A method for constructing a multimodal large model for the securities industry in an ICT-enabled executable environment

Abstract

The present application relates to a method for constructing a multimodal large model of the securities industry in an ICT executable environment, comprising: a first agent module and a second agent module, based on long-term memory and short-term memory obtained from a memory module, conduct a game between the first analysis result and the second analysis result obtained after analyzing the same task; if the first agent module and the second agent module fail to reach an agreement after the game, the third agent module performs weighted processing on the opinions discussed by the first agent module and the second agent module based on preset weights to obtain a game result; the third agent module generates an initial solution based on the task to be gambled, and obtains suggestions for the initial solution from the first agent module and the second agent module, and based on the received suggestions for the initial solution, performs adjustment analysis based on the game result to obtain a decision result.

Description

Technical Field

[0001] The present application relates to the field of big data technology, and in particular to a method for constructing a multimodal large model for the securities industry in an ICT-enabled executable environment. Background Art

[0002] The securities and futures industry is a complex and highly interactive field. The combination of large language models and multi-agent systems has opened up new possibilities for the industry. The universal large language model, capable of processing and generating natural language text, provides the industry with more intelligent information processing and decision support capabilities. It also offers a significant breakthrough in natural language communication between multiple agents, thereby improving efficiency.

[0003] However, applying large language models and multi-agent systems to the securities and futures industry also faces challenges, such as data privacy protection, system stability, and algorithm transparency. Therefore, developing an effective collaborative control algorithm and system is crucial to achieving the successful application of large language models and multi-agent systems in the securities and futures industry. Summary of the Invention

[0004] One purpose of this application is to propose a method for constructing a multimodal large model of the securities industry in an executable environment of trusted innovation.

[0005] According to one embodiment of the present application, a method for constructing a multimodal large model of the securities industry in an ICT executable environment is provided, the method comprising:

[0006] The first agent module analyzes the current task based on the acquired first information to obtain a first analysis result;

[0007] The second agent module analyzes the current task based on the acquired second information to obtain a second analysis result;

[0008] The first agent module and the second agent module, based on the long-term memory and short-term memory obtained from the memory module, conduct a game on the first analysis result and the second analysis result obtained after analyzing the same task. If the first agent module and the second agent module fail to reach a consensus after the game, the third agent module performs weighted processing on the opinions discussed by the first agent module and the second agent module based on a preset weight to obtain a game result;

[0009] The third agent module generates an initial solution based on the task to be played, obtains suggestions for the initial solution from the first agent module and the second agent module, and adjusts and analyzes the game results based on the received suggestions for the initial solution to obtain a decision result, wherein the third agent module has a higher level than the first agent module and the second agent module;

[0010] Each of the agent modules includes a large model trained based on the corresponding information; at least one of the agent modules obtains memory data whose importance ranking reaches a preset threshold from the memory module, and performs a game based on the obtained memory data; the importance of the data stored in the memory module is Determined based on a piecewise scoring function based on the weights of different events and degradation rate α l and the time difference δ E Determine that the longer the event in the database occurs, the greater the corresponding weight is, and the degradation rate α l It is used to measure the difference in importance between different events, the time difference δ E The time interval between the event occurrence time and the knowledge base search query time.

[0011] In one embodiment, the first agent module, the second agent module, and the third agent module are different agent modules, and the third agent module is a senior agent module having a higher level than the first agent module and the second agent module;

[0012] The large models in the first agent module, the second agent module and the third agent module are different functional roles played by the same large model or the same large model with different functional roles;

[0013] The first information, the second information and the third information are different types of information, and the large models in the first agent module, the second agent module and the third agent module are trained based on different types of information. When different large models make decisions on the same task, they make decisions based on the different types of information they obtain.

[0014] In one embodiment, the agent module includes a brain module, a perception module, and an action module, and the agent module includes one or more of a first agent module, a second agent module, or a third agent module;

[0015] The brain module includes giving different roles to the large language model based on Prompt engineering technology, guiding the large language model to adopt procedures to solve problems;

[0016] The perception module is used to receive and understand information from the external environment or other agent modules, and to process and interpret the input data so that the agent module can understand and respond accordingly, wherein the input data includes text, images, sounds or other forms of information;

[0017] The action module is used to perform corresponding actions based on the results obtained after reasoning and decision-making based on the information received by the perception module. The actions include controlling the movement of the robot, generating text replies, and performing one or more specific tasks.

[0018] In one embodiment, the first agent module and the second agent module perform a game based on the first analysis result and the second analysis result obtained after analyzing the same task, and obtain a game result, including:

[0019] The third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and obtains a game result when the game between the first agent module and the second agent module meets a preset game condition.

[0020] In one embodiment, the third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and when the game between the first agent module and the second agent module meets a preset game condition, obtains a game result, including:

[0021] Determining weights corresponding to different opinions and factors based on the importance of different opinions and factors generated by the first agent module and the second agent module during the game process;

[0022] The third agent module performs weighted processing on the opinions generated by the first agent module and the second agent module during the game process based on the weight, and obtains a game result when the game between the first agent module and the second agent module meets a preset number of game rounds or a preset game time.

