Uncertainty-based real-label-free multi-consultant sequential reasoning method and device

Abstract

The invention provides a real-label-free multi-consultant sequential reasoning method and device based on uncertainty, and relates to the technical field of artificial intelligence machine learning and electronic data truth value inference. According to the method, the individual uncertainty of all consultants participating in decision making in the current question is calculated by constructing a question ordered set, and the average uncertainty is obtained; and then double-path probability calculation is executed, in a weighted voting path, the score probability of each candidate option is calculated, in a Bayesian aggregation path, the posterior probability of each candidate option is calculated, the average uncertainty is used as a scaling factor to carry out linear weighted fusion, and the candidate option with the highest fusion probability is selected as a final decision reasoning result. According to the method, synchronous evolution of a consultant evaluation system and decision accuracy can be driven through internal logic self-consistency in an environment which is completely lack of external truth value labeling, the problems of cold start robustness and malicious interference are effectively solved, and the calculation overhead is remarkably optimized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence machine learning and electronic data truth reasoning technology, and more specifically, to a method and apparatus for uncertainty-based, unlabeled, multi-advisor sequential reasoning. Background Technology

[0002] In the era of big data, electronic data is widely generated from various sensors, business systems, web logs, and user-generated content, characterized by its massive scale, diverse sources, and heterogeneous forms. However, data from different channels often contains inconsistent or even contradictory information, such as differences in descriptions of the same product on different platforms or anomalies between readings from multiple sensors. Accurately inferring information reflecting the true state from complex data filled with conflict and noise has become a core challenge in the field of truth inference. In practical applications such as loan approval, medical diagnosis, or IoT monitoring, decision-makers typically rely on the opinions of multiple advisors or experts for sequential decision-making, and often cannot obtain real-world data to verify the results in real time during the decision-making process.

[0003] Existing truth inference techniques are mainly divided into aggregation methods and confidence modeling methods. In aggregation methods, majority voting is based on the egalitarian assumption, assuming all data sources have the same reliability, making the decision easily dominated by low-quality or malicious data sources. While weighted voting distinguishes the importance of data sources through weights, its static weighting strategy cannot adapt to the dynamic drift of data source confidence, and the decision accuracy drops sharply when the weights themselves are unreliable. Bayesian aggregation methods heavily rely on accurate historical prior information, facing a severe cold-start problem when dealing with emerging data sources; if prior knowledge is unavailable or inaccurate, the model will struggle to converge. Regarding confidence modeling, traditional Beta distribution models have overly rigid initialization settings, failing to effectively quantify uncertainty and distinguish between different types of trust crises caused by data scarcity or data conflict. Matrix confidence methods face the curse of dimensionality when handling multi-option decisions; parameter estimation requires massive sample sizes, failing to meet the real-time processing needs of large-scale electronic data systems.

[0004] In summary, existing technologies have significant limitations in handling truth-based reasoning with multi-source heterogeneous data. These methods generally suffer from core flaws such as static decision assumptions, heavy reliance on real-label feedback, and insufficient policy adaptability. In sequential decision-making scenarios without any real labels, traditional methods cannot dynamically assess the evolution of consultant confidence and struggle to achieve robust trust initialization in the absence of prior information. Furthermore, existing research primarily targets binary decision problems, making it difficult to effectively extend to complex multi-option scenarios prevalent in the real world. Moreover, they lack sufficient system robustness when facing malicious consultant interference or low-quality data sources, resulting in poor long-term learning performance.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] The present invention aims to provide a method and apparatus for uncertain, unlabeled, multi-advisor sequential reasoning, in order to solve the defects of existing electronic data truth inference technology in practical applications, such as static decision assumptions, heavy reliance on real label feedback, rigid initialization settings, and insufficient policy adaptability.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] An uncertainty-based, unlabeled, multi-agent sequential reasoning method includes: S1, construct an ordered set of questions, obtain advisor suggestions for each question from a preset set of advisors, and define a set of candidate options containing several discrete candidate answers, with each advisor corresponding to a confidence level; the confidence level is used to represent the probability that the advisor gives the correct advice for the question; S2, calculate the individual uncertainty of all advisors involved in the decision-making process in the current problem, and obtain the average uncertainty by averaging the individual uncertainties; S3, perform dual-path probability calculation. In the weighted voting path, the advisor's current confidence level is used as the voting weight to calculate the score probability of each candidate option. In the Bayesian aggregation path, the prior probability distribution of the candidate options and the advisor's current confidence level are combined to calculate the posterior probability of each candidate option. S4. Using the average uncertainty as a scaling factor, the posterior probability and the score probability are linearly weighted and fused, and the candidate option with the highest fusion probability is selected as the final decision reasoning result.

