A multi-agent system-based experiment scheme generation method and device

Abstract

The application provides a kind of experimental scheme generation method and device based on multi-agent system, it is related to experimental design technical field.The method comprises: using hypothesis generation agent to obtain context information corresponding to scientific research problem, and generate a plurality of target candidate hypotheses for solving scientific research problem according to context information;Using judge agent to compare a plurality of target candidate hypotheses by swiss round, and according to the comparison result, a plurality of target candidate hypotheses are screened to obtain first preferred hypothesis;Using review agent to analyze the reason of comparison result, and obtain comprehensive attribution result;Using hypothesis evolution agent to rewrite first preferred hypothesis according to comprehensive attribution result, and obtain target preferred hypothesis;The target preferred hypothesis is input to multi-scheme generation agent, and the target experimental scheme corresponding to the target preferred hypothesis is obtained.The method can effectively improve the generation efficiency and quality of experimental scheme.

Description

Technical Field

[0001] This invention relates to the field of experimental design technology, and in particular to a method and apparatus for generating experimental schemes based on a multi-agent system. Background Technology

[0002] In scientific research, researchers typically need to conduct open-ended experimental studies around specific research problems in order to optimize a complete experimental process that can solve those problems. With the groundbreaking advancements in artificial intelligence technology in recent years, some studies have utilized multi-agent systems for experimental design.

[0003] However, current experimental design methods are limited to using intelligent agents to generate a large number of experimental schemes or scientific hypotheses in an open manner, and then selecting the target experimental scheme or scientific hypothesis through various indicators such as feasibility, rationality, and potential risks. As the complexity of scientific research problems continues to increase, this simple generation-then-selection model cannot meet the current research needs in terms of efficiency and quality. There is an urgent need for a generation method that can quickly generate high-quality experimental schemes within a limited time. Summary of the Invention

[0004] This invention provides a method and apparatus for generating experimental schemes based on a multi-agent system, in order to solve the problems of low efficiency and poor quality in the generation of experimental schemes in the prior art.

[0005] This invention provides a method for generating experimental schemes based on a multi-agent system, comprising the following steps: using a hypothesis generating agent to obtain contextual information corresponding to a research problem, and generating multiple target candidate hypotheses for solving the research problem based on the contextual information; using a referee agent to perform a Swiss-system comparison of the multiple target candidate hypotheses, and filtering the multiple target candidate hypotheses based on the comparison results to obtain a first preferred hypothesis; using a review agent to perform causal analysis on the comparison results to obtain a comprehensive attribution result; using a hypothesis evolution agent to rewrite the first preferred hypothesis based on the comprehensive attribution result to obtain a target preferred hypothesis; and inputting the target preferred hypothesis into a multi-scheme generating agent to obtain a target experimental scheme corresponding to the target preferred hypothesis.

[0006] Optionally, the step of using a hypothesis generation agent to obtain contextual information corresponding to a research problem and generating multiple target candidate hypotheses for solving the research problem based on the contextual information includes: using a hypothesis generation agent to obtain contextual information corresponding to a research problem and generating multiple candidate hypotheses for solving the research problem based on the contextual information; using a hypothesis analysis agent to perform hypothesis analysis on each of the candidate hypotheses and determining a target candidate hypothesis from the multiple candidate hypotheses based on the hypothesis analysis results.

[0007] Optionally, the hypothesis analysis includes at least one of semantic analysis, mechanism analysis, and cluster analysis; the step of using a hypothesis analysis agent to perform hypothesis analysis on each of the candidate hypotheses, and determining the target candidate hypothesis from the multiple candidate hypotheses based on the hypothesis analysis results, includes: When the hypothesis analysis includes semantic analysis, multiple candidate hypotheses with semantic similarity higher than a preset similarity are merged to form the target candidate hypothesis; When the hypothesis analysis includes mechanism of action analysis, multiple candidate hypotheses with the same mechanism of action are merged to form the target candidate hypothesis. In the case where the hypothesis analysis includes cluster analysis, multiple candidate hypotheses belonging to the same cluster category are merged to form the target candidate hypothesis.

[0008] Optionally, the step of using a referee agent to perform a Swiss-system comparison of multiple target candidate hypotheses and filtering the multiple target candidate hypotheses according to the comparison results to obtain a first preferred hypothesis includes: performing pairwise comparisons of the target candidate hypotheses in multiple rounds using a Swiss-system comparison method, obtaining the target integral corresponding to each target candidate hypothesis according to the comparison results; and filtering the multiple target candidate hypotheses according to the target integral to obtain a first preferred hypothesis.

[0009] Optionally, the comparison result includes group comparison results corresponding to at least one round; the step of using a Swiss-system comparison method to perform pairwise comparisons of the target candidate hypotheses for multiple rounds to obtain the target score corresponding to each target candidate hypothesis includes: cyclically executing the following steps until the number of rounds reaches a preset number: obtaining the hypothesis score corresponding to each target candidate hypothesis; determining multiple hypothesis groups corresponding to the round based on the closeness between the hypothesis scores and the number of target candidate hypotheses; wherein each hypothesis group includes at most two target candidate hypotheses; evaluating one or two target candidate hypotheses in each hypothesis group according to a preset evaluation index to generate a group comparison result corresponding to each hypothesis group; updating the hypothesis score corresponding to each target candidate hypothesis based on the group comparison results corresponding to each hypothesis group, and using the updated hypothesis score as the target score corresponding to each target candidate hypothesis.

[0010] Optionally, the hypothesis integral is a Glicko-2 score; updating the hypothesis integral corresponding to each target candidate hypothesis based on the group comparison results corresponding to each hypothesis group includes: obtaining the uncertainty corresponding to each target candidate hypothesis; updating the Glicko-2 score based on the group comparison results corresponding to each hypothesis group and the uncertainty.