[0023] In one embodiment, the first information includes public opinion information, the second information includes behavioral information, the first agent module includes a public opinion analysis investor agent, the second agent module includes a market investor agent, and the third agent module includes an advanced decision support system agent.

[0024] According to another embodiment of the present application, a large language model control system for the securities and futures industry based on a trusted execution environment is provided, the system comprising:

[0025] a first agent module, wherein the first agent module analyzes the current task based on the acquired first information to obtain a first analysis result;

[0026] a second agent module, wherein the second agent module analyzes the current task based on the acquired second information to obtain a second analysis result;

[0027] a game module, wherein the first agent module and the second agent module conduct a game based on the long-term memory and short-term memory obtained from the memory module, using the first analysis result and the second analysis result obtained after analyzing the same task; if the first agent module and the second agent module fail to reach a consensus after the game, a third agent module performs weighted processing on the opinions discussed by the first agent module and the second agent module based on a preset weight to obtain a game result;

[0028] a decision-making module, wherein the third agent module generates an initial solution based on the task to be played, obtains suggestions for the initial solution from the first agent module and the second agent module, and adjusts and analyzes the game results based on the received suggestions for the initial solution to obtain a decision result, wherein the third agent module has a higher level than the first agent module and the second agent module;

[0029] Each of the agent modules includes a large model trained based on the corresponding information; at least one of the agent modules obtains memory data whose importance ranking reaches a preset threshold from the memory module, and performs a game based on the obtained memory data; the importance of the data stored in the memory module is Determined based on a piecewise scoring function based on the weights of different events and degradation rate α l and the time difference δ E Determine that the longer the event in the database occurs, the greater the corresponding weight is, and the degradation rate α l It is used to measure the difference in importance between different events, the time difference δ E The time interval between the event occurrence time and the knowledge base search query time.

[0030] In one embodiment, the gaming module further includes:

[0031] The third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and obtains a game result when the game between the first agent module and the second agent module meets a preset game condition.

[0032] The present application also provides a computer-readable storage medium having a computer program or instruction stored thereon, characterized in that when the computer program or instruction is executed by a processor, the steps of the algorithm and system provided in any one of the above embodiments are implemented, and the algorithm and system include: a first agent module analyzes the current task based on the acquired first information to obtain a first analysis result, and the first agent module includes a large model trained based on the first information; a second agent module analyzes the current task based on the acquired second information to obtain a second analysis result, and the second agent module includes a large model trained based on the second information; the first agent module and the second agent module play a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result; a third agent module performs adjustment analysis based on the game result to obtain a decision result, and the third agent module includes a large model trained based on the third information.

[0033] In one embodiment, a computer device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the algorithm and system provided in any one of the above embodiments are implemented, wherein the algorithm and system include: a first agent module analyzes a current task based on the acquired first information to obtain a first analysis result, and the first agent module includes a large model trained based on the first information; a second agent module analyzes the current task based on the acquired second information to obtain a second analysis result, and the second agent module includes a large model trained based on the second information; the first agent module and the second agent module play a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result; a third agent module performs adjustment analysis based on the game result to obtain a decision result, and the third agent module includes a large model trained based on the third information.

[0034] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become obvious from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0036] Figure 1 This is a flowchart of a multi-agent collaborative control algorithm and system for a large language model in the securities and futures industry provided by an embodiment of the present application;

[0037] Figure 2 This is a partial schematic diagram of a multi-agent brain provided by one embodiment of the present application;

[0038] Figure 3 is a schematic diagram of a memory module provided by one embodiment of the present application;

[0039] Figure 4 is a flowchart of decision-making based on a multi-agent module and a memory module provided in another embodiment of the present application;

[0040] Figure 5 It is a module diagram of a large language model and multi-agent collaborative control algorithm and system provided in another embodiment of the present application. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0042] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0043] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0044] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0045] In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0046] In one embodiment, Figure 1 As shown, a method for constructing a multimodal large model of the securities industry in an ICT executable environment is provided, including:

[0047] In step 101 , a first agent module analyzes a current task based on the acquired first information to obtain a first analysis result, wherein the first agent module includes a large model trained based on the first information.

[0048] In step 102 , a second agent module analyzes the current task based on the acquired second information to obtain a second analysis result, wherein the second agent module includes a large model trained based on the second information.

[0049] The first and second information mentioned above refer to different types of information. For example, the first information could be public opinion, while the second information could be market information. Thus, a general large model trained on different types of information possesses different capabilities, enabling it to analyze problems from different perspectives and derive answers based on different angles.

[0050] The first information may be public opinion information. The first agent module performs public opinion analysis based on the acquired public opinion information to obtain a first analysis result. The first agent module includes a general large model trained based on the public opinion information.

[0051] The second information may be market information. The second agent module performs market analysis based on the acquired market information to obtain a second analysis result. The second agent module includes a general large model trained based on the market information.

[0052] The current task may be a task that currently requires a decision, for example, a task of "buy", "sell" or "hold" for a certain stock.

[0053] In step 103 , the first agent module and the second agent module play a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result.