[0009] Preferably, S1 specifically comprises: Define the ordered set of problems as ; For each problem in the ordered set of problems Define a set of discrete candidate answers, denoted as the candidate option set. Then the set of candidate options for all problems is denoted as . ; From the pre-set consultant set In each question Select a subset of advisors who provide advice, denoted as . ; For each consultant ,from Select a candidate option As their response answers, and the consultants' response answers do not overlap; , The total number of candidate options; Index for candidate options; Each consultant Corresponding to a confidence level Used to indicate consultant On the issue The probability of giving the correct advice; For each consultant Initial positive evidence With negative evidence ; Based on positive evidence With negative evidence Calculate the initial trust level for each consultant. Suspicion With uncertainty The expression is: ; ; in, This represents the set non-informative prior weights; Based on each consultant's initial trust level and uncertainty, the initial confidence level is calculated using the following expression: ; in, The confidence level is the base rate, representing the probability that the advisor will randomly select the correct option when there is no prior information.

[0010] Preferably, the formula for calculating the individual uncertainty is: ; in, For the question A subset of advisors involved in decision-making The average uncertainty is such that a larger value indicates a lower reliability of the consultant's confidence level. For the consultant subset The number of consultants; Consultant On the issue Uncertainty.

[0011] Preferably, the expression for the score probability of the candidate option is: ; in, For the question Candidate options The probability of scoring; For the question Candidate options A team of consultants; , Consultants , Confidence level; For the question A subset of advisors; The expression for the posterior probability of the candidate option is: ; ; in, For the question Candidate options The posterior probability; , , Candidate options , , The prior probability; The set of candidate options for all questions; For the question Answer set; For the question The true value; Given the true value The answer set was observed at that time. The likelihood probability; Given the true value The answer set was observed at that time. The likelihood probability; For the question Chinese advisor The actual option selected.

[0012] Preferably, the expression for the linear weighted fusion is: ; in, For the question Candidate options fusion probability; For the question A subset of advisors involved in decision-making The average uncertainty; The expression for the final decision reasoning result is: ; in, For the question The final decision-making reasoning result; Represents the maximum value operator; This is the set of candidate options for all questions.

[0013] Preferably, the method further includes: asynchronously updating the advisor's positive and negative evidence based on the difference between the fusion probability corresponding to the final decision reasoning result and the probability of the candidate option currently selected by the advisor, as expressed by: , ; , ; ; in, , Consultants Updated positive and negative evidence; Consultant Conservative update value at problem t; Candidate options In the question t Conservative update value at that time; , , The questions are as follows Candidate options , , The fusion probability.

[0014] Preferably, it further includes: calculating an updated consultant confidence level based on the updated positive and negative evidence, and feeding it back to the next round of reasoning and decision-making, until all issues have been reasoned and decided.

[0015] The present invention also provides an uncertainty-based, unlabeled, multi-agent sequential reasoning device, comprising: The question ordered set construction unit is used to construct a question ordered set, obtain advisor suggestions for each question from a preset advisor set, and define a candidate option set containing several discrete candidate answers. Each advisor corresponds to a confidence level; the confidence level is used to represent the probability that the advisor gives the correct advice for the question. The uncertainty calculation unit is used to calculate the individual uncertainty of all consultants involved in the decision-making process in the current problem, and obtains the average uncertainty by averaging the individual uncertainties. The dual-path probability calculation unit is used to perform dual-path probability calculation. In the weighted voting path, the current confidence level of the consultant is used as the voting weight to calculate the score probability of each candidate option. In the Bayesian aggregation path, the prior probability distribution of the candidate options and the current confidence level of the consultant are combined to calculate the posterior probability of each candidate option. The weighted fusion unit is used to use the average uncertainty as a scaling factor to linearly weight and fuse the posterior probability and the score probability, and select the candidate option with the highest fusion probability as the final decision reasoning result.

[0016] The reasoning device also includes: The evidence update unit is used to asynchronously update the advisor's positive and negative evidence based on the difference between the fusion probability corresponding to the final decision reasoning result and the probability of the candidate option currently selected by the advisor. The confidence update unit is used to calculate the updated consultant confidence based on the updated positive and negative evidence, and feed it back to the next round of reasoning and decision-making until all issues have been reasoned and decided.

[0017] The present invention also provides an uncertainty-based, unlabeled, multi-advisor sequential reasoning device, including a processor and a memory, wherein the memory stores a computer program that can be executed by the processor to implement the uncertainty-based, unlabeled, multi-advisor sequential reasoning method described above.

[0018] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor of the device on which the computer-readable storage medium resides, implement an uncertainty-based, unlabeled, multi-advisor sequential reasoning method as described above.

