Emotion-aware multi-tool adaptive agent selection method and system

Abstract

The application provides an emotion perception multi-tool adaptive intelligent agent selection method and system, and belongs to the field of artificial intelligence, and comprises the following steps: S1, generating an enhanced query representation vector according to a user input original query and a user historical interaction record; S2, a planner performing task decomposition based on the enhanced query representation vector to generate a sub-target sequence; S3, a reasoner generating an explicit thinking chain; a Thompson sampling tool selector calculating the comprehensive score of a tool and selecting the tool with the highest comprehensive score for execution; S4, a reflector evaluating the execution result of the current tool to calculate a quality score; S5, calculating the reward of the current round of reasoning and updating the distribution parameters of the called tool, so that the tool selection strategy continuously converges to the optimal strategy after multiple rounds of iteration. The method constructs a four-stage intelligent agent circulation architecture of perception-planning-execution-reflection, and realizes the organic integration of autonomous decision-making and error recovery.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and specifically relates to a method and system for selecting an emotion-perceiving multi-tool adaptive intelligent agent. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent agent systems based on large language models have been widely applied to natural language processing tasks. In existing technologies, Retrieval Augmented Generation (RAG) architectures improve the accuracy of generated content by introducing external knowledge bases, reasoning-action collaborative frameworks such as ReAct solve complex tasks through iterative tool calls, and tool learning platforms such as ToolBench provide standardized tool call interfaces.

[0003] However, the aforementioned existing technical solutions still have the following technical shortcomings in terms of tool selection strategy:

[0004] Existing search enhancement generation systems (such as RAG, ReAct, ToolBench, etc.) have the following main shortcomings in their tool selection strategies:

[0005] 1. Rigid tool selection mechanism. Existing methods mostly use fixed rules or greedy strategies to select external tools, which cannot be dynamically adjusted according to the user's emotional state and preferences, resulting in low tool utilization and unstable response quality.

[0006] 2. Emotional information is ignored. Existing intelligent agent systems generally treat user input as pure semantic text, ignoring the emotional inclinations and personalized preferences contained within, and thus failing to provide differentiated responses that meet the user's emotional needs.

[0007] 3. Imbalance between exploration and utilization. Existing reinforcement learning methods are inefficient in the tool selection process, easily getting trapped in local optima, and unable to achieve the optimal tool combination within a limited number of inference steps.

[0008] 4. Lack of preference alignment. Existing methods lack a modeling and optimization mechanism for users' long-term preferences. Each inference is performed independently, making it impossible to accumulate preference knowledge from historical interactions.

[0009] 5. Lack of reflection mechanism. Most existing intelligent agent systems use single forward reasoning. When tool calls fail or return low-quality results, they cannot identify errors and replan, leading to the accumulation of errors. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a method for selecting an emotion-aware, multi-tool adaptive intelligent agent, comprising the following steps:

[0011] Step S1: Based on the user's original query input and the user's historical interaction records, extract semantic feature vectors, sentiment polarity feature vectors, and user preference vectors, and generate an enhanced query representation vector through weighted fusion;

[0012] Step S2: The planner decomposes the task based on the enhanced query representation vector to generate a sequence of sub-objectives;

[0013] Step S3: The inference engine generates an explicit thought chain based on working memory, the current sub-goal, and the available working set; the Thompson sampling tool selector calculates the overall score of each tool based on the distribution parameters of the historical rewards of each tool, and selects the tool with the highest overall score to execute.

[0014] Step S4: The reflector evaluates the execution result of the current tool and calculates the quality score. If it is lower than the threshold, the planner regenerates the sub-target sequence based on the historical trajectory in working memory and then executes step S3; otherwise, it executes step S5.

[0015] Step S5: Calculate the reward for this round of inference and update the distribution parameters of the invoked tool with the reward. After multiple rounds of iteration, the tool selection strategy will continue to converge toward the optimal strategy.

[0016] Beneficial effects:

[0017] 1. This invention pioneers an emotion perception tool selection mechanism, which encodes the user's emotional state into a continuous vector and injects it into the tool selection decision process, enabling the agent to select the most suitable external tool based on emotional tendencies, thereby improving the emotional adaptability of the response.

[0018] 2. This invention introduces the Thompson sampling tool selector, which combines the Bayesian probability model with the historical reward signal of the tool, to achieve a dynamic balance between exploring new tools and utilizing known efficient tools, thereby continuously improving the quality of tool selection decisions.