[0054] Because the large models included in the first and second agent modules are trained based on different types of data, the task results obtained by the first and second agent modules for processing the same task are obtained from different perspectives and may differ. In some embodiments, the results obtained by the first and second agent modules for processing the same task can be used to generate a game result. In some embodiments, the game can be a single round or multiple rounds, without limitation.

[0055] In some embodiments, the first agent module and the second agent module conduct a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result, including: the first agent module and the second agent module reach an agreement after negotiation based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain the game result, or the third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and obtains the game result when the game between the first agent module and the second agent module meets the preset game conditions.

[0056] In some embodiments, the first agent module and the second agent module analyze the same task, obtain the first analysis result and the second analysis result, and reach an agreement after negotiation to obtain a game result, including: the first agent module and the second agent module each negotiate on the first analysis result and the second analysis result based on the other party's views and the knowledge respectively acquired; in response to the third agent module determining that the first agent module and the second agent module reach an agreement, it is determined that the first agent module and the second agent module reach an agreement through negotiation to obtain a game result.

[0057] For example, a first agent module and a second agent module negotiate on a specific issue. Negotiation involves the two agents combining the other agent's perspectives with their own knowledge to ultimately reach a consensus or compromise. When both parties are satisfied and accept the agreed decision or solution, consensus is reached, and the negotiation concludes. In some embodiments, a third agent module can determine whether the two parties have reached a consensus and thus whether the game is over.

[0058] During the game, the knowledge acquired by the first and second agent modules can be the latest real-time information. Thus, the first and second agent modules can conduct game play based on the latest acquired data, making the game more reasonable and accurate. In other embodiments, the knowledge acquired by the first and second agent modules can also be historically stored information. This allows for faster acquisition of the corresponding knowledge, enabling game play based on the acquired knowledge to reach a final decision.

[0059] In some embodiments, the third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on preset weights, and obtains a game result when the game between the first agent module and the second agent module meets preset game conditions, including: determining the weights corresponding to different opinions and factors based on the importance of different opinions and factors generated by the first agent module and the second agent module during the game process; the third agent module weights the opinions generated by the first agent module and the second agent module during the game process based on the weights, and obtains a game result when the game between the first agent module and the second agent module meets a preset number of game rounds or a preset game time.

[0060] For example, after multiple rounds of negotiation between the first and second agent modules, a third agent module weights the opinions discussed by both parties using different weights. These weights can be numerical values that quantify the importance of different opinions and factors. Weights can be determined subjectively or objectively based on data analysis. These weights inform the third agent module how to comprehensively consider the opinions of all parties, thereby better valuing the importance of each. The end of this portion of the game can be determined by meeting a specific number of rounds or a time limit.

[0061] Step 104: The third agent module performs adjustment analysis based on the game result to obtain a decision result, and the third agent module includes a large model trained based on the third information.

[0062] The general large model included in the third agent module is trained based on the third information, and the first information, the second information and the third information are different types of information. For example, see Figure 2 Based on Prompt engineering technology, we can give large language models different roles and guide them to use procedures to solve problems.

[0063] In some embodiments, the first agent module, the second agent module, and the third agent module are different agent modules, and the third agent module is a high-level agent module that is higher in level than the first agent module and the second agent module.

[0064] In some embodiments, the big models in the first agent module, the second agent module and the third agent module are different functional roles played by the same big model or the same big model with different functional roles.

[0065] In some embodiments, the first information, the second information, and the third information are different types of information, and the large models in the first agent module, the second agent module, and the third agent module are trained based on different types of information. When different large models make decisions on the same task, they make decisions based on the different types of information they have obtained.

[0066] After the first and second agent modules have engaged in a game on the same task, the third agent module can further analyze the game results to arrive at the final decision. In this way, three different types of agent modules jointly make decisions on the same task, enabling decisions to be made from different perspectives, resulting in more intelligent and reasonable decisions.

[0067] In the above embodiment, the first agent module, the second agent module and the third agent module can be understood as different agent modules, which are respectively used to act as agents for different matters. The three agent modules work together to complete the decision-making process of the business.

[0068] In some embodiments, the first information includes public opinion information, the second information includes behavioral information, the first agent module includes a public opinion analysis investor agent, the second agent module includes a market investor agent, and the third agent module includes an advanced decision support system agent.

[0069] In some embodiments, the first agent module can be used to process public opinion related information. For example, the first agent module can be a public opinion analysis investor agent, which obtains public opinion information and analyzes the impact of public opinion on stock price trends.

[0070] In some embodiments, the second agent module may be a market investor agent, which performs market analysis and trading decisions based on market data and technical indicators.

[0071] Moreover, the first agent module and the second agent module can play a game based on their respective analysis results, specifically, the public opinion analysis investor Agent and the market investor Agent can play a game.

[0072] In some embodiments, the third agent module can be a decision support system agent, which extracts key information from the game results of the two agents corresponding to the first agent module and the second agent module, comprehensively considers the results of public opinion analysis and market analysis, as well as the arguments and rebuttals of the two parties in the game, and thus obtains the final decision result.