[0019] In summary, compared with the prior art, the present invention has the following beneficial effects: First, this invention achieves high-precision sequential reasoning decision-making in truth-deficient environments. By introducing an integrated decision-making model combining Bayesian and weighted voting, this scheme utilizes average uncertainty as a dynamic balancer. In the early stages of system operation, when confidence levels are unreliable, a weighted voting mechanism ensures basic decision stability. Later, when evidence is sufficient and uncertainty decreases, it automatically switches to a Bayesian aggregation mode to achieve higher inference accuracy. This adaptive switching mechanism enables the system to maintain optimized decision-making performance throughout its entire lifecycle, overcoming the shortcomings of traditional single-strategy approaches that exhibit poor adaptability at different stages.

[0020] Second, this invention provides accurate dynamic assessment of consultant capabilities and robust trust management. This solution employs a conservative confidence strategy, utilizing differences in decision probabilities rather than binary correct or incorrect feedback to update consultant evidence, effectively filtering out interference from noisy data and initial erroneous decisions. Under this mechanism, consultant positive or negative evidence is only significantly updated when the decision outcome has a significant statistical advantage. This maintains the authenticity and stability of consultant capability ranking during unsupervised learning, significantly reducing the damage to the system's trust structure caused by malicious consultants or abnormal data sources.

[0021] Third, this invention significantly optimizes computational resource utilization and system response speed. This solution abandons complex historical data backtracking and high-dimensional probability matrix operations, instead employing a simplified parameterized representation based on subjective logic. Since each advisor only needs to maintain a small number of scalar parameters, the system exhibits extremely low memory usage when handling large-scale multi-option decision-making tasks. Simultaneously, because the core algorithm logic is based on closed-form algebraic operations rather than a cumbersome iterative optimization process, the system demonstrates excellent real-time processing performance when handling high-frequency electronic data streams, reducing computation time by more than 80% compared to traditional Bayesian inference methods.

[0022] Fourth, this invention possesses excellent engineering scalability and applicability to multiple scenarios. This solution, through generalized probabilistic modeling, extends the decision-making logic from simple binary classification to the multi-option decision-making domain without increasing the algorithm's complexity order. This enables the method to be widely applied to various complex business scenarios lacking immediate feedback, such as financial credit assessment, distributed sensor network data fusion, medical multi-expert consultation systems, and truth-based reasoning on crowdsourcing platforms. It provides a general, self-evolving underlying technical framework for various intelligent decision support systems.

[0023] Fifth, this invention effectively solves the problem of trust rigidity in the cold start phase. By setting a base rate in subjective logic and explicitly quantifying uncertainty, this solution can accurately express the state of "trust not yet established" at the beginning of system operation, rather than simply assigning a fixed neutral probability. This precise modeling of the "unknown," combined with dynamic integrated decision-making, ensures that the system can still make robust decisions through collective wisdom even in an environment with zero prior information, and quickly and smoothly establish a reliable advisor evaluation system as data flows in. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating an uncertainty-based, unlabeled, multi-advisor sequential reasoning method provided in Example 1.

[0026] Figure 2 This is a schematic diagram of the overall architecture of the uncertainty-based, unlabeled, multi-advisor sequential decision method (MASDM) provided in Example 1.

[0027] Figure 3 The diagram shows the strategy switching logic and performance comparison of the decision model provided in Example 1 under different confidence levels.

[0028] Figure 4 This is a schematic diagram of an uncertainty-based, unlabeled, multi-advisor sequential reasoning device provided in Embodiment 2.

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0031] Example 1 Embodiment 1 of the present invention provides an uncertainty-based, unlabeled, multi-advisor sequential inference method, which can be implemented by an uncertainty-based, unlabeled, multi-advisor sequential inference device (hereinafter referred to as inference device), specifically, executed by one or more processors within the inference device.

[0032] In this embodiment, the inference device may be an electronic device equipped with a processor, the processor having a computer program for the uncertainty-based, unlabeled, multi-advisor sequential inference method and the computer program being executable, such as a computer, smartphone, smart tablet, workstation, etc., without limitation.

[0033] In this embodiment, "advisor" refers to a non-standard answer source that can provide candidate options for a question without actual annotation. It is an advisor, not an annotator with actual annotation capabilities.

[0034] Single-choice decision problems are a common type of decision-making where advisors provide advice by offering choices among multiple options. For example, a bank decides whether to approve a loan application, or a factory decides whether to produce a product from multiple options. In this scenario, we have a series of questions. For each question, the decision-maker must make a decision based on the advisor's advice. The decision-maker may receive conflicting advice from multiple advisors.

[0035] like Figures 1-2 As shown, an uncertainty-based, unlabeled, multi-advisor sequential reasoning method includes steps S1 to S4.