[0019] 3. This invention designs a complete four-stage intelligent agent loop architecture of perception-planning-execution-reflection. When the quality of the execution result is lower than the threshold, reflection and replanning are automatically triggered, realizing the organic integration of autonomous decision-making and error recovery.

[0020] 4. This invention integrates the TAPO four-stage preference optimization framework, extending preference learning from the response level to the tool sequence level, and achieving end-to-end optimization through four progressive stages: instruction fine-tuning, tool information optimization, tool robustness optimization, and execution quality optimization.

[0021] 5. This invention uses a working memory mechanism to store the complete reasoning trajectory, supports cross-step decision-making and context continuation, and enhances consistency and coherence in multi-turn interactions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the process of selecting an emotion-perceiving multi-tool adaptive intelligent agent according to the present invention;

[0023] Figure 2 This is a schematic diagram of the architecture of the method of the present invention;

[0024] Figure 3 This is a structural block diagram of an emotion-perceiving multi-tool adaptive intelligent agent selection system according to the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0026] Example 1

[0027] like Figure 1 As shown in the figure, an embodiment of the present invention provides an emotion-aware multi-tool adaptive agent selection method, which includes the following steps:

[0028] Step S1: Based on the user's original query input and the user's historical interaction records, extract semantic feature vectors, sentiment polarity feature vectors, and user preference vectors, and generate an enhanced query representation vector through weighted fusion;

[0029] Step S2: The planner decomposes the task based on the enhanced query representation vector to generate a sequence of sub-objectives;

[0030] Step S3: The inference engine generates an explicit thought chain based on working memory, the current sub-goal, and the available working set; the Thompson sampling tool selector calculates the overall score of each tool based on the distribution parameters of the historical rewards of each tool, and selects the tool with the highest overall score to execute.

[0031] Step S4: The reflector evaluates the execution result of the current tool and calculates the quality score. If it is lower than the threshold, the planner regenerates the sub-target sequence based on the historical trajectory in working memory and then executes step S3; otherwise, it executes step S5.

[0032] Step S5: Calculate the reward for this round of inference and update the distribution parameters of the invoked tool with the reward. After multiple rounds of iteration, the tool selection strategy will continue to converge toward the optimal strategy.

[0033] In one embodiment, step S1 above: extracting semantic feature vectors, sentiment polarity feature vectors, and user preference vectors based on the user's original query input and user historical interaction records, and generating an enhanced query representation vector through weighted fusion, specifically includes:

[0034] Step S11: Extract the semantic feature vector of the original query. ;

[0035] After receiving the user's original query input in natural language, the semantic feature vector is extracted through a pre-trained language model;

[0036] Step S12: Through the emotion encoding function extract Emotional vector Used to quantify a user's current emotional state (e.g., positive, negative, neutral, and intensity):

[0037] ;

[0038] in, Let d be the sentiment embedding matrix, and d represent the dimension of the hidden state of each token. Let i be the hidden states of the i tokens of q. L represents The number of tokens; It is the softmax activation function; For learnable projection matrices, It is an emotional vector The dimension;

[0039] Step S13: Through the preference encoding function Extracting preference vectors from user's historical interaction records ;

[0040] ;

[0041] Where u represents the user; Let j be the semantic representation of the j-th historical interaction record of user u. ; For the preferred projection matrix, Let p be the dimension of the preference vector.

[0042] Step S14: Construct an enhanced query representation ;in, and These are learnable weight parameters.

[0043] In one embodiment, step S2 above: the planner decomposes the task based on the enhanced query representation vector to generate a sequence of sub-objectives, specifically including:

[0044] Based on semantic feature vectors and context Construct sub-target sequences This decomposes the original query into a sequence of executable sub-objectives, where N is the number of sub-objectives, calculated using the following formula:

[0045] ;

[0046] ;

[0047] in, Indicates query q and context The complexity formed by the combination Indicates rounding down; Indicates taking the length; , , These are weighting coefficients, corresponding to the contribution ratios of query length, context entropy, and tool diversity in complexity calculation, respectively. The information entropy of context C is used to measure the uncertainty or richness of contextual information. This indicates the number of tools contained in the available toolset T; Maximum allowed query length;

[0048] The available toolset T may include, but is not limited to, the following tools:

[0049] t1: retrieve_atomic — retrieves common sense reasoning triples from the ATOMIC knowledge graph. The input is a natural language query, and the output is a set of relevant common sense triples.