[0011] Optionally, the step of using a review agent to perform causal analysis on the comparison results to obtain a comprehensive attribution result includes: using a review agent to obtain the group comparison results corresponding to each of the rounds; performing causal analysis on each of the group comparison results to obtain the group attribution results corresponding to each of the hypothetical groups; and summarizing and analyzing the group attribution results corresponding to each of the hypothetical groups to obtain a comprehensive attribution result.

[0012] Optionally, the step of rewriting the first preferred hypothesis using the hypothesis evolution agent based on the comprehensive attribution results to obtain the target preferred hypothesis includes: determining a rewriting strategy based on the comprehensive attribution results; wherein the rewriting strategy includes at least one of supplementing key mechanisms, adding limiting conditions, clarifying variable relationships, enhancing verifiability, and improving executability; and rewriting the first preferred hypothesis using the hypothesis evolution agent based on the rewriting strategy to obtain the target preferred hypothesis.

[0013] Optionally, the method further includes: using a retrieval agent to obtain documentary evidence; The method of using a hypothesis-generating agent to obtain contextual information corresponding to a research problem and generating multiple target candidate hypotheses for solving the research problem based on the contextual information includes: obtaining contextual information corresponding to the research problem and first literature evidence corresponding to the generated hypothesis, and generating multiple target candidate hypotheses for solving the research problem based on the contextual information and the first literature evidence. The step of inputting the target preference hypothesis into a multi-scheme generation agent to obtain a target experimental scheme corresponding to the target preference hypothesis includes: obtaining second literature evidence corresponding to the generated experimental scheme; inputting the second literature evidence and the target preference hypothesis into a multi-scheme generation agent to obtain a target experimental scheme corresponding to the target preference hypothesis.

[0014] Optionally, the step of inputting the second literature evidence and the target preference hypothesis into a multi-scheme generation agent to obtain a target experimental scheme corresponding to the target preference hypothesis includes: inputting the second literature evidence and the target preference hypothesis into a multi-scheme generation agent to generate multiple first experimental schemes; for each first experimental scheme: using a path analysis agent to perform path analysis on the first experimental scheme and generating path analysis results; calling the retrieval agent to obtain third literature evidence corresponding to the path analysis results, and supplementing the first experimental schemes according to the third literature evidence to obtain second experimental schemes; and determining a target experimental scheme from the multiple second experimental schemes according to the coverage of each second experimental scheme.

[0015] The present invention also provides an experimental scheme generation device based on a multi-agent system, comprising the following modules: The hypothesis generation module is used to obtain context information corresponding to the research problem using a hypothesis generation agent, and generate multiple target candidate hypotheses for solving the research problem based on the context information. The screening module is used to perform a Swiss-round comparison of multiple target candidate hypotheses using a referee agent, and to screen the multiple target candidate hypotheses according to the comparison results to obtain a first preferred hypothesis; The analysis module is used to perform causal analysis on the comparison results using a review agent to obtain a comprehensive attribution result; The rewriting module is used to rewrite the first preferred hypothesis using the hypothesis evolution agent based on the comprehensive attribution results to obtain the target preferred hypothesis; The experimental scheme generation module is used to input the target optimization hypothesis into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target optimization hypothesis.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the experimental scheme generation method based on a multi-agent system as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the experimental scheme generation method based on a multi-agent system as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the experimental scheme generation method based on a multi-agent system as described above.

[0019] This invention provides a method and apparatus for generating experimental schemes based on a multi-agent system. By utilizing a referee agent, a review agent, and a hypothesis evolution agent, it not only performs a preliminary screening of multiple target candidate hypotheses generated by the hypothesis generation agent through pairwise comparisons, but also conducts causal analysis based on the comparison results. Based on the comprehensive attribution results, the first preferred hypothesis obtained from the screening is rewritten with targeted optimization, resulting in a higher-quality target preferred hypothesis. This effectively overcomes the inefficiency and low quality problems caused by relying solely on the agent's own capabilities for non-targeted hypothesis generation. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts illustrating the experimental scheme generation method based on a multi-agent system provided by the present invention.

[0022] Figure 2 This is a schematic diagram of the process provided by the present invention, which uses a Swiss wheel to perform pairwise comparisons and generate round comparison results.

[0023] Figure 3 This is a detailed flowchart of step 103 provided by the present invention.

[0024] Figure 4 This is a schematic diagram of a target generation experimental scheme provided by the present invention.

[0025] Figure 5 This is the second flowchart of the experimental scheme generation method based on multi-agent systems provided by the present invention.

[0026] Figure 6 This is a schematic diagram of the experimental scheme generation device based on a multi-agent system provided by the present invention.

[0027] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] Figure 1 This is one of the flowcharts illustrating the experimental scheme generation method based on a multi-agent system provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Use the hypothesis generation agent to obtain contextual information corresponding to the research problem, and generate multiple target candidate hypotheses to solve the research problem based on the contextual information.

[0030] The contextual information corresponding to the research question can specifically include: the research question, research field, research preferences, the mechanism to be explained, experimental data, user constraints, relevant evaluation results, and other information related to the research question. This invention does not impose specific limitations on this; the more types of information included in the contextual information, the more beneficial it is for generating more accurate and relevant target candidate hypotheses. For example, the contextual information can be input or uploaded by researchers through an intelligent agent interface, or it can be obtained by the hypothesis-generating intelligent agent from a database or other storage medium.

[0031] In this embodiment of the invention, to improve the collaboration efficiency between different agents in a multi-agent system, the context information is a standardized field constructed and stored by the context modeling agent, which facilitates sharing and access by other agents in the multi-agent system. In an optional embodiment, before step 101, the method further includes: using the context modeling agent to obtain the research question and the corresponding research data; transforming the research data according to preset standardized fields to generate context information; and writing the context information into the task status shared and accessed by the multi-agent system.