[0036] S1, construct an ordered set of questions, obtain advisor suggestions for each question from a preset set of advisors, and define a set of candidate options containing several discrete candidate answers. Each advisor corresponds to a confidence level; the confidence level is used to represent the probability that the advisor gives the correct advice for the question.

[0037] like Figure 2 As shown, the present invention comprises a closed-loop system consisting of a decision model and a confidence model. During operation, it does not rely on externally provided real-time labels, but rather drives the iterative iteration of the decision model and confidence model through internal logical consistency. The entire system is deployed on an electronic computing platform with data processing capabilities to handle electronic data streams from heterogeneous data sources.

[0038] In the initial stage of system startup, an ordered set of problems faced by decision-makers is first constructed. The ordered set of problems is defined as follows: It includes several independent decision-making tasks arranged in order of time step.

[0039] For each problem in the ordered set of problems Define a set of discrete candidate answers, denoted as the candidate option set. Then the set of candidate options for all problems is denoted as . .

[0040] Then from the pre-set set of consultants In each question Select a subset of advisors who provide advice, denoted as . .

[0041] The consultant group here This includes, but is not limited to, physical sensors, automated classification algorithms, expert systems, or human evaluators.

[0042] For each consultant ,from Select a candidate option As their response answers, and the consultants' response answers do not overlap; , The total number of candidate options; Index the candidate options. The system automatically records each consultant's response to the current question. The specific options selected form the answer set for the current question. , ,satisfy , and That is, the response answers do not overlap.

[0043] During the initialization process, each consultant Corresponding to a confidence level Used to indicate consultant On the issue The probability of giving the correct advice.

[0044] Because there was a lack of historical evidence regarding consultant performance during the cold start phase, the system used a subjective logical framework for parameter setting. Specifically, for each consultant... Initial positive evidence With negative evidence For example, all of them are initialized to 0.

[0045] Based on positive evidence With negative evidence Calculate the initial trust level for each consultant. Suspicion With uncertainty The expression is: ; ; Based on each consultant's initial trust level and uncertainty, calculate the initial confidence level and construct a confidence level model, expressed as follows: ; in, The non-informative prior weight is set and can be set to a fixed value of 2 as a strength parameter of the prior distribution, used to smooth out the sharp fluctuations in confidence in the initial stage when evidence is scarce. The confidence level is the base rate, representing the probability that the advisor will randomly select the correct option when there is no prior information.

[0046] If the model has no prior knowledge about the distribution of consultant confidence, then It can be set to For example, in a question with four options, the basic rate Setting it to 0.25 indicates that the advisor... The probability of providing correct advice in a randomly selected scenario is 0.25. This method ensures that, in the absence of any prior information, the confidence model does not over- or under-trust any advisor, thus eliminating systematic bias in the initialization phase. Of course, with reasonable prior knowledge, an appropriate base rate can be set accordingly.

[0047] S2 calculates the individual uncertainty of all advisors involved in the decision-making process in the current problem, and obtains the average uncertainty by arithmetically averaging the individual uncertainties.

[0048] The formula for calculating the individual uncertainty is: ; in, For the question A subset of advisors involved in decision-making The average uncertainty is such that a larger value indicates a lower reliability of the consultant's confidence level. For the consultant subset The number of consultants; Consultant On the issue Uncertainty.

[0049] This step involves calculating the average uncertainty by taking the arithmetic mean of the individual uncertainties of all participating advisors. This parameter, as a core factor in dynamically adjusting decision weights, reflects the reliability of the current confidence assessment.

[0050] S3 performs a dual-path probability calculation. In the weighted voting path, the advisor's current confidence level is used as the voting weight to calculate the score probability of each candidate option. In the Bayesian aggregation path, the prior probability distribution of the candidate options and the advisor's current confidence level are combined to calculate the posterior probability of each candidate option.

[0051] This step is... Figure 2 The decision model in this paper is an integration of Bayesian and weighted voting. This module is responsible for aggregating answers and making decisions based on the current advisor's confidence level when no truth value is available. Its core innovation lies in adaptively selecting the aggregation strategy based on the uncertainty of the advisor's confidence level.

[0052] When the system is initialized or evidence is insufficient, the uncertainty of the consultant's confidence level is high. In this case, the framework prioritizes a weighted voting method, using the consultant's confidence level as the voting weight to calculate the weighted score for each option.

[0053] In the weighted voting path, the expression for the score probability of the candidate option is: ; in, For the question Candidate options The probability of scoring; For the question Candidate options A team of consultants; , Consultants , Confidence level; For the question A subset of advisors; It is a selection option. The consultant index, and This is the advisor index for selecting other options.

[0054] This path occurs in the early stages of system operation, i.e., during the period of average uncertainty. At higher levels, they dominate, leveraging the redundancy of collective wisdom to hedge against the risk of inaccurate assessments of individual advisors' capabilities.