[0050] t2: retrieve_conceptnet — retrieves semantic association knowledge from ConceptNet. The input is concept words or phrases, and the output is the associated concepts and their relationship types.

[0051] t3: sentiment_analyze — Analyzes the sentiment polarity of text; the input is a text fragment, and the output is the sentiment distribution.

[0052] t4: knowledge_verify — Verifies the consistency between the generated content and the knowledge base. The input is the statement to be verified, and the output is the consistency score ∈ [0,1].

[0053] t5: fact_check — checks the factual accuracy of the answer. The input is the question-answer pair, and the output is the factual accuracy rate and supporting evidence.

[0054] t6: search_web — Web search to obtain real-time information. The input is search keywords, and the output is a list of relevant web page summaries.

[0055] t7: summarize — Long text summarization, input is long text and target length, output is compressed summary;

[0056] t8: qa_generate — Generates question-and-answer answers based on context. The input is the question q and the context C, and the output is the natural language answer and the thought process chain of the current step. It is used to determine the direction of the next action.

[0057] Each sub-goal can be executed independently. Each sub-goal includes: a goal description, the type of tools required, and the expected output format. The generation process is as follows:

[0058] ;

[0059] in, The decoding and generation function of a large language model generates a text sequence autoregressively after input prompts; The planning prompt template includes instruction formats and examples for task decomposition, which are used to guide LLM in generating structured sub-goals.

[0060] In one embodiment, step S3 above: the inferencer generates an explicit thought chain based on working memory, the current sub-goal, and the available working set; the Thompson sampling tool selector calculates the comprehensive score of each tool based on the distribution parameters of the historical rewards of each tool, and selects the tool with the highest comprehensive score for execution, specifically including:

[0061] Step S31: The inference engine uses working memory Current sub-target and available toolsets Generate explicit thought chains Among them, working memory Storing complete inference trajectories to support cross-step decision-making ; For the thought chain sequence, For action sequences, For observation of the sequence;

[0062] Step S32: Use the Thompson sampling tool selector to achieve dynamic tool selection based on a Bayesian probability model, specifically including:

[0063] Step S321: Each tool Maintaining Beta distribution parameters Record historical rewards, among which, This indicates the number of times tool k has received a positive reward in history. Indicates the number of historical negative rewards; all tools at the time of the first call. and Initialize to 1;

[0064] Step S322: Independently sample utility estimates for each tool ,satisfy ;

[0065] Step S323: Combine tools to select information content rating Calculate the overall score ;in, This is used to measure the contribution of tool k to the information of the current query; its value is the relevance score of tool k to q multiplied by the historical confidence score of the tool.

[0066] ;

[0067] in, and The values ​​are all in the range [0,1]. The value also takes the range [0,1]; among which, the correlation Through calculation tools Embedding vectors and Cosine similarity of the embedded vectors, confidence level Based on the historical success rate of the tool (initially set to 0.5);

[0068] The calculation process is as follows:

[0069] (1) Obtain tool representation: Use the name and function description of tool k to concatenate a short text. (For example, "Weather query tool: Get the current weather by city name").

[0070] (2) Get the query representation: directly use the text of q.

[0071] (3) Encoding: The embedding vectors emb(e) are obtained using the same embedding model (e.g., sentence-transformers / all-MiniLM-L6-v2). ) and emb(q).

[0072] (4) Calculate emb( cosine similarity of ) and emb(q) as .

[0073] Step S324: Select The tool corresponding to the highest value will be executed.

[0074] Step S33: After the tool finishes execution, the results are stored in working memory.

[0075] In one embodiment, step S4 above: The reflector evaluates the execution result of the current tool and calculates a quality score. If the score is lower than a threshold, the planner regenerates the sub-objective sequence based on the historical trajectory in working memory and then executes step S3; otherwise, step S5 is executed, which specifically includes:

[0076] Step S41: Observation of the current tool t Calculate its quality score ;

[0077] ;

[0078] in, For correlation: , This indicates the observations from the current tool. Convert to an embedding vector. Convert the query q into an embedding vector;

[0079] For completeness: ; Current tool observation The number of sub-objectives that have been covered / satisfied. This represents the total number of sub-objectives after the current task is decomposed.

