Intelligent decision-making method, computer equipment and computer readable storage medium

Abstract

The embodiment of the invention discloses an intelligent decision-making method, computer equipment and a computer readable storage medium. The method comprises the following steps that when an intelligent agent obtains request content, the intelligent agent performs analysis operation on the request content so as to split at least one thinking path; taking each thinking path as input of a local solving model, and solving the thinking paths according to the local solving model to obtain corresponding candidate answers; summarizing candidate answers corresponding to all thinking paths to obtain a candidate answer set; and taking the candidate answer set as the input of a global meta-learner, and performing fusion processing operation on the candidate answer set according to the global meta-learner to obtain a final answer and outputting the final answer. Therefore, the request can be converted into various thinking paths by utilizing the intelligent agent, and various candidate answer sets are obtained by parallel solution; and carrying out global learning on the candidate answer set by a global meta-learner to obtain a final answer which is optimal in performance and best in conformity with expectation, and outputting the final answer.

Description

Technical Field

[0001] The present application belongs to the technical field of natural language processing, and in particular, relates to an intelligent decision-making method, a computer device, and a computer-readable storage medium. Background Art

[0002] In recent years, artificial intelligence technology has developed rapidly, especially driven by deep learning and neural networks, and many core areas have ushered in unprecedented breakthroughs. In the field of natural language processing (NLP), the emergence of large-scale language models (LLMs) has set off a technological revolution. Among them, a large number of civilian and commercial-grade language models have appeared on the market. Through the training of massive corpus, they have demonstrated excellent language understanding and generation capabilities and are widely used in tasks such as text generation, dialogue systems, machine translation, and question-answering retrieval. These models not only significantly improve the accuracy of tasks and the fluency of generated content, but also give rise to new application scenarios, such as intelligent customer service, writing assistants, and education platforms.

[0003] However, as these models are applied to more complex and dynamic scenarios, the bottleneck of single model capabilities gradually emerges. For example, traditional dialogue models often perform poorly in scenarios that require long-term dialogue coherence or deep logical reasoning; in multi-step task planning or situations that require analysis from multiple perspectives, single models are easily limited to a certain fixed idea. Therefore, how to improve the response performance of language models is a technical problem that technicians in this field need to solve urgently.

[0004] The preceding description is intended to provide general background information and does not necessarily constitute prior art. Summary of the invention

[0005] Based on this, it is necessary to propose an intelligent decision-making method, a computer device and a computer-readable storage medium to address the above problems, which can make decisions based on requests comprehensively and comprehensively.

[0006] The present application solves the technical problem by adopting the following technical solutions: The present application provides an intelligent decision-making method, comprising the following steps: when an intelligent agent obtains request content, the intelligent agent performs an analysis operation on the request content to split out at least one thinking path; each thinking path is used as the input of a local solution model, and the thinking path is solved according to the local solution model to obtain a corresponding candidate answer; the candidate answers corresponding to all thinking paths are summarized to obtain a candidate answer set; the candidate answer set is used as the input of a global meta-learner, and after the global meta-learner performs a fusion processing operation on the candidate answer set, a final answer is obtained and output.

[0007] In an optional embodiment of the present application, the intelligent agent performs an analysis operation on the request content, including: determining content attributes of the request content, the content attributes including at least one of text, pictures, videos, and audios; determining a feature extraction tool based on the content attributes, and using the feature extraction tool to extract key features from the request content; parsing the key features to determine the core intent of the request content; determining at least one reply sub-idea based on the key features and the core intent; and combining the reply sub-ideas to obtain a thinking path.

[0008] In an optional embodiment of the present application, the local solution model is composed of multiple solution sub-models; the thinking path is solved according to the local solution model, including: the solution sub-model splits the thinking path into at least one sub-thinking path, and the sub-thinking path includes multiple sub-problems; the sub-problems within the same sub-thinking path are arranged in order, and the sub-problems between different sub-thinking paths are determined by the arrangement order. The corresponding solution rounds; the solution sub-model solves the sub-problems corresponding to the same solution round between multiple sub-thinking paths in parallel to obtain corresponding sub-reply contents; all sub-reply contents under the same sub-thinking path are summarized to obtain sub-answers of the sub-thinking path; and first-order fusion processing is performed on the sub-answers between all sub-thinking paths to obtain candidate answers for the thinking path.

[0009] In an optional embodiment of the present application, the solving sub-model solves the sub-problems corresponding to the same solving round between multiple sub-thinking paths in parallel to obtain corresponding sub-reply contents, including: obtaining the core intent of the request content determined after the analysis operation, and determining the scoring calculation template of the sub-thinking path according to the core intent; after each round of solving, calculating the evaluation score of the sub-reply content according to the scoring calculation template; if the evaluation score is higher than the deep thinking threshold, generating a new sub-problem based on the sub-reply content to be added to the next round for parallel solving; and / or, after each round of solving, calculating the correlation between the sub-problem of the next round based on the sub-reply content of the current round; if the correlation is lower than the preset correlation threshold, generating a new sub-problem based on the sub-reply content to be added to the next round for parallel solving.

[0010] In an optional embodiment of the present application, the sub-model is solved in parallel to solve the sub-problems corresponding to the same solution round between multiple sub-thinking paths to obtain the corresponding sub-reply content, including: obtaining a scoring calculation template; after each round of solution, calculating the evaluation scores of all sub-reply contents according to the scoring calculation template; if there is a sub-reply content with an evaluation score lower than the retention threshold, then the sub-problem of the next round corresponding to the sub-reply content is marked as a problem to be deleted; and / or, after each round of solution, the number of sub-problems to be solved in parallel in the next round is determined; when the number of sub-problems is greater than a preset upper limit, the evaluation scores of all sub-reply contents in the current round are calculated according to the scoring calculation template; the top N sub-reply contents with the evaluation scores are retained, and the sub-problems of the next round corresponding to the sub-reply contents that are not retained are marked as problems to be deleted; N is a preset integer, N is less than or equal to the preset upper limit; and the sub-thinking path with the problem to be deleted as the root is deleted.

[0011] In an optional embodiment of the present application, a first-order fusion process is performed on the sub-answers between all sub-thinking paths, including: obtaining all sub-reply contents constituting each sub-answer, and the evaluation scores of the sub-reply contents; extracting key features in the sub-reply contents, the key features are used to characterize the core elements within the sub-reply; performing weighted scoring according to the evaluation scores, and selecting the key features with the highest scores according to the scoring results for combination to construct candidate answers; the candidate answers are composed of multiple key features, and the sum of the evaluation scores of the key features constituting the candidate answers is greater than or equal to the sum of the evaluation scores of all sub-reply contents under any sub-thinking path.

[0012] In an optional embodiment of the present application, after the global meta-learner performs a fusion processing operation on the candidate answer set, it includes: obtaining feature information of each candidate answer in the candidate answer set, and using the feature information as input to the global meta-learner for processing, the global meta-learner is a linear model, a tree model, a neural network and / or a large language model. One or more constituent models; the global meta-learner trains and / or infers the feature information to weightedly fuse all the feature information, and processes and outputs it as an answer to be output, which is composed of a combination of feature information; determines whether the answer to be output meets the output conditions; if not, uses the answer to be output and the request content as the input of the intelligent body for iteration again; if satisfied, marks the answer to be output as the final answer.