[0073] In the above embodiments, the multi-agent system can simulate the behaviors of multiple participants in the market, realize intelligent decision-making and resource allocation, and thus play an important role in securities and futures trading.

[0074] The large language model has powerful natural language processing and generation capabilities. The agent based on the large language model in the above embodiment can use the large language model as a brain module, thus the agent module has powerful natural language processing and generation capabilities. The multiple agent collaboration in the above embodiment can be the collaborative information exchange between multiple intelligent entities to jointly complete a task.

[0075] In some embodiments, the above-mentioned agent module Agent can be an intelligent entity that can perceive the environment, make decisions and execute. The agent module includes a brain module, a perception module and an action module. The brain module includes giving different roles to the large language model based on Prompt engineering technology, and guiding the large language model to adopt procedures to solve problems. The perception module is used to receive and understand information from the external environment or other agent modules, and process and interpret the input data so that the agent module can understand and respond accordingly, wherein the input data includes text, images, sounds or other forms of information. The action module is used to perform corresponding actions based on the results obtained after reasoning and decision-making based on the information received by the perception module, and the actions include one or more of controlling the movement of the robot, generating text replies, and performing specific tasks.

[0076] When a large language model uses a program to solve a problem, the program can include three parts: role identification, game theory, and multi-round iterative collaboration. The following uses the decision-making scenario in the securities and futures industry as an example to illustrate the details:

[0077] Role identification. Two agents are designed for two different types of investors: public opinion analysis and market analysis. The public opinion analysis agent can obtain public opinion information from large amounts of text data such as news and social media, and analyze its impact on stock price trends. The market investor agent can make market analysis and trading decisions based on market data and technical indicators. A high-level agent is also introduced as a decision support system to extract key information from the game results of the two agents, comprehensively considering the results of public opinion analysis and market analysis, as well as the arguments and rebuttals of both parties, to reach the final result. Multiple large language models can play different roles for different tasks and are widely used.

[0078] After identifying the different participants, the senior agent is considered the leader, initiating collaboration and generating an initial solution. Before reaching a final solution, the public opinion analysis and investment agent and the market investment agent engage in a game of speculation regarding stock trends. The public opinion analysis and investment agent may focus on company sentiment, industry dynamics, and management performance, while the market investment agent may focus on market trends, technical indicators, and trading volume.

[0079] Multi-round iterative collaboration. After multiple rounds of negotiation, the game ends either when both parties reach consensus on a specific issue, or when a specific number of rounds and time limit are reached. Based on the game results and comprehensive analysis, the public opinion analysis and market investment agents can jointly reach an investment decision, considering each other's perspectives and their own knowledge. This decision could be "buy," "sell," or "hold" for a particular stock. This decision may be reached through negotiation or by comprehensively considering both parties' opinions based on a certain weighting. The senior agent role generates an initial solution and then consults with each other participant for feedback. Participants are encouraged to evaluate the initial solution and propose revisions. The senior agent role then summarizes the solution to provide reference for the decision.

[0080] The securities and futures industry is a complex and highly interactive field. The combination of large language models and multi-agent systems opens up new possibilities for the industry. Universal large language models can process and generate natural language text, providing the industry with more intelligent information processing and decision support capabilities. They also offer a significant breakthrough in natural language communication between multiple agents, thereby improving efficiency. Furthermore, compared to single-agent systems, which lack the ability to collaborate with other agents and acquire knowledge from social interactions, multi-agent systems allow different intelligent agents (such as trading algorithms and risk management systems) to collaborate, compete, and learn from each other, achieving more efficient market transactions and risk control.

[0081] In the above-mentioned embodiments, large language models are used in the securities and futures industry for market intelligence analysis, public opinion monitoring, and financial news reporting, helping financial institutions better understand market dynamics and inform their decision-making. Furthermore, securities and futures trading involves multiple parties and complex trading decision-making processes. Therefore, multi-agent collaborative control algorithms are crucial tools for optimizing trading strategies, improving efficiency, and mitigating risks. Multi-agent systems can simulate the behavior of multiple market participants, enabling intelligent decision-making and resource allocation, thus playing a vital role in securities and futures trading.

[0082] In some embodiments, the third agent module performs adjustment analysis based on the game result to obtain a decision result, including: the third agent module generates an initial solution based on the task of the game, and obtains suggestions for the initial solution from the first agent module and the second agent module; the third agent module makes a decision based on the suggestions for the initial solution received from the first agent module and the second agent module to obtain a decision result.

[0083] Specifically, the third agent module can be a high-level agent role. This agent generates an initial solution and then consults and provides feedback to each other participant (such as the first and second agents). Participants are also encouraged to evaluate the initial solution and propose modifications. Finally, the high-level agent role summarizes the solution and provides reference for decision-making. Because the third agent module includes a large model trained based on third-party information, the third information has more advanced decision-making capabilities and is able to integrate suggestions from multiple parties to make comprehensive decisions. Therefore, the final decision reached by the third agent module is more reasonable and effective.