[0055] As evidence accumulates, the consultant's confidence level becomes more reliable, uncertainty decreases, and the system gradually tilts towards Bayesian aggregation.

[0056] In the Bayesian aggregation path, the expression for the posterior probability of the candidate option is: ; ; in, For the question Candidate options The posterior probability; , , Candidate options , , The prior probability; The set of candidate options for all questions; For the question Answer set; For the question The true value; Given the true value The answer set was observed at that time. The likelihood probability; Given the true value The answer set was observed at that time. The likelihood probability; For the question Chinese advisor The actual option selected.

[0057] When the problem When the prior probability is assumed to be a uniform distribution, it can be obtained statistically from the probability distribution of the decision result of the previous step in the sequential reasoning process.

[0058] When dealing with multi-option scenarios, this invention employs an extended probability model, assuming that when an advisor makes an incorrect choice, their error probability will be distributed to all other candidate options (excluding their chosen option) according to the prior distribution of each candidate option. This refined probability allocation logic allows the system to accurately characterize the contribution of different error types to truth inference even when there are many options.

[0059] S4. Using the average uncertainty as a scaling factor, the posterior probability and the score probability are linearly weighted and fused, and the candidate option with the highest fusion probability is selected as the final decision reasoning result.

[0060] The expression for the linear weighted fusion is: ; in, For the question Candidate options fusion probability; For the question A subset of advisors involved in decision-making The average uncertainty; In this way, a smooth transition is achieved from the initial "majority rule" to the later "probabilistic calculation". This invention selects the candidate option with the highest fusion probability as the final decision reasoning result for the current problem.

[0061] The expression for the final decision reasoning result is: ; in, For the question The final decision-making reasoning result; Represents the maximum value operator; This is the set of candidate options for all questions.

[0062] After completing the decision output, the system enters the confidence update stage, which in a preferred embodiment further includes steps S5 and S6.

[0063] S5. Based on the difference between the fusion probability corresponding to the final decision reasoning result and the probability of the candidate option currently selected by the advisor, the advisor's positive evidence and negative evidence are updated asynchronously.

[0064] Due to the lack of real-world annotations, this invention achieves robust iteration of confidence by calculating evidence values ​​for decision inference.

[0065] First, the difference between the decision probability of the candidate option and the expected probability of all remaining options distributed according to their probability weights is calculated. This difference precisely quantifies the prominence of the current advisor's recommendation among the competing options. If the system determines that the option chosen by the advisor is highly likely to be correct, the resulting update value is positive; otherwise, it is negative.

[0066] Then, asynchronous evidence updates are performed based on the sign of the conservative update value. That is, positive / negative evidence is updated based on the sign of the update value. When the update value is positive, only positive evidence is updated, and when the update value is negative, only negative evidence is updated. The two are not triggered at the same time.

[0067] The update expressions for positive and negative evidence are: , ; , ; ; in, , Consultants Updated positive and negative evidence; Consultant Conservative update value at problem t; Candidate options The conservative update value at problem t, i.e., the difference value; , , The questions are as follows Candidate options , , The fusion probability.

[0068] When the decision itself has high uncertainty (i.e., the probabilities of each option are close), the generated update value is small, and the system automatically slows down the adjustment of the assessment of the consultant's ability, thereby effectively filtering out the interference from noisy data and initial erroneous decisions in the unsupervised learning process.

[0069] S6 calculates the updated consultant confidence level based on the updated positive and negative evidence, and feeds it back to the next round of reasoning and decision-making until all issues are reasoned and decided.

[0070] The advisor confidence update expression is: ; ; in, , , Consultants question Initial level of trust, skepticism, and uncertainty; The updated advisor confidence level.

[0071] The system feeds these updated parameters back to the decision-making stage of the next question. Through this closed-loop feedback architecture of "updating the basis for decision-making and updating and optimizing decision weights," the system achieves synchronous evolution of the consultant evaluation system and decision accuracy while processing continuous electronic data streams.

[0072] Figure 2 This paper provides an overview of how the Multi-Advisor Sequential Decision-Making (MASDM) method of this invention works. MASDM is designed with two parts. The first part is a confidence model. More specifically, the confidence model builds the advisors' confidence over time. The second part is a decision model, which makes the optimal decision based on the advisors' recommendations and their current confidence levels. In particular, the confidence model and the decision model are interdependent. For each problem, the decision model requires the advisors' current confidence levels to make a decision, while updating the advisors' confidence levels is based on evidence inferred from the decisions made. Therefore, both models in MASDM improve over time without requiring fundamental truths about the problem or prior information about the advisors' confidence levels.