[0013] In an optional embodiment of the present application, a global meta-learner trains and / or infers feature information, including: retrieving additional feature information based on the request content, where the additional feature information is external information related to the request content in an existing database; and using the additional feature information and feature information as input to the global meta-learner for training and / or inference.

[0014] The present application also provides a computer device, including a processor and a memory: the processor is used to execute a computer program stored in the memory to implement the aforementioned method.

[0015] The present application also provides a computer-readable storage medium storing a computer program, which implements the aforementioned method when the computer program is executed by a processor.

[0016] The embodiments of the present application have the following beneficial effects: This application can use the intelligent agent to obtain the request content and convert it into multiple thinking paths, which are then handed over to the local solution model for parallel solving to obtain a diverse set of candidate answers; the candidate answer set is then used as the input of the global meta-learner for global learning, thereby analyzing the pros and cons of the answers under multiple thinking paths, and ultimately outputting the final answer with the best performance and the most expected after considering multiple aspects.

[0017] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically cited and described in detail with the accompanying drawings. It should be understood that the above general description and the detailed description below are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 A flowchart of an intelligent decision-making method provided by an embodiment.

[0020] Figure 2 A schematic diagram of the thinking path of chain thinking provided by an embodiment.

[0021] Figure 3 A schematic diagram of a tree-like thinking path provided by an embodiment.

[0022] Figure 4 A schematic diagram of tree-like thinking to expand the depth of thinking provided by an embodiment.

[0023] Figure 5 A schematic diagram of a second-order fusion processing flow provided by an embodiment.

[0024] Figure 6A schematic diagram of the data processing flow path of an intelligent decision-making method provided in one embodiment.

[0025] Figure 7 A schematic block diagram of the structure of a computer device provided by an embodiment. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] In the field of natural language processing, the choice of reasoning method directly affects the depth and breadth of the model's problem solving. The traditional Chain of Thought (CoT) reasoning method is a linear process. The model usually gradually develops reasoning around a single path, trying to obtain the final result through staged logical deduction. This method performs well in simple tasks, but it has limitations when facing complex problems.

[0028] For example, in scenarios that require multi-angle analysis, chain thinking often finds it difficult to consider potential solutions for different paths at the same time. In multi-step iterative tasks, once a step of reasoning is wrong or key information is omitted, the subsequent steps are often difficult to correct, resulting in the quality of the final result being affected. In addition, this linear reasoning method lacks the ability to dynamically adjust and explore diverse solutions. The current common technologies and problems are as follows: 1. Limitations of single-path reasoning; once traditional chain thinking or single-model prediction deviates in a certain step, the subsequent steps will be seriously affected and difficult to correct in time, and there is also a lack of parallel exploration capabilities for multiple solutions. 2. Simple voting or weighting is not enough to cope with high-complexity scenarios; some systems use simple voting or average weighting to fuse multiple model results, but ignore the potential differences and contextual features between model outputs, resulting in limited fusion effects. 3. Lack of multi-round iterative evaluation and pruning mechanisms; currently many multi-model or multi-path fusion methods terminate after initial fusion, lack a system for iterative optimization of errors and potential improvement space in multiple rounds, and cannot fully explore the solution space of complex problems.

[0029] In order to overcome the above technical difficulties, this application proposes an intelligent decision-making method. To clearly describe the method provided by this embodiment, please refer to Figure 1 to Figure 6 , including steps S110~S130.

[0030] This application is different from the prior art that uses CoT reasoning. This application combines the multi-path parallel reasoning ideas of Tree of Thoughts (ToT) with ensemble learning methods such as Stacking to form a multi-round reasoning and fusion framework. By performing preliminary fusion ("first-order fusion") within each thinking path and using a global meta-learner ("second-order fusion") for deep integration at the global level, the adaptability and robustness of large language models in complex multi-step tasks can be significantly improved. At the same time, with heuristic scoring and pruning strategies, it can continuously select and eliminate redundancy in multiple rounds of iterations, thereby obtaining a more comprehensive and accurate final answer while effectively controlling computing overhead.

[0031] Step S110: When the agent obtains the request content, the agent performs an analysis operation on the request content to separate out at least one thinking path.

[0032] In one embodiment, step S110: the intelligent agent performs an analysis operation on the request content, including: determining the content attributes of the request content, the content attributes including at least one of text, picture, video, and audio; determining a feature extraction tool based on the content attributes, and using the feature extraction tool to extract key features from the request content; parsing the key features to determine the core intent of the request content; determining at least one reply sub-idea based on the key features and the core intent; and combining the reply sub-ideas to obtain a thinking path.

[0033] In one embodiment, in this application, an agent is used as the input interface of the entire system. The agent is different from the input of the traditional large language model. The agent can perceive the physical / digital environment (such as a camera, API) through sensors or interfaces, while the traditional large language model intelligently processes text input. In other words, it can be preliminarily understood that the agent is a large model with parsing tools (sensors or interfaces). Therefore, the request content that the agent can accept is no longer limited to text types, and may include but is not limited to various types of content such as text, pictures, videos, and audio. The agent analyzes the request content, and it is necessary to adapt the corresponding tools for parsing according to the content attributes of the request content. Therefore, after the agent obtains the request content, the content attributes of the request content can be determined. The content attributes are used to characterize what type of data the request content belongs to, and specifically may include but is not limited to one or more of text, pictures, videos, and audio.

[0034] Furthermore, the corresponding feature extraction tool is adapted according to the content attributes. Different feature extraction tools are suitable for different content attributes, and are used to extract key features from the request content, and determine the core intent of the request content based on the analysis of the key features. Taking text content as an example, the feature extraction tool can be a bag of words model (BoW), word embedding, or a pre-trained language model (BERT, GPT), etc. Taking the bag of words model as an example, for the input text content, the keywords can be extracted, the frequency of occurrence of words in the text can be counted, and the grammar and order are ignored. In this way, the theme of the text content can be determined, that is, the core intent of the request content can be determined. For example, if the words "discount" and "offer" appear in large quantities in the text, it indicates the intention of promotion. Similarly, the feature extraction tools for image content include deep learning models, multimodal models (CLIP), etc.; the feature extraction tools for video content include key frame extraction, spatiotemporal feature extraction, etc.; the feature extraction tools for audio content include acoustic feature extraction, speech recognition (ASR), etc. In addition, the request content may be a combination of the above content attributes. When analyzing, multiple feature extraction tools can be combined for feature fusion. For example, multimodal content needs to align and fuse features (such as visual + audio of a video), and then use tools such as SVM, random forest or neural network to classify the core intent of the request content. In this way, the core intent represented by the request content and the extracted key features are formed into small problems that can be solved, that is, determined as at least one sub-idea of ​​the response.