[0084] like Figure 2 The figure shows a schematic diagram of the multi-agent brain provided in one embodiment of the present application. The multi-agent brain utilizes a large language model to simulate different roles. Through the prompt project, the large model is given different personalities to play different roles. In the prompt, you can introduce the background of the task, the goal of the task, the role played in the task, what skills you have, and provide some examples. Different roles play games for the same task, synergistically combining the advantages and knowledge of multiple brains to improve the ability to solve complex tasks. At the same time, a high-level agent is set up to conduct a specific analysis of the task, and then assign it to different agents to play games. Finally, the game results are collected and summarized to provide assistance for decision-making.

[0085] In some embodiments, the algorithm and system also include a memory module, which is used to store the public opinion information and the market information in a structured manner; the memory module includes long-term memory and short-term memory; the long-term memory is used to store historical data, and the short-term memory is used to store real-time data.

[0086] In some embodiments, the memory module can be used to store data such as real-time news and market analysis information, allowing the large language model to simulate human memory. For example, long-term memory can be used to store company reports, financial data, transaction history, etc., while short-term memory may be used to store current news, social media sentiment, real-time market data, etc.

[0087] In some embodiments, this information may be stored in a memory module in a structured form. For example, the structured data may be in a key-value pair format such as (Q, A).

[0088] In some embodiments, the algorithm and system further include: the first agent module and the second agent module, based on memory data obtained from the memory module, performing a game on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result. Because the memory module stores short-term memory and long-term memory, the first agent module and the second agent module can perform the game based on the most recent memory data obtained from the memory module.

[0089] In some embodiments, the algorithm and system further include: indexing data from the memory module based on a vector indexing method; one or more of the first agent module, the second agent module, and the third agent module indexing data from the memory module based on a vector indexing method; and one or more of the first agent module, the second agent module, and the third agent module performing a game based on the data indexed from the memory module.

[0090] In order to quickly retrieve and analyze stored information, indexing technology can be used to index data, and content can be obtained through vector retrieval, so that the required data can be quickly indexed from the memory module.

[0091] In some embodiments, the vector index-based method for indexing data from the memory module includes: the proxy module determines the similarity between the data to be indexed and the data stored in the memory module, and indexes the data from the memory module based on the similarity; wherein, the proxy module includes one or more of the first proxy module, the second proxy module and the third proxy module.

[0092] In order to quickly retrieve and analyze stored information, indexing technology is used to index data to speed up data retrieval and query efficiency. Content can be retrieved through vector retrieval, as shown in formula (1).

[0093]

[0094] As shown in formula (1), it represents the memory event m E The embedding vector of the text content and the hint query m p The cosine similarity between .

[0095] In some embodiments, the vector index-based method for indexing data from the memory module includes: the proxy module determines the index time corresponding to the indexing data, and the event occurrence time corresponding to the data stored in the memory module, and obtains events corresponding to different times from the memory module based on the index time and the event occurrence event; wherein, the proxy module includes one or more of the first proxy module, the second proxy module and the third proxy module.

[0096] As shown in formula (2) and formula (3), it is used to obtain events at different times.

[0097] δ E =t P -t E (2)

[0098] Among them, δ E Indicates the difference between the time when the event occurred in the memory unit and the query time. P With t E They represent the occurrence time and query time of events in the memory unit respectively.

[0099]

[0100] Among them, Q l Indicates the number of days, δ E Represents the difference between the time when the event occurred in the memory unit and the query time. This formula can be used to obtain the time events corresponding to the required different times.

[0101] In the above embodiment, the knowledge required for a specific task is first stored in the database. In addition, the stored knowledge information provides indexing and retrieval functions to serve the general large model. There are two retrieval methods, see formulas (1) to (3) for details. The embedding of the problem is achieved by using the vector model in formula (1) and the similarity comparison with the embedding of the knowledge base content. It is also possible to query some timely or non-timely knowledge through formulas (2) and (3).

[0102] like Figure 3 As shown in FIG, it is a schematic diagram of a memory module provided in one embodiment of the present application. Figure 3As shown in the left figure, the main functions of the memory module include data storage and management, content indexing and retrieval, content processing and analysis, and real-time updates and synchronization. The memory module can be divided into long-term memory and short-term memory. Long-term memory primarily stores data such as company financial statements, macroeconomic data, national policies, and upstream and downstream enterprise relationships. This data can provide support for investors who prioritize market trends. Short-term memory primarily stores real-time news, social network sentiment, stock and futures market data, and other data. This data can also support investors who prioritize public opinion. This allows agents with different roles to obtain timely information to support their views.

[0103] In some embodiments, the long-term memory and short-term memory information stored in the memory unit may be further processed and analyzed to provide useful information for decision making.

[0104] In some embodiments, the algorithm and system also include: sorting the memory data based on the importance of the memory data stored in the memory module; the agent module obtains the memory data whose importance ranking reaches a preset threshold from the memory module, and conducts a game based on the obtained memory data, and the agent module includes one or more of the first agent module, the second agent module and the third agent module.

[0105] The processing and analysis mentioned above involves sorting the relevant data. For large amounts of data, the general model is ranked by importance based on this data, and the top-K memory events are then retrieved and fed into the general model. This information contains insights and sentiment that are key to the specific task. The importance score is shown in Formula (4).

