the original judgment probability distribution with the re-evaluation probability distribution to select a decision corresponding to a result with high reliability based on a predetermined reliability index. The operation of the decision adjustment layer (40) will be explained in more detail later together with FIG. 8.

[0103] In step S1040, the device (100) can derive the final decision output of the artificial intelligence model based on the adjustment of the preceding step.

[0104] Such adjustments can be performed sequentially or in parallel using a control signal of the logic driver layer (20) as a trigger, and the final output of the artificial intelligence model can be flexibly controlled through a multi-stage evaluation and correction procedure depending on the situation.

[0105] At this time, the final output may be changed as a result of logit correction or path re-evaluation, and the selected final decision may be provided externally through a user interface or transmitted to a subsequent system as shown in Fig. 6 below.

[0106] FIG. 6 is an example of a user interface showing a readjusted probability distribution of a judgment candidate group based on the detection of a logical event of a logical driver according to one embodiment.

[0107] Referring to Figure 6, the base probability and adjusted probability are displayed for judgment candidates A through F, along with the driver effects for each candidate. For example, it can be seen that candidate A had an initial probability of 0.25, but the adjusted probability increased to 0.30 due to the influence of logical events such as 'Instability' and 'Phase'. On the other hand, the probability of candidate C decreased due to the same logical events.

[0108] At the bottom of Fig. 6, "Collapse Detected" is indicated as "STEP BACK TRIGGER," which means that the CollapseDriver has detected a structural inconsistency or a sudden drop in win rate, and the reliability of the current path has collapsed. At this time, the user or system is provided with an interface to perform a re-evaluation through the "Reevaluate" button.

[0109] This interface configuration intuitively visualizes the results of logit correction, while also providing a foundation to transparently review the decision adjustment flow caused by the influence of the logic driver and to induce subsequent actions manually or automatically if necessary.

[0110] Next, the logit-based probability distribution correction method of the logit correction layer (30) is described through FIG. 7.

[0111] FIG. 7 is a flowchart illustrating the steps of an operation performed by a logit correction layer (30) by a device (100) according to one embodiment. The operation of the device (100) according to the embodiment of FIG. 7 can be understood as an operation performed by a processor (120).

[0112] Each step disclosed in FIG. 7 is merely a preferred embodiment for achieving the purpose of the present invention, and some steps may be added or deleted as needed, and any one step may be included in another step. The order of each operation disclosed in FIG. 7 is arranged for ease of understanding only, and such order is not limited to a chronological order, and the order may be changed and operated differently according to the designer's choice.

[0113] Referring to FIG. 7, in step S2010, the device (100) can obtain a logit vector output from an artificial intelligence model.

[0114] In this case, the logit vector consists of logit values ​​corresponding to each of the multiple judgment candidates, and each logit value includes a score value indicating the probability that the corresponding judgment candidate will be selected. A judgment candidate refers to multiple options or response items that can be derived as a result of a classification or prediction task that an artificial intelligence model intends to perform on given input information. For example, a judgment candidate may include candidate classes of results that the artificial intelligence model intends to classify or predict regarding the input information.

[0115] In step S2020, the device (100) can calculate the logical reliability of a judgment candidate corresponding to each logit value constituting a logit vector by using a plurality of logic drivers that determine at least one logic event among the logic breakdown of an artificial intelligence model, policy-value discrepancy, and instability of a prediction result.

[0116] Here, logical reliability refers to an indicator that quantitatively evaluates whether a logit value assigned by an artificial intelligence model to a specific judgment candidate has logical reliability. The device (100) can calculate logical reliability by quantitatively evaluating whether the logit value is logically reliable by reflecting logical event information derived through a logic driver for each logit value corresponding to a judgment candidate.

[0117] For example, logical reliability can be calculated based on criteria such as whether the candidate for judgment is consistent with existing policies or value judgments without structural contradictions, a level of stability where the prediction probability is not excessively low or fluctuates rapidly, consistency with the current context, and whether the choice is based on a statistically valid pattern.