[0035] Let's take a specific example. Assume that the user inputs "how to get from Sichuan to Shanghai", the key features are the two geographical locations "Sichuan" and "Shanghai", and the core intention is "arrival". For this purpose, you can select the means of transportation to arrive, including but not limited to private cars, travel buses, high-speed trains, airplanes, ferries, etc.; you can also determine the provinces that can be passed between Sichuan and Shanghai, and the ways of arrival are divergent, you can go up to Shaanxi, go down to Guizhou, or go straight to Chongqing, and then arrive in Shanghai one by one. Among them, assuming that the means of transportation is selected as a car, and the location is selected to go to Chongqing first, then the first step of how to drive from Sichuan to Chongqing becomes the first answer sub-idea. With this as a reference, you can specifically determine multiple continuous and combinable answer sub-ideas. The answer sub-ideas are combined to obtain a thinking path, and the combination process needs to meet the problem solving is continuous and feasible. For example, the first answer sub-idea is how to drive from Sichuan to Chongqing, then the next idea should be how to start from Chongqing and reach the next sub-target point. Instead of choosing the next step from Guizhou to the next sub-target point in an incoherent way. Therefore, this series of coherent thinking steps is also a thinking path.

[0036] It is also worth noting that if faced with the above problems, the traditional chain thinking reasoning method will be centered on a single path to get the answer. Usually there are the following problems: 1. Local optimality problem: In multi-step reasoning, once a certain step of reasoning deviates or omits key information, it is difficult to correct the subsequent steps, causing the final result to deviate from the actual goal. 2. Lack of global vision: Single-path methods can usually only be carried out along preset ideas, and cannot explore multi-angle, multi-hypothesis or multi-strategy solutions in parallel, thereby limiting the diversity and comprehensiveness of the reasoning results. 3. Low adaptability: Faced with complex tasks that require dynamic adjustment or multi-dimensional analysis, the single-path reasoning method lacks flexibility and is difficult to meet the needs of complex scenarios.

[0037] To this end, this application generates thinking paths in a tree-like (ToT) manner. Multiple paths are explored in parallel, and each path represents a different hypothesis or thinking angle, overcoming the limitations of single-path reasoning. Within the path (i.e., the "first-order fusion" stage), multiple model outputs are preliminarily screened and merged, which can retain multiple ideas and prevent the explosive growth of the number of candidates within the path, thereby achieving richer and more diversified thinking.

[0038] That is, the thinking path generated by the agent. In a preferred embodiment, there is not just one, but multiple paths. Each path may explore the answer through different logical sequences, different assumptions, or different prompts. In the subsequent stages, each path may further branch to form a deeper tree structure. To show the difference, we use the example of how to get from Sichuan to Shanghai mentioned above to show Figure 2 and Figure 3 The thought chain shown, in which Figure 2 The thinking path formed by the chain thinking represented by the existing technology; Figure 3 , is the thinking path under the tree-like thinking of this application. Specifically, the entire chain from the Sichuan block to the Shanghai block, connected from beginning to end, is also a thinking path. It should be understood that Figure 2 , Figure 3These are all simple examples of thinking paths. In reality, when faced with complex problems, thinking paths will only become more complex, and there will be correlations between the answer sub-thoughts, which will be specifically reflected in the following text. This application adopts a tree-like thinking approach, so the implementation process also adopts the common process of tree-like thinking: 1. Problem decomposition: break down complex problems into several operational sub-problems or steps. 2. Path generation: Based on the language model, for each sub-problem or each stage, multiple possible solutions or intermediate reasoning results are generated to form different "branch of thinking". 3. Path evaluation: Introduce heuristics or evaluation indicators to score the rationality, feasibility or score of each path, and retain the best / most potential path. 4. Expand and repeat: At a new stage or level, use the retained path to continue to generate subsequent branches of thinking, and repeat the evaluation and screening process. 5. Integrate the answer: In the final stage, the path information obtained through evaluation and expansion is summarized to form the final answer. That is, 1. Turn the problem into a tree: each step can derive different branches of thinking (such as different writing directions, different problem-solving ideas, etc.). 2. Explore multiple possibilities in parallel: Consider multiple alternatives at the same time, rather than just taking one path. 3. Continuous screening and merging: Regularly evaluate which ideas are good and which are not suitable, retain high-quality branches and continue to deepen, and finally integrate the best answer or idea. The benefits of doing this: more flexible, more comprehensive, reducing the risk of "one-track mind" going to a dead end, and making thinking become as luxuriant as a "tree". The subsequent steps will reveal how to achieve it separately.

[0039] Step S120: Use each thinking path as the input of the local solution model, solve the thinking path according to the local solution model to obtain the corresponding candidate answer; summarize the candidate answers corresponding to all thinking paths to obtain a candidate answer set.

[0040] In one embodiment, the local solution model is composed of multiple solution sub-models, and the solution sub-models can be different types of solution models, or they can be the same type but with different parameter configurations. Specifically, the local solution model can be matched according to the core intent of the request content. For example, for the request content from Sichuan to Shanghai mentioned above, the local solution model can be a navigation-related model; similarly, if it is about the weather, it can be a weather-related model. And the local solution model can also be composed of multiple sub-models. For example, different sub-models are used to solve three different means of transportation: cars, trains, and airplanes. For solution models of the same type but with different parameter configurations, taking cars as an example, it can be a private car, a bus, or a motorcycle; the road choice can be a national highway or a highway, which is also a different parameter configuration.

[0041] It is also worth noting that although the local solution model provided by this application is composed of multiple solution sub-models, it is different from the simple fusion of existing multi-models. In the prior art, multi-model integration methods often use simple voting or weighted average methods, and there are the following problems: Insufficient information utilization: lack of in-depth analysis of the characteristics and differences of the outputs of different models, and failure to tap the potential of multi-model collaboration. Insufficient robustness: Simple fusion mechanisms cannot effectively handle high-complexity tasks and are easily affected by low-quality models or path outputs, resulting in unstable results. Lack of dynamic feedback mechanism: Most methods output results after preliminary fusion, lack of multi-round optimization strategies based on evaluation and feedback, and it is difficult to continuously improve results in complex tasks. This application will score and screen the sub-answers obtained after the solution, that is, within the sub-thinking path, the outputs of multiple solution sub-models are preliminarily screened and merged, which can not only retain multiple ideas, but also prevent the explosive growth of the number of candidates in the path, thereby achieving richer multi-dimensional thinking. At the same time, by introducing evaluation and heuristic pruning strategies after each round of fusion, low-quality or invalid paths are screened out, and paths with higher potential are further expanded. This dynamic iteration mechanism of "keeping the best and discarding the worst" can continuously optimize the mind tree structure in multiple rounds, improve the accuracy and efficiency of reasoning, and avoid wasting resources on invalid paths. This solves the problem of lack of dynamic iteration and invalid path screening in the reasoning process. The above solutions are all included in the processing of step S120, which is generally referred to as "first-order fusion". The specific implementation method will be described in detail later.

[0042] In one embodiment, step S120: solving the thinking path according to the local solution model, including: solving the sub-model to split the thinking path into at least one sub-thinking path, the sub-thinking path includes multiple sub-problems; arranging the sub-problems in the same sub-thinking path in order, and determining the corresponding solution rounds of the sub-problems between different sub-thinking paths according to the arrangement order; solving the sub-model in parallel to solve the sub-problems corresponding to the same solution round between multiple sub-thinking paths to obtain corresponding sub-reply contents; summarizing all sub-reply contents under the same sub-thinking path to obtain sub-answers of the sub-thinking path; performing first-order fusion processing on the sub-answers between all sub-thinking paths to obtain candidate answers for the thinking path.