[0080] Confidence level: , It is the historical confidence score of tool k, used to measure the probability that tool k has returned reliable or successful results when it was called in the past. =Number of times tool k was successfully invoked / Total number of times tool k was invoked;

[0081] , , The adjustable weighting coefficients corresponding to relevance, completeness, and confidence are set to 0.4, 0.35, and 0.25 respectively in this embodiment of the invention.

[0082] Step S42: When the quality score is below the threshold Then based on working memory Triggering replanning Generate new sub-target sequences and repeat steps S3-S4 until the obtained quality score is higher than the threshold. Then proceed to step S5, in this embodiment of the invention, the threshold is... Set it to 0.6.

[0083] In one embodiment, step S5 above—calculating the reward for this round of inference and updating the distribution parameters of the invoked tool with the reward—ensures that the tool selection strategy continuously converges to the optimal strategy after multiple iterations. Specifically, this includes:

[0084] Step S51: Calculate the reward ;

[0085] in, , , , These are adjustable weighting coefficients for instruction compliance rewards, tool information rewards, robustness rewards, and execution quality rewards, respectively.

[0086] Rewards for following instructions: ; yes Fraction; It is the response generated by the model; This is the desired output format;

[0087] Rewards for informational tools: ;

[0088] For robust rewards: , The score is the model response to the input with added perturbations. Scoring the model response under undisturbed input;

[0089] To implement quality awards: , The gold standard answer is the correct answer as indicated by human annotation.

[0090] Step S52: Update the reward r of this round of inference to the Beta distribution parameters of the invoked tool, where positive feedback negative feedback .

[0091] Finally, through multiple iterations of the above steps, the tool selection strategy of the entire system continuously converges towards the optimal strategy, and finally the optimal tool selection strategy is obtained.

[0092] Figure 2The diagram illustrates the architecture of the method of this invention. The perception module receives the user's original query and simultaneously feeds it into the sentiment encoder and preference encoder, which output sentiment features and user preference representations, respectively. These two representations are combined with the original query to generate a sentiment-preference-enhanced query. Document-related context fragments provided by knowledge sources are retrieved and stored in working memory. The enhanced query is also sent to working memory, forming the current memory content together with the retrieved context. The output of working memory enters the core of the agent. First, the planner decomposes the task into sub-goals. The inferencer generates thought steps based on the sub-goals. The tool enhancer uses an enhanced knowledge graph to assist decision-making and passes the results to a Thompson sampling-based tool selector. The Thompson sampling tool selector maintains the Thompson sampling Beta distribution of each tool and performs adaptive tool selection by combining tool registration candidates and the toolset. The selected tool performs the corresponding operation. The execution results enter the reflector. If the evaluation quality is below a threshold, reflection is triggered, and a posterior update is performed along the blue path. Parameters are updated according to the Beta distribution, and replanning is initiated, returning to the planner to re-decompose the sub-objectives. If the evaluation quality is not below the threshold, the final response generation module is entered. This module integrates mechanisms such as instruction fine-tuning, tool selection informativeness, tool selection robustness, and tool execution quality to generate and output the final response. Throughout the entire process, the learning and update phase continuously receives external reward signals, transforming reward feedback into parameter updates according to the Beta distribution, and continuously optimizing subsequent tool selection strategies.

[0093] Example 2

[0094] like Figure 3 As shown, this embodiment of the invention provides an emotion-aware multi-tool adaptive intelligent agent selection system, including the following modules:

[0095] The perception module 61 is used to extract semantic feature vectors, sentiment polarity feature vectors and user preference vectors based on the user's original query input and the user's historical interaction records, and generate an enhanced query representation vector through weighted fusion;

[0096] Planning module 62 is used by the planner to decompose the task based on the enhanced query representation vector and generate a sequence of sub-objectives;

[0097] Action module 63 is used by the inference engine to generate an explicit thought chain based on working memory, the current sub-goal, and the available working set; the Thompson sampling tool selector calculates the overall score of each tool based on the distribution parameters of the historical rewards of each tool, and selects the tool with the highest overall score to execute.

[0098] The reflection module 64 is used by the reflector to evaluate the execution results of the current tool and calculate the quality score. If it is lower than the threshold, the planner will regenerate the sub-objective sequence based on the historical trajectory in the working memory and then execute the action module; otherwise, the optimization module will be executed.