[0106]

[0107] in, It refers to a unified piecewise scoring function (Formula 5), which gives different events in the database a weight. The larger the value, the longer the event occurred in the database, and the smaller the value, the shorter the event occurred. l Refers to the degradation rate, α l There are two values: short and long, where α short <α long , α short You can take 0.9, α long It can be taken as 0.988. This is a measure of the importance difference between different times. E It refers to the time interval between the event occurrence time and the knowledge base search query time (as shown in Formula 2). Indicates the decreasing importance of an event over time.

[0108] Short-term memory (p1, p2, p3) can be (0.8, 0.15, 0.05), and long-term memory (p1, p2, p3) can be (0.05, 0.15, 0.8). This means that if one agent pays more attention to the data in the short-term memory, (p1, p2, p3) will be (0.8, 0.15, 0.05). The probability of taking 40 will be greater, and more attention will be paid to events that occur in a shorter time. If one agent pays more attention to the data in long-term memory, then (p1, p2, p3) can be (0.05, 0.15, 0.8). The probability of taking 80 will be higher, and more attention will be paid to events that take a longer time to occur.

[0109] The above real-time updating and synchronization requires continuous updating of the memory data in the memory module in order to ensure the timeliness of the knowledge in the database.

[0110] In some embodiments, the algorithm and system further include real-time updating and synchronization of the first, second, and third information in the memory module. For example, public opinion information and market information may be updated in real time and synchronized with the actual market. In some embodiments, the memory unit has real-time updating and synchronization capabilities to ensure that the information stored therein is up to date.

[0111] In some embodiments, the algorithm and system also include an execution module. This execution module is used to formulate specific trading decisions and convert them into actual trading activities based on the relevant information stored in the memory unit and the results analyzed by the advanced agent. It mainly includes the following parts:

[0112] Decision Making. Based on the comprehensive analysis results of the advanced agent and the information in the memory unit, the decision module can use algorithms and models to make specific trading decisions. This may involve making decisions to buy, sell, or hold specific securities.

[0113] Risk management. The decision-making module needs to consider risk management factors to ensure that the trading decisions made are consistent with the investor's risk appetite and capital management strategy. This may involve setting stop-loss points, diversifying the portfolio, and considering market liquidity.

[0114] Real-time monitoring. The decision-making module needs to monitor the results of executed transactions in real time to ensure that transactions are executed as expected. It can also optimize and adjust the decision-making process in a timely manner through learning and feedback.

[0115] like Figure 4The figure below is a flow chart illustrating a multi-agent and memory-based decision-making process in one embodiment of this application. The execution module, which uses the results of the game to make decisions about tasks, first stores data in the long-term and short-term memory modules. These data are then retrieved based on similarity and timeliness. The acquired information is then used to determine the top-K events based on their importance and fed into a general macro model for game play. Advanced agents then combine database knowledge with the results of multi-agent debate to make appropriate decisions. In the investment field, this can determine whether to "buy," "sell," or "hold" within a specific timeframe. Corresponding risk management and real-time monitoring are also performed.

[0116] like Figure 5 Figure 2 shows a modular diagram of a multi-agent collaborative control algorithm and system based on a large language model. The system comprises a multi-agent brain module, a memory module, and an execution module. The multi-agent brain module primarily leverages the natural language processing and text generation capabilities of the general large model as the system's brain. It can understand, analyze, and generate complex information in the securities and futures industry, thereby providing powerful intelligent decision-making support for the multi-agent system. The memory module primarily stores relevant information required for tasks. Because the general large model is trained on large amounts of data, which has a lag, the large model cannot learn the latest information. To address this shortcoming, the general large model can query the memory module for information. The memory module consists of long-term and short-term memory components, employing a hierarchical structure to accommodate the varying time sensitivities inherent in different types of securities and futures data. The execution module effectively integrates the analysis results of the general large model with the relevant content of the memory module to support informed decision-making.

[0117] In some embodiments, the memory module can store information related to the discussion task. This information can come from a knowledge base or a search engine, enhancing the system's timeliness. This component comprises two modules: long-term memory and short-term memory. Each agent can search the memory module for relevant information and respond accordingly based on this information and its role. The execution module primarily makes task decisions based on the results of multi-agent collaboration.

[0118] In the above-mentioned embodiments, a multi-agent collaborative control algorithm is used to solve tasks related to the securities and futures sector. Combining the natural language processing and understanding capabilities of a large language model with the decision-making and negotiation mechanisms of multi-agent collaborative control, the system is able to comprehensively consider more information and interests, leveraging collective intelligence to improve decision accuracy and global optimization capabilities. In tasks where public opinion and market conditions influence investment decisions, the system performs well not only on historical data but also on future data. Furthermore, the multi-agent collaborative control algorithm can define different roles for discussion based on different tasks, and has a wide range of applications, such as international crude oil price analysis, futures trading, and other fields, demonstrating strong versatility and applicability.