[0043] In one embodiment, taking the above example as an example, Figure 3The three feasible route options from Sichuan to Shanghai are three thinking paths. For each thinking path, the local solution model can be used to solve it separately. And according to the previous description of the local solution model, the local solution model is composed of multiple sub-models. Each sub-model has its own configuration process, so when solving the thinking path, the sub-model will split into sub-thinking paths. For example, the sub-model is configured to use private cars to take different roads: sub-model A is configured to take the highway, and sub-model B is configured to take the national highway. Figure 3 Take the middle thinking path as an example. Sub-model A needs to determine how to use highways to connect the routes of Sichuan→Chongqing→Hubei→Anhui→Jiangsu→Shanghai; sub-model B determines the situation of connecting the national highway in this order. Therefore, connecting the path with highways and connecting the path with national highways will further subdivide the thinking path into two sub-thinking paths. At the same time, the sub-thinking path is composed of multiple sub-problems. Taking the sub-thinking path of sub-model A as an example, it can be specifically divided into five sub-problems: how to use highways to connect Sichuan→Chongqing, Chongqing→Hubei, Hubei→Anhui, Anhui→Jiangsu, and Jiangsu→Shanghai. The five sub-problems are arranged in order with each other; similarly, for sub-model B, there are also five similar sub-problems of how to connect with national highways.

[0044] It can be seen that there is a relevant order between the sub-problems, so the corresponding solving round can be determined according to the arrangement order of the sub-problems. For example, for the sub-problem of Sichuan→Chongqing, the sub-problems in sub-model A and the sub-problems in sub-model B can belong to the same solving round. The remaining sub-problems can be arranged according to this rule. It is worth noting that the solving round is not a forced binding of the corresponding order, but represents the corresponding thinking depth. The same thinking depth can be called the same solving round. There is a correlation between the rounds, and each is connected to each other at the level. Therefore, in the solving process, the sub-problems corresponding to the same solving round can be solved in parallel to obtain the corresponding sub-answer content, that is, the problems of the same thinking depth are solved in parallel. In this way, the multi-path parallel exploration characteristics of tree-like thinking are utilized, which is not limited to a single linear reasoning chain, and multiple "thinking paths" are unfolded at the same time.

[0045] In one embodiment, the sub-model is solved in parallel to solve the sub-problems corresponding to the same solution round between multiple sub-thinking paths to obtain corresponding sub-reply contents, including: obtaining the core intent of the request content determined after the analysis operation, and determining the scoring calculation template of the sub-thinking path according to the core intent; after each round of solution, calculating the evaluation score of the sub-reply content according to the scoring calculation template; if the evaluation score is higher than the deep thinking threshold, generating a new sub-problem based on the sub-reply content to be added to the next round for parallel solution; and / or, after each round of solution, calculating the correlation between the sub-problem of the next round and the sub-problem of the current round according to the sub-reply content; if the correlation is lower than the preset correlation threshold, generating a new sub-problem based on the sub-reply content to be added to the next round for parallel solution.

[0046] In one embodiment, the sub-model is solved in parallel to solve the sub-problems corresponding to the same solution round between multiple sub-thinking paths to obtain the corresponding sub-reply content, including: obtaining a scoring calculation template; after each round of solution, calculating the evaluation scores of all sub-reply contents according to the scoring calculation template; if there is a sub-reply content with an evaluation score lower than the retention threshold, then the sub-problem of the next round corresponding to the sub-reply content is marked as a problem to be deleted; and / or, after each round of solution, the number of sub-problems to be solved in parallel in the next round is determined; when the number of sub-problems is greater than a preset upper limit, the evaluation scores of all sub-reply contents in the current round are calculated according to the scoring calculation template; the top N sub-reply contents with the evaluation scores are retained, and the sub-problems of the next round corresponding to the sub-reply contents that are not retained are marked as problems to be deleted; N is a preset integer, N is less than or equal to the preset upper limit; the sub-thinking path with the problem to be deleted as the root is deleted.

[0047] In one embodiment, the method and process of solving the sub-model to obtain the sub-answer content specifically belongs to the scope of the prior art, and needs to be determined specifically according to the specific problem, so it will not be expanded here. Scoring screening requires specific determination of the sub-answer content, so the core intention of the request content determined after the analysis operation is obtained before scoring, and the scoring calculation template of the sub-thinking path is determined according to the core intention. For example, taking the example of Sichuan to Shanghai cited in the previous article as an example, its core intention is navigation, and its scoring template can be the time spent on the journey, or the total toll required, etc. Similarly, for other intentions, the corresponding scoring calculation template is matched. The scoring calculation template can be a pre-configured requirement. For example, for the composition intention, the matching scoring calculation template needs to evaluate the fluency of its writing, the rationality of the wording, etc.; it can also be a unique standard set by the user alone or proposed in addition. For example, in the scenario of generating a composition, if the user needs to reduce or increase rhetoric, this standard can also be included in the scoring calculation template.

[0048] Furthermore, for the specific scoring process, not only can the content itself be processed, but also the sub-answer content can be feature extracted (such as text embedding, answer confidence, logical coherence score, output length, etc.), and the extracted features can be used to calculate the evaluation score of the sub-answer content according to the scoring calculation template. For example, if the problem is a math problem: solve sub-model A to calculate "10", solve sub-model B to calculate "8", solve sub-model C to calculate "12", and extract their intermediate reasoning embedding or step features respectively. Note that this is not to ask for their results, but the process of obtaining the results and some gain information in the middle. The calculation process can use heuristic algorithms to calculate the evaluation score, or it can be calculated by obtaining the real label, and there is no restriction on this. For example, assuming that the request content is a task with a standard answer (such as real labels or annotated data), each sub-answer content and its features can be scored directly. If there is no label, it can be scored based on heuristics (such as matching degree with the topic keywords, similarity with the existing knowledge base, logical consistency, etc.). Thus, digital indicators are provided for the subsequent fusion process to measure which candidate is better or more reasonable. If the result is only a "generation" task, self-supervision or rules can also be used to evaluate the quality of the answer.

[0049] After each round of solving, the evaluation score of the sub-answer content is calculated according to the scoring calculation template, and the evaluation score is used to determine whether a new sub-question needs to be generated in the sub-answer content of the current round. For this, please refer to Figure 4 The schematic diagram is shown in Figure 1. It is worth noting that the newly generated sub-questions may not be just one, but a series of sub-questions, forming a form similar to a sub-thinking path. In other words, on the high-scoring sub-answers, new sub-thinking paths are explored to conduct deeper thinking and obtain richer answers.

[0050] like Figure 4 As shown in the figure, the sub-answers obtained by solving each sub-problem of the original sub-thinking path represented by the blue squares have undergone a total of four rounds of solving processes from top to bottom. When the sub-answers obtained in the first round are scored, the evaluation score is higher than the preset deep thinking threshold, which means that the sub-answers obtained in the first round can be further expanded to generate new sub-problems to be solved simultaneously in the next round of solving. For example, Figure 4In the second round of solving, in addition to the sub-answer content on the original blue sub-thinking path, two additional sub-answer contents were obtained. However, they are distinguished by green and red, and it can be seen that the green sub-answer content also has the sub-answer content of the third round; while the red sub-answer content ends here. The green sub-answer content corresponds to its evaluation score that is higher than the deep thinking threshold again, so it can be deeply thought about again today and evening, and new sub-questions can be excavated to join the third round of parallel solving. The evaluation score corresponding to the red sub-answer content is lower than the deep thinking threshold, which means that this idea is no longer worth deep thinking. In this case, the answer content can be discarded, that is, no new sub-questions will be excavated in the future.