[0032] In a further optional embodiment, the target candidate hypothesis is not a candidate hypothesis directly generated by the hypothesis generation agent, but rather a candidate hypothesis first generated by the hypothesis generation agent and then filtered by the hypothesis analysis agent. Specifically, the process of step 101 above may include: using the hypothesis generation agent to obtain contextual information corresponding to the research problem, and generating multiple candidate hypotheses to solve the research problem based on the contextual information; using the hypothesis analysis agent to perform hypothesis analysis on each candidate hypothesis, and determining the target candidate hypothesis from multiple candidate hypotheses based on the hypothesis analysis results. The role of the hypothesis analysis agent is to perform multi-dimensional analysis on each candidate hypothesis, which may specifically include at least one of semantic analysis, mechanism analysis, cluster analysis, and quality analysis. Specifically, when using the hypothesis analysis agent to perform analysis in different dimensions, an abstract definition corresponding to that dimension of analysis can be added to the prompts of the hypothesis analysis agent, relying on the agent's capabilities for analysis; this invention does not specifically limit this. For example, semantic analysis can be performed by comparing the research object, core variables, variable relationships, and expected results of the candidate hypotheses to summarize the theme or key concepts, thereby identifying multiple candidate hypotheses that differ only in expression but have the same substantive content. Mechanism analysis can be performed by comparing candidate hypotheses in terms of their subjects, intermediate links, causal paths, and final effects, thereby identifying multiple candidate hypotheses with the same or similar mechanisms of action. Furthermore, multiple candidate hypotheses can be grouped according to research direction and mechanism of action, and cluster analysis can be conducted based on the principle of "similarity within groups and differences between groups," thus obtaining multiple candidate hypotheses belonging to the same cluster category.

[0033] Furthermore, by conducting multi-dimensional analysis of each candidate hypothesis, effective screening and merging can be performed on the large number of candidate hypotheses generated by the hypothesis-generating agent to obtain higher-quality target candidate hypotheses. Specifically, when hypothesis analysis includes semantic analysis, multiple candidate hypotheses with semantic similarity higher than a preset similarity are merged as target candidate hypotheses; when hypothesis analysis includes mechanism analysis, multiple candidate hypotheses with the same mechanism are merged as target candidate hypotheses; when hypothesis analysis includes cluster analysis, multiple candidate hypotheses belonging to the same cluster category are merged as target candidate hypotheses. Through the above-mentioned multiple hypothesis analyses, candidate hypotheses with the same or similar expression level and the same or similar substantive mechanism can be identified. By merging multiple similar candidate hypotheses, on the one hand, the efficiency of subsequent experimental scheme generation can be improved, and on the other hand, multiple candidate hypotheses belonging to the same mechanism but with their own corresponding flaws can be merged to obtain target candidate hypotheses that are more complete in terms of a single mechanism, thereby improving the overall quality of the subsequently generated experimental schemes. In addition, when hypothesis analysis includes quality analysis, logical flaw identification, experimental feasibility verification, evidence gap identification, mechanism integrity verification, and executability verification can be performed for each candidate hypothesis. Based on the quality analysis results, candidate hypotheses with lower quality can be eliminated, and only candidate hypotheses with relatively higher quality can be retained as target candidate hypotheses.

[0034] In an optional embodiment, the multi-agent system of the present invention further includes a retrieval agent for acquiring documentary evidence to serve as knowledge support for other agents when performing corresponding steps. Specifically, step 101 may further include: acquiring contextual information corresponding to the research question and first documentary evidence corresponding to the generated hypothesis; and generating multiple target candidate hypotheses for solving the research question based on the contextual information and the first documentary evidence. That is, after receiving a call request or query request from other agents, the retrieval agent can perform multi-source retrieval from a local literature database, online data source, or metadata based on the call request or query request, and merge, deduplicate, standardize metadata, filter, and rearrange the retrieval results to remove documentary evidence that is irrelevant to the call request or query request, has a mismatched evidence type, is unreliable in its source, is duplicated, or lacks key metadata. The rearrangement may be based on at least one of the following: text relevance, entity matching degree, literature authority, publication time, citation information, and evidence relevance.

[0035] Step 102: Use the referee agent to perform a Swiss round comparison of multiple target candidate hypotheses, and filter the multiple target candidate hypotheses according to the comparison results to obtain the first preferred hypothesis.

[0036] Specifically, in one optional embodiment, a Swiss-round comparison method can be used to perform multiple rounds of pairwise comparisons on the target candidate hypotheses, and the target score corresponding to each target candidate hypothesis can be obtained based on the comparison results. The target candidate hypotheses are then filtered according to their target scores to obtain a first preferred hypothesis. It should be noted that in this embodiment of the invention, multiple target candidate hypotheses are not directly further optimized based on the comprehensive evaluation results of multiple dimensions. This is because when the number of target candidate hypotheses is large, directly performing a comprehensive multi-dimensional evaluation on each target candidate hypothesis would consume a large amount of computational resources, and the ranking method based on comprehensive scores cannot reflect the fine-grained quality differences between different target candidate hypotheses, thus failing to provide support for subsequent attribution. However, while performing pairwise comparisons on every two target candidate hypotheses can effectively reflect the fine-grained quality differences between each pair of target candidate hypotheses, it would undoubtedly increase computational costs in a large candidate pool. Therefore, this invention uses a Swiss-round comparison method to perform multiple rounds of pairwise comparisons, which avoids performing complete pairwise comparisons on all target candidate hypotheses, significantly reducing the comparison computational cost in a large candidate pool, while also providing fine-grained support for subsequent attribution through a small number of pairwise comparisons.