[0119] The algorithms and systems provided in the above embodiments can be applied to securities and futures scenarios. For example, the algorithms and systems provided in this application can be a multi-agent collaborative control algorithm and system for a large language model in the securities and futures industry. The algorithm and system mainly include three core structures: a multi-agent brain module, a memory module, and an execution module. Among them, the multi-agent brain module gives the large language model different personalities to play different roles for discussion or game. In the discussion of the impact of public opinion and market conditions on investment decisions in the securities industry, the multi-agent brain module can include two agents, representing the characteristics, decision-making preferences, and behavior patterns of two different types of investors, public opinion analysis and market analysis respectively. At the same time, a high-level agent can be set up to summarize the results of the discussion.

[0120] Current large language models suffer from the limitations of hallucinations and information latency, while the securities and futures industry places high demands on stability, security, and real-time information. Furthermore, single-agent systems are relatively isolated and lack the ability to communicate and learn from other agents to improve performance. To address these shortcomings, we propose a multi-agent collaborative control algorithm and system for large language models in the securities and futures industry. Large language models are given different personalities to play different roles, or multiple large language models are used to discuss or negotiate the same task. Ultimately, a higher-level agent summarizes the discussion or negotiation. The system also incorporates a memory module to store task-specific information, such as current mainstream industry data, company reports, and news. During discussions and debates, multiple agents can access the information they need from the memory module to support their own views. The execution module primarily makes decisions based on the results of the multi-agent discussions or debates.

[0121] The present application also provides a computer-readable storage medium having a computer program or instruction stored thereon, characterized in that when the computer program or instruction is executed by a processor, the steps of the algorithm and system provided in any one of the above embodiments are implemented, and the algorithm and system include: a first agent module analyzes the current task based on the acquired first information to obtain a first analysis result, and the first agent module includes a large model trained based on the first information; a second agent module analyzes the current task based on the acquired second information to obtain a second analysis result, and the second agent module includes a large model trained based on the second information; the first agent module and the second agent module play a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result; a third agent module performs adjustment analysis based on the game result to obtain a decision result, and the third agent module includes a large model trained based on the third information.

[0122] In one embodiment, a computer device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the algorithm and system provided in any one of the above embodiments are implemented, wherein the algorithm and system include: a first agent module analyzes a current task based on the acquired first information to obtain a first analysis result, and the first agent module includes a large model trained based on the first information; a second agent module analyzes the current task based on the acquired second information to obtain a second analysis result, and the second agent module includes a large model trained based on the second information; the first agent module and the second agent module play a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result; a third agent module performs adjustment analysis based on the game result to obtain a decision result, and the third agent module includes a large model trained based on the third information.

[0123] In one embodiment, a computer program product is also provided, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the steps of the algorithm and system provided in any one of the above embodiments are implemented. The algorithm and system include: a first agent module analyzes the current task based on the acquired first information to obtain a first analysis result, and the first agent module includes a large model trained based on the first information; a second agent module analyzes the current task based on the acquired second information to obtain a second analysis result, and the second agent module includes a large model trained based on the second information; the first agent module and the second agent module play a game based on the first analysis result and the second analysis result obtained after analyzing the same task to obtain a game result; the third agent module adjusts and analyzes the game result to obtain a decision result, and the third agent module includes a large model trained based on the third information.

[0124] It should be understood that although Figure 1-5 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1-5 At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The order of execution of these steps or stages is not necessarily one by one, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0125] It should also be understood that the first, second, third, fourth and various numerical numbers involved in this document are only distinctions made for the convenience of description and are not intended to limit the scope of this application.

[0126] It should be understood that the term "and / or" in this document simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0127] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meanings as those commonly understood by those skilled in the art to which this application belongs. In the event of any inconsistency, the meanings described in this specification or derived from the contents described in this specification shall prevail.

[0128] Those skilled in the art will understand that all or part of the processes in the algorithms and systems of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned algorithms and systems.

[0129] It should be understood that the processor mentioned in the embodiments of the present application may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0130] It should also be understood that the memory mentioned in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DR RAM).

[0131] It should be noted that when the processor is a general-purpose processor, DSP, ASIC, FPGA or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, the memory (storage module) is integrated into the processor.

[0132] It should be noted that the memory described herein is intended to include, but is not limited to, these and any other suitable types of memory. The technical features of the above embodiments can be combined in any manner. To simplify the description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there are no contradictions in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for constructing a multimodal large model of the securities industry in an ICT executable environment, characterized by: The method comprises: The first agent module analyzes the current task based on the acquired first information to obtain a first analysis result; The second agent module analyzes the current task based on the acquired second information to obtain a second analysis result; The first agent module and the second agent module, based on the long-term memory and short-term memory obtained from the memory module, conduct a game on the first analysis result and the second analysis result obtained after analyzing the same task. If the first agent module and the second agent module fail to reach a consensus after the game, the third agent module performs weighted processing on the opinions discussed by the first agent module and the second agent module based on a preset weight to obtain a game result; The third agent module generates an initial solution based on the task to be played, obtains suggestions for the initial solution from the first agent module and the second agent module, and adjusts and analyzes the game results based on the received suggestions for the initial solution to obtain a decision result, wherein the third agent module has a higher level than the first agent module and the second agent module; Among them, the first agent module, the second agent module and the third agent module respectively include large models trained based on the first information, the second information and the third information; at least one of the agent modules obtains memory data whose importance ranking reaches a preset threshold from the memory module, and performs a game based on the obtained memory data; the importance of the data stored in the memory module Determined based on a piecewise scoring function based on the weights of different events and degradation rate and time difference Determine that the longer the event in the database occurs, the greater the corresponding weight. Used to measure the difference in importance between different events, the time difference The time interval between the event occurrence time and the knowledge base search query time.