[0051] And we can see that by setting the evaluation score and the deep thinking threshold, Figure 4 For example, originally there were only four sub-reply contents represented by blue blocks. Due to the expansion of deep mining, four additional sub-reply contents were added, which expanded the number of sub-reply contents and generated more comprehensive and rich answers.

[0052] However, it is worth noting that the method provided in this application uses tree-like thinking to expand the breadth of thinking. However, the solution itself still has the following challenges: Redundancy of invalid paths: A large number of generated thinking paths may contain low-quality or invalid solutions. If there is a lack of effective screening mechanism, it will lead to waste of computing resources and degradation of system performance. Path explosion problem: With the increase of reasoning steps, the number of paths may grow exponentially. Traditional methods find it difficult to control computational complexity while ensuring the quality of the results. Insufficient interaction between paths: Each path is often explored independently, lacking information sharing and comprehensive analysis across paths, and failing to fully utilize the potential complementarity between paths.

[0053] At the same time, existing reasoning and fusion methods usually lack dynamic iteration and feedback mechanisms, which are manifested as: Lack of iterative optimization: In complex reasoning tasks, the results of the initial reasoning may have room for improvement, but the existing technology lacks a flexible multi-round iteration mechanism to optimize the solution. Insufficient pruning strategy: Invalid paths will continue to occupy resources in subsequent reasoning, and low-quality paths cannot be effectively pruned based on heuristic evaluation. Unclear improvement direction: In multi-round reasoning, there is a lack of targeted strategies to prioritize the exploration of high-potential paths, which easily leads to invalid exploration iterations.

[0054] To this end, after each round of solving, the correlation between the sub-answer content of the current round and the sub-questions of the next round can be calculated. Correlation represents the degree of association between the sub-answer content or sub-questions. If the correlation is lower than the preset correlation threshold, a new sub-question is generated based on the sub-answer content to be added to the next round for parallel solving. On the contrary, if the correlation is higher than the threshold, it indicates that there is a correlation between the sub-answer content or sub-questions. In the process of solving or subsequent fusion, the mutual influence is considered to improve the potential complementarity of the data. Through the above-mentioned setting, after each round of fusion, evaluation and heuristic pruning strategies are introduced to screen out low-quality or invalid paths, while deeply expanding paths with higher potential. And through this dynamic iteration mechanism of "keeping the best and discarding the worst", the thinking tree structure can be continuously optimized in multiple rounds, the accuracy and efficiency of reasoning can be improved, and resources can be avoided from being wasted on invalid paths.

[0055] In addition to digging into new sub-questions, the quality of each sub-answer content can also be determined. The evaluation scores of all sub-answer contents are calculated according to the scoring calculation template; if the evaluation score of a sub-answer content is lower than the retention threshold, the sub-questions of the next round corresponding to the sub-answer content are marked as questions to be deleted, so that they can be deleted later. This ensures the quality of the sub-answer content obtained in the end.

[0056] As mentioned above, there may be an explosion in the number of branches in tree-like thinking. The explosion in number will lead to an increase in computing power requirements and the generation of a large number of low-quality answers. For this reason, it is necessary to control the thinking branches. In this regard, after each round of solving, the number of sub-problems that need to be solved in parallel in the next round can be determined. When the number of sub-problems is greater than the preset upper limit, the number of branches needs to be controlled to ensure that all the sub-problems solved are high-quality sub-problems. And it is known that the sub-problems of each round are related to the sub-answer content of the previous round, so it can be determined whether the subsequent sub-thinking path needs to be retained based on the evaluation score of the sub-answer content of the previous round. In this regard, the sub-answer content of the current round can be arranged according to the level of its evaluation score, and the sub-questions corresponding to the sub-answer content with the lowest score can be deleted. Only the N sub-answer contents with the highest evaluation score are retained, that is, only the subsequent sub-questions corresponding to the N sub-answer contents are solved in the next round. This ensures that computing power is not wasted and the number of answers is increased. N is a preset integer, and N is less than or equal to the preset upper limit. By setting N to or equal to the preset upper limit, it is not only ensured that the number of sub-problems to be solved in parallel in the next round will not exceed the upper limit of the system's computing power, but also that there is computing power redundancy in subsequent rounds, so that subsequent rounds can continue to conduct in-depth thinking and mining based on the polymer's reply content.

[0057] The specific means for deleting subproblems is to mark the subproblems to be deleted as problems to be deleted. Before starting to solve the next round, the subsequent subproblems with the problem to be deleted as the root, that is, the corresponding sub-thinking path branches, will be deleted from the tree-like thinking. That is, the subproblems after the sub-thinking path where the deleted problem is located will no longer be solved. Dynamic iteration and heuristic pruning are used to ensure high efficiency, screen out low-quality or invalid paths, and deeply expand paths with higher potential. This dynamic iteration mechanism of "keeping the best and discarding the worst" can continuously optimize the thinking tree structure in multiple rounds, improve the accuracy and efficiency of reasoning, and avoid wasting resources on invalid paths.

[0058] In one embodiment, a first-order fusion process is performed on the sub-answers between all sub-thinking paths, including: obtaining all sub-reply contents constituting each sub-answer, and the evaluation scores of the sub-reply contents; extracting key features in the sub-reply contents, the key features being used to characterize the core elements within the sub-reply; performing weighted scoring according to the evaluation scores, and selecting the key features with the highest scores for combination based on the scoring results to construct candidate answers; the candidate answers are composed of multiple key features, and the sum of the evaluation scores of the key features constituting the candidate answers is greater than or equal to the sum of the evaluation scores of all sub-reply contents under any sub-thinking path.

[0059] In one embodiment, as described above, a thinking path passes through a local solution model and will be split into multiple sub-thinking paths by the solution sub-model in the local solution model; each sub-thinking path will obtain a series of sub-answer contents by solving the corresponding solution sub-model. Figure 3 If any of the thinking paths in the above is solved by the local solution model, the following will be obtained: Figure 4 The sub-reply contents are represented by blue and green blocks as shown. However, the sub-reply contents correspond to different sub-thinking paths, and cannot be simply integrated to output answers that meet the logic and output conditions. In other words, the simple fusion method of multiple models in the prior art cannot be adopted. For this reason, the present application performs hierarchical deep fusion within the path for the sub-reply contents of multiple sub-thinking paths under one thinking path, so as to significantly improve the adaptability and reasoning quality to complex multi-step tasks. This process is called first-order fusion processing.

[0060] The first-order fusion processing can be handled by the local aggregator pre-configured in the local solution model. The local aggregator is used to perform preliminary fusion / weighting / voting within the "thinking path": using the aforementioned evaluation scores, and the sub-answer content itself or the features extracted from it as key inputs, the multiple sub-answer contents generated by multiple models in the same path are combined, or the best one is selected. For example, in the example of the math problem in the previous article, we can get: the final output of path 1: after local fusion, we guess the answer = 12, credibility = 0.85; the final output of path 2: after local fusion, the answer = 8, credibility = 0.65; path N and so on. In this way, each thinking path is realized, and the multiple sub-answer contents obtained by the local solution model are fused and output as an answer, which is called the candidate answer corresponding to the thinking path.