[0118] Therefore, logical reliability is a value that quantifies whether a judgment candidate is derived within a valid logical flow when selected, separate from the magnitude of the logit value, and can be used as an important criterion for logit correction or judgment path re-evaluation in subsequent stages.

[0119] The logical reliability calculated at this time can be adjusted according to the results of one or more interventions among CollapseDriver, SurpriseDriver, and ControlDriver, and is used to correct the selection reliability of the judgment candidate corresponding to each logit value.

[0120] For example, the device (100) may evaluate whether the logic of the logit value corresponding to the judgment candidate has collapsed according to the judgment criteria of the CollapseDriver, which is a logic driver that takes the relative logit value deviation or structural distribution distortion between judgment candidates in the distribution of the logit vector as a judgment criterion, and may calculate the logic reliability of the judgment candidate based on the evaluation.

[0121] For example, the device (100) can evaluate whether there is prediction instability corresponding to a judgment candidate based on the judgment criteria of a SurpriseDriver, which is a logic driver that uses whether the logit value of a judgment candidate deviates from a predefined range of occurrence probability based on the previously stored output statistics for each judgment candidate, and can calculate the logic reliability of the judgment candidate based on the evaluation.

[0122] For example, the device (100) can evaluate whether there is a policy-value mismatch corresponding to a judgment candidate based on the judgment criteria of a ControlDriver, which is a logic driver that uses the output of a prediction result corresponding to a policy or objective function as a judgment criterion, and calculate the logic reliability of the judgment candidate based on the evaluation.

[0123] In step S2030, the device (100) can correct the logit vector by applying differential weights to each logit value according to the magnitude of the logical reliability calculated for each logit value, and generate a rebalanced distribution vector based on the corrected logit vector.

[0124] For example, the device (100) can generate a corrected logit vector by assigning weights to each logit value included in the logit vector by multiplying a logit value corresponding to a judgment candidate whose logical reliability calculated for each judgment candidate is greater than or equal to a predetermined threshold by an amplification factor (e.g., a weight greater than 1) and multiplying a logit value corresponding to a judgment candidate whose logical reliability calculated for each judgment candidate is less than the predetermined threshold by a damping factor (e.g., a weight greater than 0 and less than 1). Additionally, the device (100) can generate a rebalanced distribution vector by applying a normalization function (e.g., a softmax function, etc.) to the corrected logit vector to convert it into a probability distribution for each judgment candidate.

[0125] For example, the above process can be implemented as shown in the following mathematical formula 1.

[0126] [Mathematical Formula 1]

[0127]

[0128] ( is the rebalanced distribution vector, is the corrected logit vector, is the logit vector output from the artificial intelligence model, is the logical reliability calculated by a single logical driver, is a weight that is differentially set according to the magnitude of the confidence calculated by a single logic driver,)

[0129] According to Equation 1, the value of logical reliability can vary based on the reliability results calculated by logical drivers such as CollapseDriver, SurpriseDriver, and ControlDriver. In other words, since the criteria for detecting logical events for the same input may differ for each driver, the logical reliability calculated for each decision candidate may also appear differently. For example, CollapseDriver may react sensitively to structural contradictions and calculate a rapidly low reliability, whereas ControlDriver may assign a high reliability to the same decision candidate within stable patterns.

[0130] Accordingly, the weight assigned to the judgment candidate can be dynamically adjusted according to the value of the logical reliability of the specific judgment candidate calculated by the logic driver used. Thus, the weight applied to each judgment candidate can be set to reflect the judgment result of the most appropriate logic driver in a specific situation, or the value of the driver with the highest priority among the reliability values ​​calculated by multiple logic drivers can be selected or combined using a weighted average method. In this way, the device (100) can perform fine output correction considering the characteristics of the input situation and can effectively suppress the inclination toward unstable or irrational judgments.