[0073] Large-scale simulation experiments and validation on the MNIST dataset demonstrate that the MASDM method significantly outperforms existing methods in terms of decision accuracy, cold-start robustness, and dynamic environment adaptability. It effectively solves key technical challenges such as sequential decision-making without real labels, dynamic policy adaptation, and rigidity of trust initialization, providing an innovative solution for multi-advisor decision-making systems.

[0074] In the early stages of system operation, due to the limited accumulation of evidence parameters by each consultant, the average uncertainty was low. It is at a relatively high level (close to 1). At this point, the fusion probability... This is primarily influenced by the weighted voting path. As the number of issues processed increases, the positive or negative performance of consultants is continuously solidified into evidentiary parameters through conservative update values, leading to a gradual decrease in individual uncertainty. When the average uncertainty drops to a low level, the system automatically shifts its decision focus towards the Bayesian aggregation path.

[0075] In specific practical applications, such as data fusion tasks in distributed sensor networks, this invention achieves real-time truth inference of readings from multiple heterogeneous sensors by executing the above steps. When a sensor's reading deviates from the consensus of most sensors due to aging or environmental interference, the system uses a conservative update mechanism to accumulate the difference in its negative evidence, thereby automatically reducing its weight in subsequent decisions. Since the system only needs to store two evidence parameters for each advisor and the current base rate, without maintaining a large historical decision matrix, it exhibits extremely low memory usage and excellent real-time processing performance when handling large-scale sensor networks.

[0076] In another application scenario, such as expert consultation systems within crowdsourcing platforms, this invention automatically builds expert competency profiles by executing sequential decision-making logic, without requiring manual verification of expert answers. Through the base case setting in the subjective logic, the system can still make robust decisions based on collective wisdom even in environments with zero prior information. As the task progresses, the system tracks the dynamic changes in expert competence in real time through a closed-loop feedback mechanism. For example, in scenarios where increased expert experience leads to improved accuracy, the system continuously accumulates positive evidence to achieve real-time optimization of expert confidence levels, maintaining the long-term reliability of the decision-making system.

[0077] Furthermore, this invention employs an incremental update mechanism when processing electronic data streams. Because the core logic of the algorithm is based on closed-form algebraic operations rather than a cumbersome iterative optimization process, the system exhibits extremely high response speed when processing high-frequency data. In image classification experiments simulating the MNIST dataset, the system treats different classifiers as advisors and performs sequential decision-making while completely masking the true labels. Results show that the MASDM method, through dynamic ensemble decision-making, achieves a classification accuracy approaching the upper limit of supervised learning, and its computation time is reduced by more than 80% compared to traditional Bayesian inference methods.

[0078] In another preferred embodiment, comparative experiments were conducted on the Bayesian aggregation method (BYS) and the weighted voting method (WV) to explain the motivation behind the dual-path strategy design in the method of the present invention.

[0079] Specifically, we examined the performance of the Bayesian aggregation method (BYS) and the weighted voting method (WV) at different confidence levels under 24 conditions. To launch the experiment, we created a set of simulated advisors. Each advisor was assigned a true accuracy, sampled from an extended rectified Gaussian distribution (ERGd). The reason for using ERGd is that different ERGd means can be set. and ERGd standard deviation To simulate different distributions of consultants' actual accuracy rates, in order to comprehensively evaluate our method. Furthermore, and Indicates the information used to generate accuracy. The mean and standard deviation of the ERGd distribution.

[0080] For example, if This indicates that the consultant There is an 80% probability of choosing the correct option, while the consultant... There is a 20% chance that one of the options will be randomly selected incorrectly.

[0081] In addition, for each consultant We also generate confidence levels for practical use using Bayesian aggregation and weighted voting methods. .like Figure 3 As shown, let This represents the distance coefficient, indicating how closely the consultant's confidence level approximates the true accuracy rate, ranging from 0 to 100. Additionally, we randomly generate a probability variable. .

[0082] Confidence The calculation method is as follows: ; in, For true confidence level; along with The increase, From a random value Gradually approaching true accuracy This simulates the transition of the confidence model from unreliable to reliable. Furthermore, in... For each stage, from 0 to 100 (increasing by 1 each time), the consultant needs to aggregate 10,000 questions using the Bayesian aggregation method (BYS) and the weighted voting method (WV). Additionally, the number of correct decisions is the result of the experiment.

[0083] Figure 3 The results of comparative experiments on the Bayesian aggregation method (BYS) and the weighted voting method (WV) under different conditions are presented. Figure 3 The X-axis represents the distance coefficient. The value ranges from 0 to 100, while the Y-axis displays the average accuracy observed for these two methods on 10,000 questions. Figure 3 The curve in the figure includes a semi-transparent area, which represents the error band of the 95% confidence interval. This represents the total number of candidate options.