[0099] The optimization module 65 is used to calculate the reward for this round of inference and update the distribution parameters of the called tool with the reward. After multiple rounds of iteration, the tool selection strategy continues to converge toward the optimal strategy.

[0100] An emotion-aware multi-tool adaptive agent selection device includes one or more electronic devices, wherein the one or more electronic devices are used to implement the emotion-aware multi-tool adaptive agent selection method.

[0101] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement an emotion-aware multi-tool adaptive agent selection method.

[0102] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for selecting an emotion-perceiving, multi-tool adaptive intelligent agent, characterized in that, include: Step S1: Based on the user's original query input and the user's historical interaction records, extract semantic feature vectors, sentiment polarity feature vectors, and user preference vectors, and generate an enhanced query representation vector through weighted fusion; Step S2: The planner decomposes the task based on the enhanced query representation vector to generate a sequence of sub-objectives; Step S3: The inference engine generates an explicit thought chain based on working memory, the current sub-goal, and the available working set; Thompson's sampling tool selector calculates the overall score of each tool based on the distribution parameters of its historical rewards, and selects the tool with the highest overall score to execute. Step S4: The reflector evaluates the execution result of the current tool and calculates the quality score. If it is lower than the threshold, the planner regenerates the sub-target sequence based on the historical trajectory in working memory and then executes step S3; otherwise, it executes step S5. include: Step S41: Observation of the current tool t Calculate its quality score ; ； in, For correlation: , This indicates the observations from the current tool. Convert to an embedded vector. , convert the query q into an embedding vector; For completeness: ; Current tool observation The number of sub-objectives that have been covered / satisfied. This represents the total number of sub-objectives after the current task is decomposed. Confidence level: , It is the historical confidence level of tool k; These are adjustable weighting coefficients corresponding to relevance, completeness, and confidence, respectively. Step S42: When the quality score is below the threshold, then based on working memory... Triggering replanning Generate new sub-target sequences and execute steps S3-S4 until the quality score is higher than the threshold, then proceed to step S5; Step S5: Calculate the reward for this round of inference and update the distribution parameters of the invoked tool with the reward. After multiple rounds of iteration, the tool selection strategy will continue to converge toward the optimal strategy.

2. The emotion-perception multi-tool adaptive intelligent agent selection method according to claim 1, characterized in that, Step S1: Based on the user's original query input and the user's historical interaction records, extract semantic feature vectors, sentiment polarity feature vectors, and user preference vectors, and generate an enhanced query representation vector through weighted fusion, specifically including: Step S11: Extract the semantic feature vector of the original query. ; Step S12: Through the emotion encoding function extract Emotional vector : ； in, Let d be the sentiment embedding matrix, and d represent the dimension of the hidden state of each token. Let i be the hidden states of the i tokens of q. L represents The number of tokens; It is the softmax activation function; For learnable projection matrices, It is an emotional vector The dimension; Step S13: Through the preference encoding function Extract the preference vector from the user's historical interaction records. ; ； Where u represents the user; Let j be the semantic representation of the j-th historical interaction record of user u. ; For the preferred projection matrix, Let p be the dimension of the preference vector. Step S14: Construct an enhanced query representation ;in, and These are learnable weight parameters.

3. The emotion-perception multi-tool adaptive intelligent agent selection method according to claim 2, characterized in that, Step S2: The planner decomposes the task based on the enhanced query representation vector to generate a sequence of sub-objectives, specifically including: Based on semantic feature vectors and context Construct sub-target sequences This decomposes the original query into a sequence of executable sub-objectives, where N is the number of sub-objectives, calculated using the following formula: ； ； in, Indicates query q and context The complexity formed by the combination Indicates rounding down; Indicates taking the length; , , These are weighting coefficients, corresponding to the contribution ratios of query length, context entropy, and tool diversity in complexity calculation, respectively. Representing context Information entropy is used to measure the uncertainty or richness of contextual information; L_max represents the number of tools contained in the available toolset T; L_max is the maximum allowed query length. Each sub-objective can be executed independently, and its generation process is as follows: ； in, The decoding and generation function of a large language model generates a text sequence through autoregression after input prompts; The planning prompt template includes instruction formats and examples for task decomposition, which guides LLM in generating structured sub-goals.