[0131] Meanwhile, if the artificial intelligence model used in the present invention is configured as an ensemble structure including a plurality of neural networks, the device (100) constructs an integrated logit vector output from the artificial intelligence model based on the output logit vector of each neural network and weight coefficients according to the input according to the following mathematical formula 2, and the integrated logit vector is of the mathematical formula 1 It can be used as.

[0132] [Mathematical Formula 2]

[0133]

[0134] ( is the combined logit vector output from multiple neural networks, m is the total number of multiple neural networks, is the logit vector output by the j-th neural network model, is the weight coefficient assigned to the j-th neural network for input x)

[0135] According to Equation 2, the logit vectors output by each neural network are assigned differential weights based on the characteristics of the input to form an integrated logit vector. This structure helps to produce more appropriate and stable outputs in specific decision situations by reflecting the expertise and reliability of each neural network according to the input situation.

[0136] In step S2040, the device (100) can derive the final output of the artificial intelligence model based on the distribution vector readjusted in the previous step.

[0137] At this time, the readjusted distribution vector includes a selection probability corresponding to each of the multiple judgment candidates, and the device (100) can select the judgment candidate having the highest probability value among them as the final result, or set multiple judgment candidates satisfying a predetermined probability threshold or higher as a candidate group and determine the final output through interaction with a subsequent system or user.

[0138] Since this final output is a result reconstructed based on reliability through the correction procedure of the logit correction layer (30), it may differ from the result output by the original artificial intelligence model and can be interpreted as a more precise output that reflects the judgment result of the logic driver.

[0139] The device (100) can display the final output result externally through a user interface or transmit it to another system module for use in subsequent processing. This configuration helps to provide consistent and reliable results even in uncertain situations or under complex judgment conditions.

[0140] Next, with reference to FIG. 8, the path evaluation and re-determination mechanism performed by the decision coordination layer (40) will be described.

[0141] FIG. 8 is a flowchart illustrating the steps of an operation performed by a decision coordination layer (40) by a device (100) according to one embodiment. The operation of the device (100) according to the embodiment of FIG. 7 can be understood as an operation performed by a processor (120).

[0142] Each step disclosed in FIG. 8 is merely a preferred embodiment for achieving the purpose of the present invention, and some steps may be added or deleted as needed, and any one step may be included in another step. The order of each operation disclosed in FIG. 8 is arranged for ease of understanding only, and such order is not limited to a chronological order, and the order may be changed and operated differently according to the designer's choice.

[0143] Referring to FIG. 8, in step S3010, the device (100) can obtain a logit vector output from an artificial intelligence model.

[0144] In this case, the logit vector consists of logit values ​​corresponding to each of the multiple judgment candidates, and each logit value includes a score value indicating the probability that the corresponding judgment candidate will be selected. A judgment candidate refers to multiple options or response items that can be derived as a result of a classification or prediction task that an artificial intelligence model intends to perform on given input information. For example, a judgment candidate may include candidate classes of results that the artificial intelligence model intends to classify or predict regarding the input information.

[0145] In step S3020, the device (100) can determine whether there is a possibility of failure of a decision path corresponding to the original judgment probability distribution generated for the logit vector according to a predefined criterion by using a plurality of logic drivers that determine at least one logic event among logic breakdown of the artificial intelligence model, policy-value mismatch and instability of the prediction result.

[0146] For example, the device (100) can generate an original judgment probability distribution by applying a softmax function to a logit vector to calculate a probability value for each judgment candidate. The original judgment probability distribution contains the judgment result finally derived by the artificial intelligence model for the input and the confidence level therein.

[0147] Here, a decision path refers to the flow of inference or prediction procedures internally performed by an artificial intelligence model to select a specific judgment candidate based on input information. This may include a computation path within the artificial intelligence model, an activated neural network path, or decision logic to reach the top candidate within a probability distribution derived based on a logit vector. For example, the device (100) can extract a decision path by identifying a judgment candidate whose probability value in the original judgment probability distribution is greater than or equal to a predetermined threshold, and by tracing the internal layers of the artificial intelligence model until the identified judgment candidate is derived to construct the path of the main node that reaches the identified judgment candidate.