[0037] In a further optional embodiment, the comparison results include group comparison results corresponding to at least one round; the process of performing Swiss-system comparison and generating round comparison results is as follows: Figure 2 As shown, repeat the following steps in a loop until the preset number of rounds is reached: Step 201: Obtain the hypothesis integral for each target candidate hypothesis; Step 202: Based on the closeness between the integrals of each hypothesis and the number of target candidate hypotheses, determine multiple hypothesis groups corresponding to each round; wherein each hypothesis group includes at most two target candidate hypotheses; Step 203: Evaluate one or two target candidate hypotheses in each hypothesis group according to the preset evaluation indicators, and generate the group comparison results corresponding to each hypothesis group; Step 204: Based on the group comparison results of each hypothesis group, update the hypothesis integrals corresponding to each target candidate hypothesis, and use the updated hypothesis integrals as the target integrals corresponding to each target candidate hypothesis.

[0038] It is understandable that comparing two target candidate hypotheses with more similar quality yields more meaningful comparison results. Therefore, in step 202 of this embodiment, two target candidate hypotheses with a higher degree of similarity are preferentially grouped into one hypothesis group. This effectively improves the information value of the comparison in each round, helping the subsequent review agent to conduct effective causal analysis based on the comparison results of each round. For example, in a certain round, the scores of target candidate hypothesis 1, target candidate hypothesis 2, target candidate hypothesis 3, target candidate hypothesis 4, and target candidate hypothesis 5 are 3, 8, 10, 5, and 2 points respectively. When grouping the hypothesis groups, target candidate hypothesis 2 and target candidate hypothesis 3, corresponding to 8 and 10 points respectively, are grouped into one hypothesis group, and target candidate hypothesis 1 and target candidate hypothesis 5, corresponding to 3 and 2 points respectively, are grouped into another. Since there are a total of 5 target candidate hypotheses, target candidate hypothesis 4, corresponding to 5 points, is grouped with the bye position. It can be seen that this embodiment of the invention, by dividing the hypothesis groups based on the similarity between hypothesis scores, can effectively avoid the situation of comparing 10 points and 2 points. Understandably, comparing the target candidate hypotheses corresponding to 10 points and 2 points, since the hypotheses scores differ significantly, it indicates a large gap in the quality of the two target candidate hypotheses. The group comparison results are usually obvious, and even if subsequent analysis is based on these group comparison results, it is difficult to find more granular information. However, comparing the target candidate hypotheses corresponding to 10 points and 8 points, since the quality of the two target candidate hypotheses is similar, the comparison results can better illustrate the advantages and disadvantages of fine granularity, thus yielding more valuable group comparison results.

[0039] In step 203, the preset evaluation index can be a comprehensive index obtained by weighting the scores corresponding to the following multiple dimensions: scientific rigor, evidence support, feasibility, testability, relevance, novelty, and potential impact. The selection of the number of dimensions and the setting of the corresponding weights can be set according to actual needs, and this invention does not impose specific limitations on this. For example, taking the aforementioned five target candidate hypotheses as an example, the group comparison results can be that target candidate hypothesis 2 has stronger scientific rigor than target candidate hypothesis 3, higher evidence support than target candidate hypothesis 3, and both have feasibility and testability.

[0040] Specifically, assuming the score is the Glicko-2 score, which is a competitive scoring algorithm that introduces scoring bias and score volatility, the comparison between scientific hypotheses is relatively complex and difficult to represent through a simple score accumulation process (such as adding one point for winning). This embodiment of the invention combines the Glicko-2 scoring system with a Swiss-system comparison method, utilizing the concept of uncertainty in the Glicko-2 score to adjust the update magnitude of each target candidate hypothesis in the score update, thereby making the target scores corresponding to each target candidate hypothesis more reasonable. In an optional embodiment, the process of updating the hypothesis scores in step 204 above may specifically include: obtaining the uncertainty corresponding to each target candidate hypothesis; updating the Glicko-2 score based on the group comparison results and uncertainties of each hypothesis group to obtain the target scores corresponding to each target candidate hypothesis.

[0041] Step 103: Use the review agent to perform causal analysis on the comparison results and obtain comprehensive attribution results.

[0042] It should be noted that this step is crucial for the present invention to improve optimization efficiency in a targeted manner, and it works synergistically with the pairwise comparisons in step 102. Specifically, as mentioned earlier, the comparison results actually include the group comparison results corresponding to at least one round. By performing causal analysis on each group comparison result, the reasons for the success or failure of various target candidate hypotheses can be obtained from the pairwise comparisons of multiple hypothesis groups, thereby providing a clear optimization direction for the subsequent process of rewriting the first optimization hypothesis.

[0043] In an optional embodiment, the process of step 103 described above is as follows: Figure 3 As shown, it specifically includes: Step 301: Use the review agent to obtain the group comparison results for each round; Step 302: Perform causal analysis on the comparison results of each group to obtain the group attribution results corresponding to each hypothetical group; Step 303: Summarize and analyze the group attribution results corresponding to each hypothesis group to obtain the comprehensive attribution results.

[0044] The comprehensive attribution result can include failure modes, improvement priorities, win-loss patterns, and evidence gaps, etc. Specifically, failure modes can be determined based on common causes among the attribution results of each group, and improvement priorities can be determined based on the statistical results of each common cause (for example, if the statistical proportion of failures caused by the absence of hypothesis A is the highest, then the improvement priority of improving A is relatively high). In this embodiment of the invention, the abstract definitions of failure modes, improvement priorities, win-loss patterns, and evidence gaps can be pre-input as prompt words along with the group comparison results into the review agent, so that the review agent can determine the comprehensive attribution result based on its understanding of the abstract definitions and its own capabilities.

