A large model dynamic valuation control method and system
By constructing event sequences and game heatmaps, and combining them with a multi-agent architecture to simulate the event game process, the problem that existing valuation methods cannot adapt to dynamic market environments is solved, and real-time accuracy and environmental adaptability of product valuation are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU ZICHUAN INFORMATION TECH CO LTD
- Filing Date
- 2026-03-20
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241362A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of valuation control technology, and more specifically to a dynamic valuation control method and system for large models. Background Technology
[0002] Currently, with the rapid development of big data and artificial intelligence technologies, valuation methods based on large models have been widely used in finance, investment, and other fields. Existing valuation methods typically rely on static financial indicators, industry comparisons, or discounted cash flow models. These methods cannot adapt to highly dynamic and complex market environments. Based solely on historical data and fixed rules, they struggle to capture the dynamic impact of environmental factors such as market events and competitive behavior on product value, resulting in lagging and significant biases in product valuation results.
[0003] The existing technology has the following problems: based on historical data, the product value is updated at a low frequency and cannot respond to the impact of market events in real time; it adopts a single analysis method and ignores the game between different events, which makes it unable to adapt to real-time changes in the environment during the analysis process, resulting in a large deviation between the valuation results and the actual value; it only analyzes the static value of the product and does not consider the impact of environmental changes on the product value, which makes the valuation results unable to reflect the real value of the product in real time; in order to solve at least one of the above problems, this application proposes a large-scale model dynamic valuation control method and system. Summary of the Invention
[0004] To address the shortcomings of existing technologies, the purpose of this application is to provide a dynamic estimation control method and system for large models, which can effectively solve the problems in the background technology. The specific technical solution of this application is as follows: A dynamic valuation control method for large models includes: Based on the pre-acquired environmental data, the environmental changes are analyzed using a pre-set event parsing model to extract the corresponding environmental events and construct an event sequence. For each event in the event sequence, features are extracted using a preset feature extraction model to construct a corresponding event feature vector. Based on the event feature vector, the event game process is simulated through a pre-constructed multi-agent architecture, the game value of each agent is calculated, and the corresponding game heatmap is generated according to the game value. In response to the preset valuation task, the target feature vector of the target product is analyzed and mapped onto the game heat map. A dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor to obtain the dynamic valuation result, so as to dynamically control the valuation process.
[0005] Specifically, the step of analyzing environmental changes based on pre-acquired environmental data using a preset event parsing model, parsing corresponding environmental events, and constructing an event sequence includes: The pre-acquired environmental data is preprocessed to obtain the first environmental data; Based on the first environmental data, the environmental changes are analyzed through a preset event parsing model to obtain event information and cluster it to obtain an event set; Arrange the events in the event set in chronological order to obtain the event sequence.
[0006] Specifically, based on the first environmental data, the environmental changes are analyzed using a preset event parsing model to obtain event information and cluster it to obtain an event set, including: Using a pre-defined event parsing model, technical and semantic analyses are performed on the first environmental data to obtain event information. Calculate the association keys between event information, cluster the event information based on the association keys, and treat each cluster as an event to obtain an event set.
[0007] Specifically, for each event in the event sequence, features are extracted using a preset feature extraction model to construct a corresponding event feature vector, including: Each event in the event sequence is parsed into different event components, resulting in a set of event components; Based on the event component set, the corresponding event component features are extracted using a preset feature extraction model to obtain the event component feature set; According to the corresponding event component dimensions, the event component feature sets are fused to construct the corresponding event feature vector.
[0008] Specifically, based on the event feature vector, the event game process is simulated through a pre-built multi-agent architecture, the game value of each agent is calculated, and a corresponding game heatmap is generated according to the game value, including: Based on the analysis of event feature vectors, an event type is constructed, and a multi-agent architecture is built, with each agent responsible for one event type. In the multi-agent architecture, each agent simulates and analyzes the event state, simulates the event game process, calculates the game value corresponding to each agent, and obtains the game value sequence. The game value sequence is mapped to a heatmap according to time order and agent type to construct a game heatmap.