[0148] The probability of failure refers to the likelihood that incorrect predictions, low reliability, ethical conflicts, or discrepancies with user requirements will occur when an artificial intelligence model selects a specific decision path, and this can be evaluated through statistical failure rates, the frequency of logical events, or compliance with external standards.

[0149] A predefined criterion refers to a mathematical or logical condition set in advance to evaluate the probability of failure of a decision path, and may include, for example, the following examples.

[0150] For example, the device (100) can evaluate whether the logic of a decision path corresponding to the original decision probability distribution has collapsed according to the judgment criteria of a CollapseDriver, which is a logic driver that takes the relative logit value deviation or structural distribution distortion between decision candidates in the distribution of the logit vector as a judgment criterion, and determine the possibility of failure of the decision path based on the evaluation.

[0151] For example, the device (100) can evaluate whether there is instability in the prediction result corresponding to the original judgment probability distribution based on the judgment criteria of a SurpriseDriver, which is a logic driver that uses whether the logit value of a judgment candidate within the original judgment probability distribution deviates from a predefined range of occurrence probability based on the previously stored output statistical information for each judgment candidate, and can determine the possibility of failure of the decision path based on the evaluation.

[0152] For example, the device (100) can evaluate whether there is a policy-value mismatch of the prediction result corresponding to the original judgment probability distribution according to the judgment criteria of the ControlDriver, which is a logic driver that uses the judgment criteria for whether the output of the prediction result that matches the policy or objective function is a judgment criterion, and determine the possibility of failure of the decision path based on the evaluation.

[0153] In step S3030, if the device (100) is determined to have a possibility of failure in the preceding step, it can generate a re-evaluation path from a top node within the decision-making structure of the artificial intelligence model and derive a re-evaluation probability distribution based on the re-evaluation path.

[0154] At this time, the device (100) can perform more effective correction decisions by identifying key decision points that have influenced the decision structure of the artificial intelligence model and establishing a re-evaluation path centered on them. For example, the device (100) can establish a re-evaluation path by tracking key decision points through backpropagation-based attribution or attention weight analysis. The device (100) can generate a new logit vector based on the internal representation generated at the node of the path along the re-evaluation path thus established, and derive a re-evaluation probability distribution based thereon.

[0155] When establishing such a re-evaluation path, it is important to determine the depth of the re-evaluation. If an excessively shallow path is selected, it may fail to accurately correct the problem because it cannot sufficiently trace back to the cause of the error in the original judgment; conversely, if an excessively deep path is selected, it may affect higher-level generalization nodes unrelated to the original input, leading to increased unnecessary computations and reduced judgment stability.

[0156] Accordingly, the device (100) can minimize judgment errors in the model and secure both efficiency and reliability of re-evaluation by setting the re-evaluation depth based on the following mathematical formula 3, which comprehensively considers the degree of win rate collapse, instability indicators within the path, and severity of driver detection events, so that it can return to an appropriate depth centered on the point where the cause of judgment failure occurred and perform re-inference.

[0157] [Mathematical Formula 3]

[0158]

[0159] ( is the depth of the reevaluation path that determines how many steps prior to reevaluation in the decision path, is the magnitude of win rate collapse, indicating the degree to which the win rate drops sharply after the AI ​​model's decision, is the number of times the Principal Variation Rank changed while traversing the decision path, is the sum of the number of error detections detected by the logical driver, is the weighting coefficient for each term)

[0160] If mathematical equation 3 is defined in simpler terms, it is as follows mathematical equation 4.

[0161] [Mathematical Formula 4]

[0162]

[0163] (d: Depth of reevaluation path, : Post-decision winrate collapse magnitude, : PVR fluctuation count (Principal Variation Rank fluctuation count), E: Number of logic collapse / error events detected by the logic driver, Experimental-based weights (Tunable weight coefficients)

[0164] In step S3040, the device (100) can compare the original judgment probability distribution with the re-evaluation probability distribution and select a probability distribution corresponding to a result with high reliability based on a predetermined reliability index.