[0045] In summary, the embodiments of the present invention, through the collaboration between the review agent and the referee agent, can perform effective attribution analysis by combining the pairwise comparison results of multiple hypothesis groups during the pairwise comparison process. By summarizing the group attribution results corresponding to multiple hypothesis groups, a more comprehensive and higher-dimensional integrated attribution result can be obtained, providing a clear iterative direction for the subsequent rewriting process.

[0046] Step 104: Using the hypothesis evolution agent, rewrite the first optimal hypothesis based on the comprehensive attribution results to obtain the target optimal hypothesis.

[0047] Specifically, the rewriting process may include: determining a rewriting strategy based on the comprehensive attribution results; wherein the rewriting strategy includes at least one of supplementing key mechanisms, adding limiting conditions, clarifying variable relationships, enhancing verifiability, and improving executability; and using a hypothesis evolution agent to rewrite the first preferred hypothesis according to the rewriting strategy to obtain the target preferred hypothesis. For example, the hypothesis evolution agent will first identify defects based on the comprehensive attribution results and then generate targeted rewriting strategies based on these defects. For example, when the comprehensive attribution results indicate an incomplete mechanism description, a rewriting strategy can be generated that indicates key mechanisms such as supplementing the acting subject, intermediate links, and causal pathways; when the comprehensive attribution results indicate an overly broad research scope, a rewriting strategy can be generated that indicates limiting the research subjects, adding applicable conditions, and supplementing the research stages; when the comprehensive attribution results indicate unclear variable definitions, a rewriting strategy can be generated that indicates clarifying the relationship between variables, such as clarifying the independent or dependent variables; when the comprehensive attribution results indicate a lack of verifiable predictions, a rewriting strategy can be generated that indicates adding observation indicators, adding control conditions, and adding falsifiable results to enhance verifiability; and when the comprehensive attribution results indicate high implementation difficulty, a rewriting strategy can be generated that indicates combining available experimental methods and resources to improve feasibility.

[0048] In this invention, multiple rewriting strategies are prioritized to maximize the diversity of target-optimized hypotheses and provide researchers with more options. However, to prevent the generation of invalid target-optimized hypotheses during the diverse rewriting process, in an optional embodiment of the invention, the process of rewriting the first-optimized hypothesis using a hypothesis evolution agent according to the rewriting strategies specifically includes: rewriting the first-optimized hypothesis using the hypothesis evolution agent according to multiple rewriting strategies to obtain multiple second-optimized hypotheses; in response to the existence of multiple second-optimized hypotheses belonging to the same mechanism of action or the same research direction, the step of rewriting the first-optimized hypothesis using the hypothesis evolution agent according to multiple rewriting strategies is re-executed based on differentiated prompts. Specifically, the same key causal path can be identified as the same mechanism of action, and the same research objective, core question, and verification direction can be identified as the same research direction. In other words, this embodiment of the invention can determine multiple rewriting strategies based on comprehensive attribution results and perform differentiated rewriting based on each rewriting strategy, maximizing the diversity of target-optimized hypotheses.

[0049] Step 105: Input the target optimization hypothesis into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target optimization hypothesis.

[0050] In the field of scientific research, the target optimization hypothesis can only provide a guiding direction with scientific value, but it does not translate into specific solutions on how to achieve it. Therefore, obtaining only the target optimization hypothesis is far from sufficient. In this embodiment of the invention, an executable target experimental plan for researchers is further generated based on the target optimization hypothesis, transforming the guiding direction into a concrete and feasible plan, which greatly improves the operability and feasibility of the overall process.

[0051] It should be noted that the multiple target optimization hypotheses obtained from step 104 can be directly input into the scheme generation agent to obtain the target experimental schemes corresponding to each target optimization hypothesis. Alternatively, the target optimization hypotheses obtained from step 104 can be first displayed to researchers through the agent's interactive interface, and then researchers can select one or more to generate subsequent target experimental schemes. This invention does not impose specific limitations on this approach.

[0052] In an optional embodiment, similar to the process of calling the retrieval agent in 101 above, the generation of the target experimental scheme can also be coordinated with the retrieval agent, and step 105 specifically includes: obtaining second literature evidence corresponding to the generated experimental scheme; inputting the second literature evidence and the target preferred hypothesis into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target preferred hypothesis.

[0053] In a further optional embodiment, the process of inputting the second literature evidence and the target preference hypothesis into the multi-scheme generating agent to generate the target experimental scheme can be as follows: Figure 4 As shown, it includes: Step 401: Input the second literature evidence and the target optimization hypothesis into the multi-scheme generation agent to generate multiple first experimental schemes; Step 402: For each first experimental scheme: use the path analysis agent to perform path analysis on the first experimental scheme and generate path analysis results; call the retrieval agent to obtain the third literature evidence corresponding to the path analysis results, and supplement the first experimental scheme according to the third literature evidence to obtain the second experimental scheme. Step 403: Based on the coverage of each second experimental scheme, determine the target experimental scheme from multiple second experimental schemes.

[0054] Each first experimental scheme includes at least an experimental model, intervention conditions, indicator observations, and judgment criteria, so that researchers can intuitively understand the specific implementation process of the experiment based on the first experimental scheme.

[0055] As can be seen, the embodiments of the present invention utilize a path analysis agent to perform path analysis on the generated first experimental scheme, and effectively supplement the scheme by combining the third literature data corresponding to the path analysis results, thus significantly improving the completeness of the experimental scheme. Furthermore, by screening based on the coverage of each second experimental scheme, it is possible to ensure that multiple target experimental schemes covering different experimental models, intervention conditions, indicator observations, judgment criteria, and mechanisms of action are obtained first.