4. The emotion-perception multi-tool adaptive intelligent agent selection method according to claim 3, characterized in that, Step S3: The inference engine generates an explicit thought chain based on working memory, the current sub-goal, and the available working set; the Thompson sampling tool selector calculates the comprehensive score of each tool based on the distribution parameters of its historical rewards, and selects the tool with the highest comprehensive score for execution, specifically including: Step S31: The inference engine uses working memory Current sub-target and available toolsets Generate explicit thought chains ; Step S32: Use the Thompson sampling tool selector to achieve dynamic tool selection based on a Bayesian probability model; Step S33: After the tool finishes execution, the results are stored in working memory.

5. The emotion-perception multi-tool adaptive intelligent agent selection method according to claim 4, characterized in that, Step S32: Using the Thompson sampling tool selector to achieve dynamic tool selection based on a Bayesian probability model, specifically includes: Step S321: Each tool Maintaining Beta distribution parameters Record historical rewards, among which This indicates the number of times tool k has received a positive reward in history. Indicates the number of negative rewards in history; Step S322: Independently sample utility estimates for each tool ,satisfy ; Step S323: Combine tools to select information content rating Calculate the overall score ;in, This is used to measure the contribution of tool k to the information of the current query; its value is the relevance score of tool k to q multiplied by the historical confidence score of the tool. ； in, and The values ​​are all in the range [0,1]. The value also takes place in the range [0,1]; among which, the correlation... Through calculation tools Embedding vectors and Cosine similarity of the embedded vectors, confidence level Based on the tool's historical call success rate; Step S324: Select The tool corresponding to the highest value will be executed.

6. The emotion-perception multi-tool adaptive intelligent agent selection method according to claim 3, characterized in that, Step S5: Calculate the reward for this round of inference and update the distribution parameters of the invoked tool with the reward. After multiple iterations, the tool selection strategy continuously converges to the optimal strategy. Specifically, this includes: Step S51: Calculate the reward ; in, , , , These are adjustable weighting coefficients for instruction compliance rewards, tool information rewards, robustness rewards, and execution quality rewards, respectively. Rewards for following instructions: ; yes Fraction; It is the response generated by the model; This is the desired output format; Rewards for informational tools: ; For robust rewards: , The score is the model response to the input with added perturbations. Scoring the model response under undisturbed input; To implement quality awards: , The gold standard answer, i.e., the correct answer as marked by human annotation; Step S52: Update the reward r of this round of inference to the Beta distribution parameters of the invoked tool, where positive feedback negative feedback .

7. An emotion-perceiving multi-tool adaptive intelligent agent selection system, characterized in that, Includes the following modules: The perception module is used to extract semantic feature vectors, sentiment polarity feature vectors, and user preference vectors based on the user's original query input and the user's historical interaction records, and generate an enhanced query representation vector through weighted fusion. The planning module is used by the planner to decompose the task based on the enhanced query representation vector and generate a sequence of sub-objectives; The action module is used by the inference engine to generate explicit thought chains based on working memory, the current sub-goal, and the available working set; Thompson's sampling tool selector calculates the overall score of each tool based on the distribution parameters of its historical rewards, and selects the tool with the highest overall score to execute. The reflection module is used by the reflector to evaluate the execution results of the current tool and calculate a quality score. If the score is below a threshold, the planner regenerates the sub-objective sequence based on the historical trajectory in working memory and then executes the action module; otherwise, the optimization module is executed. This includes: Current observation of tool t Calculate its quality score ; ； in, For correlation: , This indicates the observations from the current tool. Convert to an embedded vector. , convert the query q into an embedding vector; For completeness: ; Current tool observation The number of sub-objectives that have been covered / satisfied. This represents the total number of sub-objectives after the current task is decomposed. Confidence level: , It is the historical confidence level of tool k; These are adjustable weighting coefficients corresponding to relevance, completeness, and confidence, respectively. When the quality score is below the threshold, it is based on working memory. Triggering replanning Generate a new sub-target sequence and execute the actions and reflections modules until the quality score is higher than the threshold, then proceed to execute the optimization module. The optimization module is used to calculate the reward for this round of inference and update the distribution parameters of the invoked tool with the reward. After multiple rounds of iteration, the tool selection strategy continues to converge toward the optimal strategy.

8. An emotion-perceiving multi-tool adaptive intelligent agent selection device, characterized in that, It includes one or more electronic devices, wherein the one or more electronic devices are used to implement the method of any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 6.

Images