[0165] Reliability indicators can include various indicators that quantitatively represent the stability of decision-making results and the certainty of predictions based on probability distributions.

[0166] For example, the device (100) can calculate a reliability index using the maximum value within each probability distribution, the cumulative probability value for the top multiple judgment candidates, the inverse of the entropy of the probability distribution, or the average value of the logical reliability corresponding to the judgment candidate. The device (100) can select the distribution with higher reliability between the original judgment distribution and the re-evaluation distribution according to a composite index composed of a combination of these elements or a predefined criterion function, and select the judgment candidate corresponding to the maximum value of the distribution as the final decision.

[0167] For example, the device (100) can calculate a Shannon entropy or KL Divergence-based logical reliability statistic for each of the original judgment distribution and the re-evaluation distribution, and select a probability distribution corresponding to the distribution with higher reliability among the two distributions.

[0168] In step S3050, the device (100) can derive the final output of the artificial intelligence model based on the probability distribution selected in the preceding step. At this time, the selected probability distribution is a distribution that has a higher reliability based on a predetermined reliability index among the original judgment probability distribution or the re-evaluation probability distribution.

[0169] This final output is not the result produced by existing AI models, but rather the outcome of logic driver-based error detection, decision path reconstruction, and reliability re-evaluation; therefore, it can be interpreted as a more precise and stable decision. In particular, since the re-evaluation path is constructed through contribution analysis of key decision points and recursive review within the decision structure, this output can be provided to users or external systems in a form that ensures logical consistency and result reliability.

[0170] The device (100) can display the final output result externally through a user interface or transmit it to another system module for use in subsequent processing. This configuration helps to provide consistent and reliable results even in uncertain situations or under complex judgment conditions.

[0171] According to the above-described embodiment, the present invention can be extended beyond the existing method that simply relies on the statistical output of a neural network to an active decision system capable of structurally adjusting and supplementing the execution decision flow through a logic driver layer (LDL) applicable in a plug-in manner.

[0172] In particular, the present invention can perform the role of guiding a judgment direction appropriate to the situation in real time by operating as a plugin between the output of an artificial intelligence model and the final decision-making process. Through this, the artificial intelligence model to which the present invention is applied can implement high-level cognitive functions at the execution stage, such as dynamic changes in judgment strategies based on context, goal-based behavior guidance, and pre-detection and response to potential failures.

[0173] Specifically, the present invention adjusts the search probability by applying logical weights to the logits generated by the neural network through the Logic-Weighted Distribution for Symbolic Signal Kernel (LWD-SSK), thereby enabling a more explainable and user-intended action selection.

[0174] In addition, if structural errors or logical breakdowns are detected during the judgment process through the Reevaluation Mechanism (Re:eval), an alternative path different from the existing path is additionally generated and compared with the other; by reflecting the path that yielded the superior result, a highly reliable decision can be derived.

[0175] As such, the method proposed by the present invention can be easily applied without changing the structure of existing neural network models, and thus can provide the effect of significantly improving the consistency and stability of judgment quality, as well as the autonomy and resilience of the system, in various application environments.

[0176] The various embodiments of this document and the terms used therein are not intended to limit the technical features described in this document to specific embodiments, and should be understood to include various modifications, equivalents, or substitutions of said embodiments. In connection with the description of the drawings, similar reference numerals may be used for similar or related components. The singular form of a noun corresponding to an item may include one or more items unless the relevant context clearly indicates otherwise.

[0177] In this document, each of the phrases such as “A or B,” “at least one of A and B,” “at least one of A or B,” “A, B or C,” “at least one of A, B and C,” and “at least one of A, B, or C” may include all possible combinations of items listed together in the corresponding phrase. Terms such as “1,” “2,” or “first” or “second” may be used simply to distinguish a component from another component and do not limit the components in any other aspect (e.g., importance or order). Where any (e.g., 1st) component is referred to as “coupled” or “connected” to another (e.g., 2nd) component, with or without the terms “functionally” or “communicationly,” it means that the component may be connected to the other component directly (e.g., wired), wirelessly, or through a third component.