[0056] Below, with Figure 5 The specific process of the experimental scheme generation method based on a multi-agent system provided in this embodiment of the invention will be illustrated by way of example. Figure 5 As shown, it includes: Step 501: Use the context modeling agent to obtain the research question and the research data corresponding to the research question; transform the research data according to the preset standardized fields to generate context information, and write the context information into the task status shared and accessed by the multi-agent system; Step 502: Use the retrieval agent to obtain first-document evidence corresponding to the generated hypothesis; Step 503: Use the hypothesis generation agent to obtain contextual information corresponding to the research problem, and generate multiple candidate hypotheses to solve the research problem based on the contextual information and the first literature evidence. Step 504: Use the hypothesis analysis agent to perform hypothesis analysis on each candidate hypothesis, and determine the target candidate hypothesis from multiple candidate hypotheses based on the hypothesis analysis results; Step 505: Use the referee agent to perform a Swiss round comparison of multiple target candidate hypotheses, and filter the multiple target candidate hypotheses according to the comparison results to obtain the first preferred hypothesis; Step 506: Use the review agent to perform causal analysis on the comparison results and obtain comprehensive attribution results; Step 507: Using the hypothesis evolution agent, the first preferred hypothesis is rewritten based on the comprehensive attribution results to obtain the second preferred hypothesis; Step 508: Use the second preferred hypothesis, which belongs to a different mechanism of action or a different research direction, as the target preferred hypothesis; Step 509: Call the retrieval agent to obtain second literature evidence corresponding to the generated experimental plan; Step 510: Input the second literature evidence and the target preference hypothesis into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target preference hypothesis.

[0057] In summary, the experimental scheme generation method based on a multi-agent system provided in this embodiment of the invention not only performs a preliminary screening of multiple target candidate hypotheses generated by the hypothesis generating agent through pairwise comparisons, but also conducts causal analysis based on the comparison results. Based on the comprehensive attribution results, the first preferred hypothesis obtained from the screening is rewritten with targeted optimization, resulting in a higher-quality target preferred hypothesis. This effectively overcomes the inefficiency and low quality problems caused by relying solely on the ability of the agent itself to generate non-targeted hypotheses.

[0058] The experimental scheme generation device based on a multi-agent system provided by the present invention is described below. The experimental scheme generation device based on a multi-agent system described below and the experimental scheme generation method based on a multi-agent system described above can be referred to in correspondence.

[0059] like Figure 6 As shown, the experimental scheme generation device 600 based on a multi-agent system provided in this embodiment of the invention includes: Hypothesis generation module 601 is used to obtain context information corresponding to the research problem using a hypothesis generation agent, and generate multiple target candidate hypotheses for solving the research problem based on the context information; The screening module 602 is used to perform a Swiss round comparison of multiple target candidate hypotheses using a referee agent, and to screen the multiple target candidate hypotheses according to the comparison results to obtain a first preferred hypothesis; Analysis module 603 is used to perform causal analysis on the comparison results using a review agent to obtain a comprehensive attribution result; The rewriting module 604 is used to rewrite the first preferred hypothesis using the hypothesis evolution agent based on the comprehensive attribution results to obtain the target preferred hypothesis; The experimental scheme generation module 605 is used to input the target preference hypothesis into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target preference hypothesis.

[0060] In an optional embodiment of the present invention, the hypothesis generation module 601 is further configured to: use a hypothesis generation agent to obtain context information corresponding to the research problem, and generate multiple candidate hypotheses for solving the research problem based on the context information; use a hypothesis analysis agent to perform hypothesis analysis on each of the candidate hypotheses, and determine a target candidate hypothesis from the multiple candidate hypotheses based on the hypothesis analysis results.

[0061] In an optional embodiment of the present invention, the hypothesis analysis includes at least one of semantic analysis, mechanism of action analysis, and cluster analysis; the hypothesis generation module 601 is further configured to, when the hypothesis analysis includes semantic analysis, merge multiple candidate hypotheses with semantic similarity higher than a preset similarity to serve as the target candidate hypothesis; when the hypothesis analysis includes mechanism of action analysis, merge multiple candidate hypotheses with the same mechanism of action to serve as the target candidate hypothesis; and when the hypothesis analysis includes cluster analysis, merge multiple candidate hypotheses belonging to the same cluster category to serve as the target candidate hypothesis.

[0062] In an optional embodiment of the present invention, the screening module 602 is further configured to perform pairwise comparisons of the target candidate hypothesis in multiple rounds using a Swiss-system comparison method, generate a round comparison result for each round, and screen the multiple target candidate hypotheses according to the round comparison results corresponding to each round to obtain a first preferred hypothesis.

[0063] In an optional embodiment of the present invention, the comparison result includes group comparison results corresponding to at least one round; the filtering module 602 is further configured to repeatedly execute the following steps until the number of rounds reaches a preset number: obtaining the hypothesis score corresponding to each of the target candidate hypotheses; determining multiple hypothesis groups corresponding to the rounds based on the closeness between the hypothesis scores and the number of target candidate hypotheses; wherein each hypothesis group includes at most two target candidate hypotheses; evaluating one or two target candidate hypotheses in each hypothesis group according to a preset evaluation index to generate a group comparison result corresponding to each hypothesis group; updating the hypothesis scores corresponding to each target candidate hypothesis based on the group comparison results corresponding to each hypothesis group, and using the updated hypothesis scores as the target scores corresponding to each target candidate hypothesis.

[0064] In an optional embodiment of the present invention, the hypothesis integral is a Glicko-2 score; the filtering module 602 is further configured to obtain the uncertainty corresponding to each of the target candidate hypotheses; and update the Glicko-2 score according to the group comparison results corresponding to each hypothesis group and the uncertainty, so as to obtain the target integral corresponding to each of the target candidate hypotheses.