[0061] Therefore, in step S120, a tree-like thinking approach is adopted. After problem decomposition, thinking expansion and answer integration, candidate answers for multiple thinking paths are generated in the local situation. The following progress can be achieved: 1. Parallel thinking effectively reduces the probability of failure due to "dead ends" or being limited to a certain wrong reasoning chain. 2. Strong scalability. Each path can be further refined and corrected in subsequent stages, allowing the model to generate more diversified answer candidates. 3. Adapt to complex tasks, such as strategy planning (games, planning tasks, etc.), multi-step reasoning (logic questions, mathematical proofs), or creative writing (multiple storyline directions).

[0062] Step S130: taking the candidate answer set as the input of the global meta-learner, and performing a fusion processing operation on the candidate answer set according to the global meta-learner to obtain a final answer and output it.

[0063] In one embodiment, it can be seen in the previous description that after the intelligent agent thinking decomposition in step S110 and the local solution model local solution and first-order fusion in step S120, a candidate answer set summarized by the candidate answers sought by multiple ideas will be obtained. Each candidate answer in the candidate answer set corresponds to a possible answer under different ideas. Each answer has its own advantages and disadvantages, but each is not enough to constitute the final answer. For this reason, the present application, based on the first-order fusion of step S120, introduces a global meta-learner (such as Stacking) for "second-order fusion". The global meta-learner deeply explores the complementarity between the outputs and features of each path, and comprehensively learns the difference information between different behaviors / models, so as to achieve a more accurate and robust global optimal fusion.

[0064] For a clear description of the second-order fusion process, please refer to Figure 5 .

[0065] Step S210: Obtain feature information of each candidate answer in the candidate answer set.

[0066] Step S220: Retrieve additional feature information according to the request content.

[0067] In one embodiment, in the original Stacking technology processing flow, the global meta-learner processes the prediction results of multiple base learners as input. However, it can be seen that each candidate answer in the candidate answer set is actually composed of sub-answer content, and the reasoning process of the sub-answer content, scoring and other comprehensive data. For this purpose, it is necessary to extract the feature information therein, so as to facilitate the global meta-learner to process. The feature information includes but is not limited to the evaluation score, reasoning process, confidence level, etc. of the first-order fusion stage.

[0068] At the same time, additional feature information can be retrieved based on the request content. Additional feature information is external information related to the request content in the existing database, such as search engine results, external knowledge base matching, conversation context information, etc., to enrich the feature dimension. For example, the question is "Will it rain tomorrow if the humidity is 80%?", where "humidity 80%" is known information in the request content. According to the reasoning process of the first-order fusion stage, the relationship between humidity and rainfall can be determined. However, in reality, rainfall is also related to other data, such as tomorrow's air pressure data mentioned in the context of the chat, previous weather data retrieved from the weather database, and various weather radar images obtained based on the retrieval. The above information, except for humidity itself, and information related to rainfall are all additional feature information. It can be obtained based on the correlation of the obtained results. Therefore, the prediction of a single data surface is limited. If the remaining related data can be obtained in addition to this and processed together, the output result will be more accurate.

[0069] However, it is worth noting that the reference Figure 5 , step S220 of obtaining additional feature information is not actually necessary, but hidden or optional. Because the quality of the additional feature information obtained based on the correlation is usually difficult to guarantee, it may affect the final output answer. Therefore, the scheme described in this application is only in the case of a preferred embodiment; in other implementations, step S220 may not be included, and correspondingly, no additional feature information is input in the subsequent step S230.

[0070] Step S230: using the additional feature information and the feature information as input to a global meta-learner for training and / or inference.

[0071] In one embodiment, the global meta-learner is a model composed of one or more of a linear model, a tree model, a neural network and / or a large language model. The global meta-learner can automatically learn how to weight or combine the results of different paths to obtain a more credible and comprehensive output. The second-order fusion processing of the global meta-learner can refer to the fusion processing of the sub-answer content of multiple sub-thinking paths in the first-order fusion processing. The global meta-learner can perform secondary training or inference on the feature information to weightedly fuse all the feature information, and the final output fusion answer, or an answer distribution (if it is classified / multiple choice), is called the answer to be output, and the answer to be output is composed of a combination of feature information. The answer to be output may also be accompanied by uncertainty estimates (such as confidence intervals, the subjective probability of the model's own answer), to help subsequent links determine whether to do further exploration.

[0072] Step S240: Obtain the answer to be output processed and output by the global meta-learner; and determine whether the answer to be output meets the output condition.

[0073] In one embodiment, the output condition is specifically set according to the specific request content, for example, it may be whether the content quality of the answer to be output reaches a set threshold; or whether sufficient in-depth thinking and sufficient depth of exploration have been carried out during the processing of the first-order fusion and the second-order fusion; it may also be whether the answer obtained in the second-order fusion process is the global optimal.

[0074] If not, then execute step S250: take the answer to be output and the request content as the input of the agent, and return to step S110.

[0075] In one embodiment, if the quality of the global fusion result is not enough or has not reached the set threshold, a new expansion or introduction of a new path will be performed based on the existing optimal path. This will return to step S110 to regenerate the thinking path and iterate again. However, it is worth noting that when the intelligent agent is input again during the iteration process, the sub-reply content or thinking path that determines the reason for not meeting the output condition during the judgment process should be pruned, so as to directly discard those paths with too low scores in the next round. Alternatively, if the global result still has room for improvement or requires deeper reasoning, the optimal path can be retained or the new path can be re-expanded according to the evaluation & pruning strategy, so as to continue iterating until a satisfactory answer is finally obtained.

[0076] If satisfied, execute step S260: mark the answer to be output as the final answer and output it.

[0077] Taken together, this application adopts a hierarchical integration approach in design: first-order fusion is for the integration of multiple models within a single path; second-order fusion is responsible for the joint learning of global multiple paths. This architecture can flexibly embed multiple models and heuristic algorithms, and can also be used with different types of meta-learners for second-order fusion, with good scalability; and can be combined with the mainstream AI technology stack to achieve unified support for tasks of different scales and complexities.

[0078] For easier understanding, this application proposes a specific example to describe the implementation process of the solution provided by this application. Figure 6 . Figure 6 The data processing flow path in the method of the present application is provided, wherein the rounded rectangle is the input request content and the final output answer; the square is the intermediate data in the processing process; and the circle is the data model used in the processing. And Figure 6 The steps are separated by dotted lines, corresponding to the processing process of steps S110 to S130 in the previous text.

[0079] The following is an example of a travel planning scenario to demonstrate the implementation process of the solution provided by this application. Assume that the user inputs: "I want to go to Japan for 5 days, with a budget of US$2,000, and want to visit Tokyo, Kyoto, and Hiroshima. Can you give me a detailed itinerary plan that includes both popular attractions and some local cultural experiences?" After the agent receives the message, it can start the thinking splitting of step S110. The agent will split the planning requirements and may first generate several different travel ideas from the perspective of the outline to obtain three thinking paths: 1. Thinking Path 1: Day 1: Tokyo sightseeing; Day 2: Tokyo -> Kyoto (Shinkansen); Day 3: Cultural experience around Kyoto; Day 4: Kyoto -> Hiroshima, visit the atomic bomb site and park; Day 5: Return to Tokyo, end of the trip.