[0178] As used in this document, the term "module" may include a unit implemented in hardware, software, or firmware, and may be used interchangeably with terms such as logic, logic block, component, or circuit. A module may be a component formed integrally, or a minimum unit of a component or part thereof that performs one or more functions. For example, according to one embodiment, a module may be implemented in the form of an application-specific integrated circuit (ASIC).

[0179] Various embodiments of this document may be implemented as software (e.g., a program) comprising one or more instructions stored in a storage medium (e.g., memory) that can be read by a device (e.g., an electronic device). The storage medium may include random access memory (RAM), a memory buffer, a hard drive, a database, erasable programmable read-only memory (EPROM), electrically erasable read-only memory (EEPROM), read-only memory (ROM), and / or the like.

[0180] Additionally, the processor of the embodiments of this document may call at least one instruction among one or more instructions stored from a storage medium and execute it. This enables the device to operate to perform at least one function according to at least one called instruction. Such one or more instructions may include code generated by a compiler or code that can be executed by an interpreter. The processor may be a general-purpose processor, a Field Programmable Gate Array (FPGA), an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), and / or the like.

[0181] A device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-transitory' simply means that the storage medium is a tangible device and does not contain signals (e.g., electromagnetic waves), and this term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily.

[0182] Methods according to the various embodiments disclosed in this document may be provided as part of a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store (e.g., Play Store) or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product may be temporarily stored or temporarily created on a device-readable storage medium, such as a manufacturer's server, an application store's server, or the server's memory.

[0183] According to various embodiments, each component (e.g., module or program) of the described components may include a singular or multiple entities. According to various embodiments, one or more of the components or operations of the aforementioned components may be omitted, or one or more other components or operations may be added. Generally or additionally, multiple components (e.g., module or program) may be integrated into a single component. In this case, the integrated component may perform one or more functions of each of the components of the multiple components in the same or similar manner as they were performed by the corresponding component among the multiple components prior to integration. According to various embodiments, operations performed by a module, program, or other component may be executed sequentially, in parallel, iteratively, or heuristically; one or more of the operations may be executed in a different order; omitted; or one or more other operations may be added. Explanation of the symbols

[0184] 100: Device 110: Memory 120: Processor 130: Input / Output Interface 140: Communication interface