[0065] In an optional embodiment of the present invention, the analysis module 603 is further configured to: use the review agent to obtain the round comparison results corresponding to each round; perform causal analysis on each round comparison result to obtain the attribution results corresponding to each round comparison result; and summarize and analyze the attribution results corresponding to each round to obtain a comprehensive attribution result.

[0066] In an optional embodiment of the present invention, the rewriting module 604 is further configured to determine a rewriting strategy based on the comprehensive attribution result; wherein the rewriting strategy includes at least one of supplementing key mechanisms, adding limiting conditions, clarifying variable relationships, enhancing verifiability, and improving executability; and using a hypothesis evolution agent to rewrite the first preferred hypothesis according to the rewriting strategy to obtain the target preferred hypothesis.

[0067] In an optional embodiment of the present invention, the experimental scheme generation device 600 further includes a retrieval module for obtaining literature evidence using a retrieval agent; the hypothesis generation module 601 is further configured to obtain context information corresponding to the research problem and first literature evidence corresponding to the generated hypothesis, and generate multiple target candidate hypotheses for solving the research problem based on the context information and the first literature evidence; the experimental scheme generation module 605 is further configured to obtain second literature evidence corresponding to the generated experimental scheme; and input the second literature evidence and the target preferred hypothesis into the multi-scheme generation agent to obtain a target experimental scheme corresponding to the target preferred hypothesis.

[0068] In an optional embodiment of the present invention, the experimental scheme generation module 605 is further configured to: input the second literature evidence and the target preferred hypothesis into a multi-scheme generation agent to generate multiple first experimental schemes; for each first experimental scheme: use a path analysis agent to perform path analysis on the first experimental scheme and generate path analysis results; call the retrieval agent to obtain third literature evidence corresponding to the path analysis results, and supplement the first experimental scheme according to the third literature evidence to obtain a second experimental scheme; and determine the target experimental scheme from the multiple second experimental schemes according to the coverage of each second experimental scheme.

[0069] In summary, the experimental scheme generation device based on a multi-agent system provided in this embodiment of the invention utilizes a referee agent, a review agent, and a hypothesis evolution agent. It not only performs a preliminary screening of multiple target candidate hypotheses generated by the hypothesis generation agent through pairwise comparisons, but also conducts causal analysis based on the comparison results. Based on the comprehensive attribution results, it performs targeted optimization and rewriting of the first preferred hypothesis obtained from the screening, resulting in a higher-quality target preferred hypothesis. This effectively overcomes the inefficiency and low quality problems caused by relying solely on the agent's own capabilities for non-directional hypothesis generation.

[0070] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communications interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute an experimental scheme generation method based on a multi-agent system. This method includes: using a hypothesis generating agent to obtain context information corresponding to a research problem, and generating multiple target candidate hypotheses for solving the research problem based on the context information; using a referee agent to perform a Swiss-system comparison of the multiple target candidate hypotheses, and filtering the multiple target candidate hypotheses based on the comparison results to obtain a first preferred hypothesis; using a review agent to perform causal analysis on the comparison results to obtain a comprehensive attribution result; using a hypothesis evolution agent to rewrite the first preferred hypothesis based on the comprehensive attribution result to obtain a target preferred hypothesis; and inputting the target preferred hypothesis into a multi-scheme generating agent to obtain a target experimental scheme corresponding to the target preferred hypothesis.

[0071] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0072] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the experimental scheme generation method based on a multi-agent system provided by the above methods. The method includes: using a hypothesis generation agent to obtain context information corresponding to a research problem, and generating multiple target candidate hypotheses for solving the research problem based on the context information; using a referee agent to perform a Swiss-system comparison of the multiple target candidate hypotheses, and filtering the multiple target candidate hypotheses based on the comparison results to obtain a first preferred hypothesis; using a review agent to perform causal analysis on the comparison results to obtain a comprehensive attribution result; using a hypothesis evolution agent to rewrite the first preferred hypothesis based on the comprehensive attribution result to obtain a target preferred hypothesis; and inputting the target preferred hypothesis into a multi-scheme generation agent to obtain a target experimental scheme corresponding to the target preferred hypothesis.

[0073] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the experimental scheme generation method based on a multi-agent system provided by the above methods. The method includes: using a hypothesis generating agent to obtain context information corresponding to a research problem, and generating multiple target candidate hypotheses for solving the research problem based on the context information; using a referee agent to perform a Swiss-system comparison of the multiple target candidate hypotheses, and filtering the multiple target candidate hypotheses based on the comparison results to obtain a first preferred hypothesis; using a review agent to perform causal analysis on the comparison results to obtain a comprehensive attribution result; using a hypothesis evolution agent to rewrite the first preferred hypothesis based on the comprehensive attribution result to obtain a target preferred hypothesis; and inputting the target preferred hypothesis into a multi-scheme generating agent to obtain a target experimental scheme corresponding to the target preferred hypothesis.

[0074] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0075] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating experimental schemes based on multi-agent systems, characterized in that, include: The hypothesis-generating agent acquires contextual information corresponding to the research problem and generates multiple target candidate hypotheses for solving the research problem based on the contextual information. The referee agent performs a Swiss round comparison on multiple target candidate hypotheses, and filters the multiple target candidate hypotheses based on the comparison results to obtain the first preferred hypothesis; The comparison results are analyzed using a review agent to obtain a comprehensive attribution result. The first preferred hypothesis is rewritten using the hypothesis evolution agent based on the comprehensive attribution results to obtain the target preferred hypothesis; The target optimization hypothesis is input into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target optimization hypothesis.

2. The method for generating experimental schemes based on multi-agent systems according to claim 1, characterized in that, The process of using a hypothesis-generating agent to acquire contextual information corresponding to a research problem, and generating multiple target candidate hypotheses for solving the research problem based on the contextual information, includes: The hypothesis-generating agent acquires contextual information corresponding to the research problem and generates multiple candidate hypotheses to solve the research problem based on the contextual information. The hypothesis analysis agent performs hypothesis analysis on each of the candidate hypotheses, and determines the target candidate hypothesis from the multiple candidate hypotheses based on the hypothesis analysis results.