[0084] In addition, we use Let represent the average number of consultants per question in each experiment. For each question, we randomly selected an average of 14 consultants from all consultants. Figure 3The results of four sets of comparative experiments are presented. The following results show that when When the size is very small, the weighted voting method (WV) generally outperforms the Bayesian aggregation method (BYS), although the advantage of the weighted voting method is not significant in some cases. However, as... As the confidence level increases, the Bayesian aggregation method (BYS) demonstrates superior performance compared to the weighted voting method (WV). Therefore, we combine the Bayesian aggregation method and the weighted voting method to accommodate different levels of advisor confidence and improve overall aggregation performance. Experimental data show that this adaptive switching logic enables the system to maintain optimized decision-making performance at different confidence and reliability stages, and its long-term average accuracy is significantly better than any single fixed strategy.

[0085] In summary, compared with the prior art, the present invention has the following beneficial effects: This invention constructs a deeply coupled architecture of decision-making and confidence models, quantifies uncertainty using subjective logic, and employs a conservative update strategy to build a highly efficient, robust, and self-learning decision-making system without requiring any manual annotation costs. This scheme systematically solves core technical challenges such as cold-start robustness, dynamic policy adaptation, and unsupervised continuous learning through specific probabilistic path fusion logic and asynchronous evidence update algorithms, providing a complete technical path for solving conflict inference and truth discovery problems in complex big data environments. The method is not dependent on a specific hardware platform; through real-time streaming processing of electronic data, it provides decision-makers with high-confidence option recommendations and simultaneously outputs reliability reports from various advisors, providing accurate data support for subsequent resource scheduling or expert selection.

[0086] Example 2 like Figure 4 As shown, the second embodiment of the present invention also provides an uncertainty-based, unlabeled, multi-advisor sequential reasoning device, comprising: The question ordered set construction unit is used to construct a question ordered set, obtain advisor suggestions for each question from a preset advisor set, and define a candidate option set containing several discrete candidate answers. Each advisor corresponds to a confidence level; the confidence level is used to represent the probability that the advisor gives the correct advice for the question. The uncertainty calculation unit is used to calculate the individual uncertainty of all consultants involved in the decision-making process in the current problem, and obtains the average uncertainty by averaging the individual uncertainties. The dual-path probability calculation unit is used to perform dual-path probability calculation. In the weighted voting path, the current confidence level of the consultant is used as the voting weight to calculate the score probability of each candidate option. In the Bayesian aggregation path, the prior probability distribution of the candidate options and the current confidence level of the consultant are combined to calculate the posterior probability of each candidate option. The weighted fusion unit is used to use the average uncertainty as a scaling factor to linearly weight and fuse the posterior probability and the score probability, and select the candidate option with the highest fusion probability as the final decision reasoning result.

[0087] Also includes: The evidence update unit is used to asynchronously update the advisor's positive and negative evidence based on the difference between the fusion probability corresponding to the final decision reasoning result and the probability of the candidate option currently selected by the advisor. The confidence update unit is used to calculate the updated consultant confidence based on the updated positive and negative evidence, and feed it back to the next round of reasoning and decision-making until all issues have been reasoned and decided.

[0088] Example 3 The third embodiment of the present invention also provides an uncertainty-based, unlabeled, multi-advisor sequential reasoning device, which includes a memory and a processor. The memory stores a computer program that can be executed by the processor to implement the uncertainty-based, unlabeled, multi-advisor sequential reasoning method described above.

[0089] Example 4 The fourth embodiment of the present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, they implement the above-described uncertainty-based, unlabeled, multi-advisor sequential inference method.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for unlabeled, multi-agent sequential reasoning based on uncertainty, characterized in that: include: S1, construct an ordered set of questions, obtain advisor suggestions for each question from a preset set of advisors, and define a set of candidate options containing several discrete candidate answers, with each advisor corresponding to a confidence level; the confidence level is used to represent the probability that the advisor gives the correct advice for the question; S2, calculate the individual uncertainty of all advisors involved in the decision-making process in the current problem, and obtain the average uncertainty by averaging the individual uncertainties; S3, perform dual-path probability calculation. In the weighted voting path, the advisor's current confidence level is used as the voting weight to calculate the score probability of each candidate option. In the Bayesian aggregation path, the prior probability distribution of the candidate options and the advisor's current confidence level are combined to calculate the posterior probability of each candidate option. S4. Using the average uncertainty as a scaling factor, the posterior probability and the score probability are linearly weighted and fused, and the candidate option with the highest fusion probability is selected as the final decision reasoning result.