[0080] 2. Thinking Path 2: Day 1: Tokyo’s major attractions (Asakusa, Akihabara); Day 2: Tokyo cultural exploration (Kimono experience, tea ceremony); Day 3: Tokyo -> Hiroshima, visit Miyajima; Day 4: Hiroshima -> Kyoto; Day 5: Kyoto’s famous temples and Gion historical district, end of the trip.

[0081] 3. Thinking Path 3: Day 1: Go to Kyoto to experience traditional culture first; Day 2: Kyoto -> Osaka or Kobe for transfer; Day 3: Osaka -> Hiroshima; Day 4: Hiroshima -> Tokyo; Day 5: Free travel in Tokyo, the trip ends.

[0082] At this time, the intelligent agent generated several parallel solutions for the same problem "How to arrange the itinerary reasonably?" with different start and end orders, transportation methods, and attractions selection.

[0083] Afterwards, each thought path may be handed over to the local solution model for processing to enter the first-order fusion processing process corresponding to step S120. Figure 6 It is not shown in the figure, but according to the previous description, the local solution model actually consists of multiple solution sub-models with different configurations and / or types. Assume that solution sub-model A focuses on comprehensive information; solution sub-model B focuses on cultural experience; solution sub-model C focuses on convenient transportation. Therefore, even for the arrangement under thinking path 1, different sub-thinking paths will be generated and corresponding sub-answer contents will be obtained, which can be specifically: Solve sub-model A: It arranges detailed information such as hotel recommendations, how to buy subway cards, and a list of popular restaurants. It also gives a budget estimate: about $1,800 (including transportation and accommodation).

[0084] Solution for sub-model B: The description of traditional cultural activities in Kyoto is strengthened, such as copying sutras, kimono dressing, and wagashi making experience. There is less description of modern cultural experiences in Tokyo. Budget estimate: about $1,900.

[0085] Solving sub-model C: Detailed list of money-saving options for JR Pass, night buses / night trains, and a very tight schedule. Budget estimate: about $1,700, but some cultural activities may be sacrificed.

[0086] Similarly, thinking path 2 and thinking path 3 will also generate multiple sub-plans. The sub-responses can be scored later. Based on the request content, the following features can be extracted from the sub-response content for subsequent scoring: budget (numeric value); number of cultural activities (count); whether popular attractions are included (Boolean value); daily itinerary congestion (score); model confidence, etc.

[0087] Because the request content can be determined to be travel planning, and travel planning has no unique correct solution, so this is a task without a standard answer. For this evaluation score calculation template, you can match it to score each plan based on heuristics, such as: feasibility (whether the transportation connection between cities is reasonable); diversity (whether it takes into account modern cities and traditional culture); budget (whether it is close to US$2,000, not overspending or too wasteful). In this way, you can get the score of the response content of each solution sub-model, for example: the solution sub-sub-scheme of solution sub-model A: comprehensive score = 8.5 / 10; the sub-scheme of solution sub-model B: comprehensive score = 7.5 / 10 (high culturality, but lacks the modern experience of Tokyo); the sub-scheme of solution sub-model C: comprehensive score = 7.0 / 10 (transportation is economical but slightly compact).

[0088] After that, the local aggregator in the local solution model can be used to perform first-order fusion output. In the local solution model corresponding to thinking path 1, the three sub-answers obtained by solving sub-models A, B, and C, as well as their characteristics and scores, are used as inputs for preliminary weighting or voting fusion, so as to finally merge / join together a hybrid solution, which is called the candidate answer of thinking path 1: during the day, the cultural experience of solving sub-models A / B is combined with famous attractions; at night, the traffic refers to the money-saving solution of solving sub-model C; the budget is between 1,700 and 1,900 yuan, and the comprehensive score or confidence is the highest.

[0089] Output candidate answers: "After local integration, the five-day itinerary is: the first two days in Tokyo (taking into account popular cultural activities), then take the Shinkansen to Kyoto for 2 nights, experience the tea ceremony & kimono dress-up, and the last day go to Hiroshima to experience history and culture. Estimated cost is US$1,800." Overall score: 0.85.

[0090] Similarly, thinking path 2 and thinking path 3 will also perform the same "first-order fusion" within their own local solution models to generate their own candidate answers. For example, the candidate answer A2 of thinking path 2 has a budget of $1,900 and a comprehensive score of 0.78; the candidate answer A3 of thinking path 3 has a budget of $1,700 and a comprehensive score of 0.80. The sum of these is the candidate answer set.

[0091] The candidate answer set will be used as the input of the global meta-learner to enter the second-order fusion processing process corresponding to step S130. The candidate answer set includes not only the above-mentioned schemes, but also the reasoning process, reasoning score and corresponding features. As mentioned above, additional feature information can also be obtained, such as: real-time air ticket prices / accommodation (possibly connected to a third-party API); season information (for example, August is the summer vacation in Japan, and the number of tourists is large); user personal preferences (such as whether you like historical buildings, whether you need to check in for food, etc.). And package these additional information into new features, and use them together with the candidate answer set as the input of the global meta-learner.

[0092] The global meta-learner (such as linear regression or a small model) performs second-order fusion processing, using the feature information of the candidate answer set and additional feature information as training data to learn: In previous user feedback or example data, which features best indicate whether the trip is good? How to assign weights to these three options, or directly select the best one? After calculation or training, the global meta-learner may determine that the "candidate answer of thinking path 1" has the highest score, and the external features (convenient transportation, rich attractions) are also good, so it is the best.

[0093] And before output, the global meta-learner can check whether further refinement is needed. If the user is not satisfied with the plan, or wants a lower budget / more cities, he can go back to step S110 again, and perform the next round of expansion based on the well-performing parts of "Thinking Path 1" and "Thinking Path 2", and cut off the poorly performing paths or sub-models. Let different ideas or algorithms complement each other to obtain more robust and accurate results. If the answer to be output already meets the output conditions, a global optimal itinerary can be given directly, and the following will be given: uncertainty score (such as confidence interval 0.85~0.9) The satisfaction of the plan with user needs (budget, location, cultural experience) is such as: "Recommended plan: 5-day itinerary, costing US$1,800~1,900, convenient transportation connection, and both popular attractions and cultural activities; satisfaction score 0.87." Finally, the optimal solution selected after the second-order fusion is output to the user and may be presented in the form of text / list / chart on the front end: “Travel itinerary: Day 1: Arrive in Tokyo->Visit Sensoji Temple, stroll around Akihabara, and experience izakaya culture at night; Day 2: Deepen Tokyo cultural exploration, choose tea ceremony / kimono / manga cafe, and take the Shinkansen to Kyoto in the evening; Day 3: Full day in Kyoto, Kiyomizu Temple, Arashiyama Bamboo Grove, kimono experience, etc.; Day 4: Kyoto->Hiroshima, visit the atomic bomb site in the morning, go to Miyajima to see the shrine and taste seafood in the afternoon; Day 5: Return to Tokyo, end of the trip.” From the above description, we can see that ToT can develop multiple possible routes in the thinking dimension of a problem; Stacking can integrate the insights of multiple different experts or models in a prediction or decision. Putting them together is a "double insurance" way of thinking: 1. Use ToT to develop a thinking path: Generate multiple possible solution directions or ideas for each angle or solution step of the problem. Each idea is like "Expert A", "Expert B", "Expert C"... except that these "experts" are candidate solutions formed in different directions of thinking chains.