3. The method for generating experimental schemes based on multi-agent systems according to claim 1, characterized in that, The process of using a referee agent to perform a Swiss-system comparison of multiple target candidate hypotheses, and then filtering the multiple target candidate hypotheses based on the comparison results to obtain a first preferred hypothesis, includes: The target candidate hypothesis is compared pairwise in multiple rounds using a Swiss-system comparison method, and the target integral corresponding to each target candidate hypothesis is obtained based on the comparison results. The first preferred hypothesis is obtained by filtering multiple target candidate hypotheses according to the target integral.

4. The method for generating experimental schemes based on multi-agent systems according to claim 3, characterized in that, The comparison results include group comparison results corresponding to at least one round; the method of using Swiss-system comparison to perform pairwise comparisons on the target candidate hypotheses for multiple rounds, and obtaining the target integral corresponding to each target candidate hypothesis based on the comparison results, includes: Repeat the following steps until the preset number of rounds is reached: Obtain the hypothesis integral for each of the target candidate hypotheses; Based on the proximity between the integrals of each hypothesis and the number of target candidate hypotheses, multiple hypothesis groups corresponding to each round are determined; wherein each hypothesis group includes at most two target candidate hypotheses. The target candidate hypotheses of each hypothesis group are evaluated according to the preset evaluation indicators, and the group comparison results corresponding to each hypothesis group are generated. Based on the group comparison results corresponding to each set of hypotheses, update the hypothesis integrals corresponding to each target candidate hypothesis, and use the updated hypothesis integrals as the target integrals corresponding to each target candidate hypothesis.

5. The method for generating experimental schemes based on multi-agent systems according to claim 4, characterized in that, The assumed integral is the Glicko-2 rating; The step of updating the hypothesis integral corresponding to each target candidate hypothesis based on the group comparison results corresponding to each hypothesis group includes: Obtain the uncertainty corresponding to each of the target candidate hypotheses; The Glicko-2 score is updated based on the group comparison results corresponding to each of the hypotheses and the uncertainty.

6. The method for generating experimental schemes based on multi-agent systems according to claim 4, characterized in that, The step of using a review agent to perform causal analysis on the comparison results to obtain a comprehensive attribution result includes: The evaluation agent is used to obtain the group comparison results corresponding to each round. Each group comparison result is analyzed for cause, and the group attribution results corresponding to each hypothetical group are obtained. The group attribution results corresponding to each of the hypothetical groups are summarized and analyzed to obtain the comprehensive attribution results.

7. The method for generating experimental schemes based on multi-agent systems according to claim 1, characterized in that, The process of rewriting the first preferred hypothesis using the hypothetical evolutionary agent based on the comprehensive attribution results to obtain the target preferred hypothesis includes: The rewriting strategy is determined based on the comprehensive attribution results; wherein the rewriting strategy includes at least one of the following: supplementing key mechanisms, adding limiting conditions, clarifying variable relationships, enhancing verifiability, and improving executability; The first preferred hypothesis is rewritten using the hypothesis evolution agent according to the rewriting strategy to obtain the target preferred hypothesis.

8. The method for generating experimental schemes based on multi-agent systems according to claim 1, characterized in that, The method also includes: Utilizing search agents to obtain documentary evidence; The process of using a hypothesis-generating agent to acquire contextual information corresponding to a research problem, and generating multiple target candidate hypotheses for solving the research problem based on the contextual information, includes: Obtain contextual information corresponding to the research question and first literature evidence corresponding to the generated hypothesis, and generate multiple target candidate hypotheses to solve the research question based on the contextual information and the first literature evidence; The step of inputting the target preference hypothesis into a multi-solution generation agent to obtain a target experimental scheme corresponding to the target preference hypothesis includes: Obtain secondary literature evidence corresponding to the generated experimental design; The second literature evidence and the target preference hypothesis are input into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target preference hypothesis.

9. The method for generating experimental schemes based on multi-agent systems according to claim 8, characterized in that, The step of inputting the second literature evidence and the target optimization hypothesis into a multi-solution generation agent to obtain a target experimental scheme corresponding to the target optimization hypothesis includes: The second literature evidence and the target preference hypothesis are input into the multi-scheme generation agent to generate multiple first experimental schemes; For each of the first experimental schemes: the path analysis agent is used to perform path analysis on the first experimental scheme and generate path analysis results; the retrieval agent is called to obtain third-party literature evidence corresponding to the path analysis results, and the first experimental scheme is supplemented based on the third-party literature evidence to obtain the second experimental scheme; Based on the coverage of each of the second experimental schemes, the target experimental scheme is determined from the multiple second experimental schemes.

10. An experimental scheme generation device based on a multi-agent system, characterized in that, include: The hypothesis generation module is used to obtain context information corresponding to the research problem using a hypothesis generation agent, and generate multiple target candidate hypotheses for solving the research problem based on the context information. The screening module is used to perform a Swiss-round comparison of multiple target candidate hypotheses using a referee agent, and to screen the multiple target candidate hypotheses according to the comparison results to obtain a first preferred hypothesis; The analysis module is used to perform causal analysis on the comparison results using a review agent to obtain a comprehensive attribution result; The rewriting module is used to rewrite the first preferred hypothesis using the hypothesis evolution agent based on the comprehensive attribution results to obtain the target preferred hypothesis; The experimental scheme generation module is used to input the target optimization hypothesis into the multi-scheme generation agent to obtain the target experimental scheme corresponding to the target optimization hypothesis.

Images