2. The method according to claim 1, characterized in that The importance of the data stored in the memory module in in, It is a unified piecewise scoring function that gives different events in the database a weight. The larger the value, the longer the event occurred in the database, and the smaller the value, the shorter the event occurred. is the degradation rate, The time when the event occurred and inquiry time time interval, Indicates the decreasing importance of an event over time.

3. The method according to claim 1, characterized in that The first agent module, the second agent module and the third agent module are different agent modules, and the third agent module is a senior agent module having a higher level than the first agent module and the second agent module; The large models in the first agent module, the second agent module and the third agent module are different functional roles played by the same large model or the same large model with different functional roles; The first information, the second information and the third information are different types of information, and the large models in the first agent module, the second agent module and the third agent module are trained based on different types of information. When different large models make decisions on the same task, they make decisions based on the different types of information they obtain.

4. The method according to claim 1, wherein The agent module includes a brain module, a perception module and an action module, and the agent module includes one or more of a first agent module, a second agent module or a third agent module; The brain module includes giving different roles to the large language model based on Prompt engineering technology, guiding the large language model to adopt procedures to solve problems; The perception module is used to receive and understand information from the external environment or other agent modules, and to process and interpret the input data so that the agent module can understand and respond accordingly, wherein the input data includes text, images, sounds or other forms of information; The action module is used to perform corresponding actions based on the results obtained after reasoning and decision-making based on the information received by the perception module. The actions include controlling the movement of the robot, generating text replies, and performing one or more specific tasks.

5. The method according to claim 1, wherein The first agent module and the second agent module perform a game based on the first analysis result and the second analysis result obtained after analyzing the same task, and obtain a game result, including: The third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and obtains a game result when the game between the first agent module and the second agent module meets a preset game condition.

6. The method according to claim 5, characterized in that The third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and when the game between the first agent module and the second agent module meets a preset game condition, obtains a game result, including: Determining weights corresponding to different opinions and factors based on the importance of different opinions and factors generated by the first agent module and the second agent module during the game process; The third agent module performs weighted processing on the opinions generated by the first agent module and the second agent module during the game process based on the weight, and obtains a game result when the game between the first agent module and the second agent module meets a preset number of game rounds or a preset game time.

7. The method according to any one of claims 1 to 6, characterized in that The first information includes public opinion information, the second information includes behavioral information, the first agent module includes a public opinion analysis investor agent, the second agent module includes a market investor agent, and the third agent module includes an advanced decision support system agent; and / or, the decision includes one or more of decision making, risk management and real-time monitoring.

8. The method according to claim 1, characterized in that The method further includes a memory module, the memory module being configured to store one or more of the first information, the second information, and the third information in a structured manner, the agent module obtaining memory data corresponding to the first information, the second information, and the third information from the memory module, and performing a game based on the obtained memory data; The memory module includes long-term memory and short-term memory; The long-term memory is used to store historical data, and the short-term memory is used to store real-time data.

9. The large language model control system for the securities and futures industry is characterized by: The system comprises: a first agent module, wherein the first agent module analyzes the current task based on the acquired first information to obtain a first analysis result; a second agent module, wherein the second agent module analyzes the current task based on the acquired second information to obtain a second analysis result; a game module, wherein the first agent module and the second agent module conduct a game based on the long-term memory and short-term memory obtained from the memory module, using the first analysis result and the second analysis result obtained after analyzing the same task; if the first agent module and the second agent module fail to reach a consensus after the game, a third agent module performs weighted processing on the opinions discussed by the first agent module and the second agent module based on a preset weight to obtain a game result; a decision-making module, wherein the third agent module generates an initial solution based on the task to be played, obtains suggestions for the initial solution from the first agent module and the second agent module, and adjusts and analyzes the game results based on the received suggestions for the initial solution to obtain a decision result, wherein the third agent module has a higher level than the first agent module and the second agent module; Each of the agent modules includes a large model trained based on the corresponding information; at least one of the agent modules obtains memory data whose importance ranking reaches a preset threshold from the memory module, and performs a game based on the obtained memory data; the importance of the data stored in the memory module is Determined based on a piecewise scoring function based on the weights of different events and degradation rate and time difference Determine that the longer the event in the database occurs, the greater the corresponding weight. Used to measure the difference in importance between different events, the time difference The time interval between the event occurrence time and the knowledge base search query time.

10. The system according to claim 9, characterized in that The game module also includes: The third agent module weights the opinions corresponding to the first analysis result and the second analysis result based on a preset weight, and obtains a game result when the game between the first agent module and the second agent module meets a preset game condition.

Images