2. The uncertainty-based, unlabeled, multi-agent sequential reasoning method according to claim 1, characterized in that... S1 specifically refers to: Define the ordered set of problems as ; For each problem in the ordered set of problems Define a set of discrete candidate answers, denoted as the candidate option set. Then the set of candidate options for all problems is denoted as . ; From the pre-set consultant set In each question Select a subset of advisors who provide advice, denoted as . ; For each consultant ,from Select a candidate option As their response answers, and the consultants' response answers do not overlap; , The total number of candidate options; Index for candidate options; Each consultant Corresponding to a confidence level Used to indicate consultant On the issue The probability of giving the correct advice; For each consultant Initial positive evidence With negative evidence ; Based on positive evidence With negative evidence Calculate the initial trust level for each consultant. Suspicion With uncertainty The expression is: ； ； in, This represents the set non-informative prior weights; Based on each consultant's initial trust level and uncertainty, the initial confidence level is calculated using the following expression: ； in, The confidence level is the base rate, representing the probability that the advisor will randomly select the correct option when there is no prior information.

3. The uncertainty-based, unlabeled, multi-agent sequential reasoning method according to claim 2, characterized in that... The formula for calculating the individual uncertainty is: ； in, For the question A subset of advisors involved in decision-making The average uncertainty is such that a larger value indicates a lower reliability of the consultant's confidence level. For the consultant subset The number of consultants; Consultant On the issue Uncertainty.

4. The uncertainty-based, unlabeled, multi-agent sequential reasoning method according to claim 2, characterized in that... The expression for the score probability of the candidate option is: ； in, For the question Candidate options The probability of scoring; For the question Candidate options A team of consultants; , Consultants , Confidence level; For the question A subset of advisors; The expression for the posterior probability of the candidate option is: ； ； in, For the question Candidate options The posterior probability; , , Candidate options , , The prior probability; The set of candidate options for all questions; For the question Answer set; For the question The true value; Given the true value The answer set was observed at that time. The likelihood probability; Given the true value The answer set was observed at that time. The likelihood probability; For the question Chinese advisor The actual option selected.

5. The uncertainty-based, unlabeled, multi-agent sequential reasoning method according to claim 4, characterized in that... The expression for the linear weighted fusion is: ； in, For the question Candidate options fusion probability; For the question A subset of advisors involved in decision-making The average uncertainty; The expression for the final decision reasoning result is: ； in, For the question The final decision-making reasoning result; Represents the maximum value operator; This is the set of candidate options for all questions.

6. The uncertainty-based, unlabeled, multi-agent sequential reasoning method according to claim 2, characterized in that... It also includes: based on the difference between the fusion probability corresponding to the final decision reasoning result and the probability of the candidate option currently selected by the advisor, asynchronously updating the advisor's positive and negative evidence, expressed as: , ； , ； ； in, , Consultants Updated positive and negative evidence; Consultant Conservative update value at problem t; Candidate options Conservative update value at problem t; , , The questions are as follows Candidate options , , The fusion probability.

7. The uncertainty-based, unlabeled, multi-agent sequential reasoning method according to claim 6, characterized in that... It also includes: calculating updated consultant confidence based on updated positive and negative evidence, and feeding it back to the next round of reasoning and decision-making, until all issues have been reasoned and decided.

8. An uncertainty-based, unlabeled, multi-advisor sequential reasoning apparatus for implementing the uncertainty-based, unlabeled, multi-advisor sequential reasoning method as described in any one of claims 1-7, characterized in that, include: The question ordered set construction unit is used to construct a question ordered set, obtain advisor suggestions for each question from a preset advisor set, and define a candidate option set containing several discrete candidate answers. Each advisor corresponds to a confidence level; the confidence level is used to represent the probability that the advisor gives the correct advice for the question. The uncertainty calculation unit is used to calculate the individual uncertainty of all consultants involved in the decision-making process in the current problem, and obtains the average uncertainty by averaging the individual uncertainties. The dual-path probability calculation unit is used to perform dual-path probability calculation. In the weighted voting path, the current confidence level of the consultant is used as the voting weight to calculate the score probability of each candidate option. In the Bayesian aggregation path, the prior probability distribution of the candidate options and the current confidence level of the consultant are combined to calculate the posterior probability of each candidate option. The weighted fusion unit is used to use the average uncertainty as a scaling factor to linearly weight and fuse the posterior probability and the score probability, and select the candidate option with the highest fusion probability as the final decision reasoning result.

9. A multi-advisor sequential reasoning device based on uncertainty and without real-world annotations according to claim 8, characterized in that... It also includes: The evidence update unit is used to asynchronously update the advisor's positive and negative evidence based on the difference between the fusion probability corresponding to the final decision reasoning result and the probability of the candidate option currently selected by the advisor.

10. The uncertainty-based, unlabeled, multi-advisor sequential reasoning device according to claim 9, characterized in that... It also includes: The confidence update unit is used to calculate the updated consultant confidence based on the updated positive and negative evidence, and feed it back to the next round of reasoning and decision-making until all issues have been reasoned and decided.

Images