Claims

Claim 1 A method performed by an artificial intelligence control device operated by a processor comprises: acquiring a logit vector output from a predetermined artificial intelligence model; using a plurality of logic drivers that determine at least one logic event among logic collapse, policy-value mismatch, and instability of a prediction result of the artificial intelligence model, determining whether a decision path corresponding to an original judgment probability distribution generated for the logit vector has a possibility of failure according to a predefined criterion; if a possibility of failure is determined, generating a re-evaluation path from a higher node on an internal layer different from the decision path corresponding to the original judgment probability distribution within the decision structure of the artificial intelligence model, and deriving a re-evaluation probability distribution based on the re-evaluation path; comparing the original judgment probability distribution with the re-evaluation probability distribution derived by the re-evaluation path generated according to the determination result of the plurality of logic drivers, and selecting a probability distribution corresponding to a result with high reliability based on a predetermined reliability index as an output result of the artificial intelligence model. A method comprising an operation of deriving a final output of the artificial intelligence model based on the selected probability distribution, wherein the plurality of logic drivers include: a first logic driver that determines whether the logic of the logit value of the judgment candidate has collapsed based on a judgment criterion of relative logit value deviation or structural distribution distortion among the judgment candidates; a second logic driver that determines whether there is a policy-value discrepancy based on a judgment criterion of whether the output of the artificial intelligence model conforms to a predefined policy or objective function; and a third logic driver that determines whether there is instability of the logit value of the judgment candidate based on a judgment criterion of whether it deviates from a predefined probability range of occurrence based on output statistical information for each judgment candidate stored previously. Claim 2 A method according to claim 1, wherein the operation of determining whether the possibility of failure exists includes identifying a judgment candidate whose probability value among the original judgment probability distributions is greater than or equal to a predetermined threshold, and tracing the internal layers of the artificial intelligence model until the identified judgment candidate is derived to construct a path of a major node reaching the identified judgment candidate, thereby extracting the decision path. Claim 3 A method according to claim 1, characterized in that the plurality of logic drivers are configured to determine whether there is a structural abnormality in the decision path of the artificial intelligence model. Claim 4 A method according to claim 1, wherein each of the plurality of logic drivers performs a determination operation according to an independent judgment criterion, and the decision path determines whether to reconfigure based on each determination result for each of the plurality of logic drivers. Claim 5 A method according to claim 1, wherein the selection of the probability distribution is performed by reflecting the structural reliability of the decision path based on the determination results of the plurality of logic drivers. Claim 6 A method according to claim 1, wherein the operation of deriving the re-evaluation probability distribution includes the operation of determining the major judgment point that influenced the upper nodes within the decision-making structure of the artificial intelligence model. Claim 7 In claim 6, the operation of deriving the re-evaluation probability distribution includes tracking the key decision point through backpropagation-based contribution analysis or attention weight analysis to establish a re-evaluation path, and deriving the re-evaluation probability distribution based on the re-evaluation path. Claim 8 In claim 7, the operation of deriving the re-evaluation probability distribution includes the operation of setting the depth of the re-evaluation path based on the following mathematical formula 1, [Mathematical Formula 1] ( is the depth of the reevaluation path that determines how many steps prior to reevaluation in the decision path, is the magnitude of win rate collapse, indicating the degree to which the win rate drops sharply after the AI ​​model's decision, is the number of times the Principal Variation Rank changed while traversing the decision path, is the sum of the number of error detections detected by the logical driver, is the weighting coefficient for each term) method. Claim 9 A method according to claim 1, wherein the operation of selecting the decision includes the operation of calculating a reliability index based on the value of Shannon entropy or KL Divergence for each of the original judgment probability distribution and the re-evaluation probability distribution. Claim 10 The method comprises: a memory containing instructions; and a processor that performs a predetermined operation based on the instructions, wherein the operation of the processor includes: an operation of obtaining a logit vector output from a predetermined artificial intelligence model; an operation of determining whether there is a possibility of failure according to a predefined criterion for a decision path corresponding to an original judgment probability distribution generated for the logit vector, by using a plurality of logic drivers that determine at least one logic event among logic collapse, policy-value mismatch, and instability of the prediction result of the artificial intelligence model; an operation of, if a possibility of failure is determined, generating a re-evaluation path from a higher node on an internal layer different from the decision path corresponding to the original judgment probability distribution within the decision structure of the artificial intelligence model, and deriving a re-evaluation probability distribution based on the re-evaluation path; and an operation of comparing the original judgment probability distribution with the re-evaluation probability distribution derived by the re-evaluation path generated according to the determination result of the plurality of logic drivers, and selecting a probability distribution corresponding to a result with high reliability based on a predetermined reliability index as an output result of the artificial intelligence model. An artificial intelligence control device comprising an operation to derive a final output of the artificial intelligence model based on the selected probability distribution, wherein the plurality of logic drivers include: a first logic driver that determines whether the logic of the logit value of the judgment candidate has collapsed based on a judgment criterion of relative logit value deviation or structural distribution distortion among the judgment candidates; a second logic driver that determines whether there is a policy-value mismatch based on a judgment criterion of whether the output of the artificial intelligence model conforms to a predefined policy or objective function; and a third logic driver that determines whether there is instability of the logit value of the judgment candidate based on a judgment criterion of whether it deviates from a predefined probability range of occurrence based on previously stored output statistical information for each judgment candidate.

Images