[0094] 2. Use Stacking to fuse the results of these paths: Summarize the final answer (or intermediate analysis conclusion, confidence score, etc.) of each thinking path into a "fuser". The "fuser" is like a commander-in-chief, who will learn how to optimally refer to and integrate these different ideas to make the final conclusion more robust.

[0095] In other words, ToT+Stacking allows the system to explore widely at the thinking level (avoiding taking only one path) while deeply integrating at the result level (not just a veto or simple averaging, but intelligent weighted integration).

[0096] Therefore, the present application can use the intelligent agent to obtain the request content and convert it into multiple thinking paths, which are then handed over to the local solution model for parallel solution to obtain a variety of candidate answer sets; the candidate answer set is then used as the input of the global meta-learner for global learning, thereby analyzing the pros and cons of the answers under multiple thinking paths, so as to finally output the final answer with the best performance and the most expected after multiple considerations. The method provided by the present application has high application value in the fields of multi-model collaboration, task planning, creative text generation, strategic reasoning, etc., and can cope with uncertain and high-dimensional task scenarios, providing a new idea and technical path to solve the limitations of current large-scale language models. Through hierarchical design, this patent can fully utilize the generation and understanding capabilities brought by large-scale language models when solving high-dimensional and multi-step reasoning problems, and can combine the advantages of integrated learning to achieve a more flexible, accurate and efficient problem-solving process. It has broad application potential in a variety of scenarios such as complex question and answer, strategy planning, text generation, scientific research reasoning, etc., and provides important technical support and innovative ideas for the development of next-generation intelligent systems.

[0097] Figure 7 FIG. 1 shows an internal structure diagram of a computer device in an embodiment. The computer device may be a terminal or a server. Figure 7 As shown, the computer device includes a processor, a memory and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement the intelligent decision-making method. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement the intelligent decision-making method. Those skilled in the art can understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0098] In one embodiment, the present application further proposes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the method described in any of the aforementioned embodiments.

[0099] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0100] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. An intelligent decision-making method, characterized in that: The steps include: When the agent obtains the request content, the agent performs an analysis operation on the request content to separate at least one thinking path; Taking each of the thought paths as input of a local solution model, solving the thought paths according to the local solution model to obtain corresponding candidate answers; summarizing the candidate answers corresponding to all the thought paths to obtain a candidate answer set; The candidate answer set is used as an input of a global meta-learner, and after the global meta-learner performs a fusion processing operation on the candidate answer set, a final answer is obtained and output.

2. The intelligent decision-making method according to claim 1, characterized in that: The agent performs an analysis operation on the request content, including: Determine a content attribute of the request content, where the content attribute includes at least one of text, picture, video, and audio; Determining a feature extraction tool according to the content attributes, and extracting key features from the request content using the feature extraction tool; Analyze the key features to determine the core intent of the request content; determine at least one reply sub-thought based on the key features and the core intent; The answer sub-thoughts are combined to obtain the thinking path.

3. The intelligent decision-making method according to claim 1, characterized in that: The local solution model is composed of multiple solution sub-models; Solving the thinking path according to the local solution model includes: The solving sub-model splits the thinking path into at least one sub-thinking path, and the sub-thinking path includes multiple sub-problems; the sub-problems in the same sub-thinking path are arranged in order, and the sub-problems between different sub-thinking paths determine the corresponding solving rounds according to the arrangement order; The solving sub-model solves the sub-problems corresponding to the same solving round between the multiple sub-thinking paths in parallel to obtain corresponding sub-answer contents; summarizes all the sub-answer contents under the same sub-thinking path to obtain the sub-answer of the sub-thinking path; A first-order fusion process is performed on the sub-answers between all the sub-thinking paths to obtain candidate answers for the thinking paths.

4. The intelligent decision-making method according to claim 3, characterized in that: The solving sub-model solves the sub-problems corresponding to the same solving round between the plurality of sub-thinking paths in parallel to obtain corresponding sub-answer contents, including: Obtaining the core intent of the request content determined after the analysis operation, and determining the scoring calculation template of the sub-thinking path according to the core intent; after each round of solving, calculating the evaluation score of the sub-answer content according to the scoring calculation template; if the evaluation score is higher than the deep thinking threshold, generating a new sub-question according to the sub-answer content to add to the next round for parallel solving; and / or, After each round of solving, the correlation between the sub-answer content of the current round and the sub-problem of the next round is calculated; if the correlation is lower than the preset correlation threshold, a new sub-problem is generated based on the sub-answer content to be added to the next round for parallel solving.

5. The intelligent decision-making method according to claim 3, characterized in that: The solving sub-model solves the sub-problems corresponding to the same solving round between the plurality of sub-thinking paths in parallel to obtain corresponding sub-answer contents, including: Get the scoring calculation template; After each round of solving, the evaluation scores of all the sub-answer contents are calculated according to the scoring calculation template; if there is a sub-answer content whose evaluation score is lower than the retention threshold, the sub-question of the next round corresponding to the sub-answer content is marked as a question to be deleted; and / or, After each round of solving, determine the number of sub-problems to be solved in parallel in the next round; when the number of sub-problems is greater than the preset upper limit, calculate the evaluation scores of all the sub-answer contents in the current round according to the scoring calculation template; retain the N sub-answer contents with the highest evaluation scores, and mark the sub-problems of the next round corresponding to the sub-answer contents that are not retained as the problems to be deleted; N is a preset integer, which is less than or equal to the preset upper limit; Delete the sub-thinking path with the problem to be deleted as the root.

6. The intelligent decision-making method according to claim 3, characterized in that: The first-order fusion processing is performed on the sub-answers between all the sub-thinking paths, including: Obtain all sub-answer contents constituting each sub-answer and evaluation scores of the sub-answer contents; extract key features in the sub-answer contents, wherein the key features are used to characterize core elements in the sub-answers; A weighted score is performed according to the evaluation scores, and key features with the highest scores are selected and combined according to the scoring results to construct the candidate answer; the candidate answer is composed of multiple key features, and the sum of the evaluation scores of the key features constituting the candidate answer is greater than or equal to the sum of the evaluation scores of all the sub-reply contents under any of the sub-thinking paths.

7. The intelligent decision-making method according to claim 1, characterized in that: After the global meta-learner performs a fusion processing operation on the candidate answer set, the method includes: Acquire feature information of each candidate answer in the candidate answer set, and use the feature information as input of the global meta-learner for processing, wherein the global meta-learner is one or more constituent models of a linear model, a tree model, a neural network and / or a large language model; The global meta-learner trains and / or infers the feature information to weightedly fuse all the feature information, and processes and outputs the answer to be output, wherein the answer to be output is composed of a combination of the feature information; Determine whether the answer to be output meets the output condition; If not satisfied, the answer to be output and the request content are used as input of the agent to iterate again; If satisfied, the answer to be output is marked as the final answer.

8. The intelligent decision-making method according to claim 7, characterized in that: The global meta-learner trains and / or infers the feature information, including: Retrieving and acquiring additional feature information according to the request content, where the additional feature information is external information related to the request content in an existing database; The additional feature information and the feature information are used as inputs of the global meta-learner for training and / or inference.

9. A computer device, characterized in that: including a processor and a memory; The processor is configured to execute the computer program stored in the memory to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images