[0009] Specifically, in the multi-agent architecture, each agent simulates and analyzes the event state, simulates the event game process, calculates the game value corresponding to each agent, and obtains a sequence of game values, including: Based on each agent in the multi-agent architecture, the event state is simulated and analyzed to obtain the corresponding set of decision actions; Based on the set of decision actions, simulate environmental changes to obtain an environmental state vector; By combining the environmental state vector and the set of decision actions, the event game process is simulated, the contribution of the corresponding agent to the environmental state is analyzed, and the game value corresponding to each agent is calculated. The game values are arranged in chronological order to obtain a game value sequence.
[0010] Specifically, in response to a preset valuation task, the target feature vector of the target product is parsed, mapped onto the game heatmap, and a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor to obtain the dynamic valuation result, including: In response to a pre-defined valuation task, the valuation information of the target product is analyzed using a pre-defined task analysis model to obtain the target feature vector. According to the time correspondence, the target feature vector is mapped onto the game heatmap, and the game vector is calculated; By combining the game vectors, a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor, and the dynamic valuation result is obtained.
[0011] Specifically, the step of mapping the target feature vector to the game heatmap according to the time correspondence and calculating the game vector includes: Analyze the valuation time points from the target feature vector; Based on the valuation time points, heatmap slices with preset time windows are extracted from the game heatmap. The agent game values in the heatmap slices are analyzed, the corresponding game features are extracted, and game vectors are constructed.
[0012] Specifically, based on the game vector, a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor, resulting in a dynamic valuation result, including: Based on the target feature vector, the basic value of the target product is analyzed through a preset basic valuation model, and the corresponding basic valuation is calculated. Based on the game vector, the impact of the agent's game state on value is analyzed through a preset dynamic adjustment model, and the corresponding dynamic adjustment factor is calculated. The basic valuation and the valuation adjusted by the dynamic adjustment factor are weighted and fused to obtain the dynamic valuation result. The weights in the weighting and fusion process are calculated based on the sensitivity information in the game vector.
[0013] A large-scale model dynamic valuation control system, used to implement the aforementioned large-scale model dynamic valuation control method, includes: The environmental analysis module analyzes environmental changes based on pre-acquired environmental data using a preset event parsing model, identifies corresponding environmental events, and constructs an event sequence. The event analysis module extracts features for each event in the event sequence using a preset feature extraction model and constructs a corresponding event feature vector. The game heatmap construction module, based on the event feature vector, simulates the event game process through a pre-built multi-agent architecture, calculates the game value of each agent, and generates the corresponding game heatmap according to the game value; The dynamic valuation module, in response to a preset valuation task, analyzes the target feature vector of the target product, maps it to the game heatmap, configures a dynamic valuation mechanism to calculate the corresponding basic valuation and dynamic adjustment factor, and obtains the dynamic valuation result to dynamically control the valuation process.
[0014] The beneficial effects of this application are as follows: Event sequences are obtained by analyzing multi-source environmental data using an event analysis model; a multi-agent architecture is constructed, with each agent representing a type of event; the contribution of various events to the environmental state is analyzed through simulated decision-making and game theory processes, generating a game heatmap; the basic valuation and dynamic adjustment factor are calculated separately by combining the game heatmap and a dynamic valuation mechanism, and then fused to obtain the dynamic valuation result; by combining real-time environmental events with multi-agent game theory, the accuracy of environmental state analysis results is improved; the basic valuation and dynamic adjustment factor are calculated separately based on the game heatmap, and the impact of environmental changes on product value is incorporated into the valuation process, improving the accuracy and environmental adaptability of the valuation results. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating a large-scale model dynamic valuation control method as described in this application. Figure 2 This is a schematic diagram of the event set in the embodiments of this application; Figure 3 This is a schematic diagram of the game heatmap in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of a large-scale dynamic valuation control system in an embodiment of this application. Detailed Implementation
[0016] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.
[0017] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.
[0018] Hereinafter, the terms "first," "second," and other generic terms are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.
[0019] refer to Figure 1 The image shows a specific implementation of a large-scale model dynamic estimation control method according to this application, including: S101. Based on the pre-acquired environmental data, analyze the environmental changes through a preset event parsing model, parse out the corresponding environmental events, and construct an event sequence; S102. For each event in the event sequence, feature extraction is performed using a preset feature extraction model to construct a corresponding event feature vector. S103. Based on the event feature vector, simulate the event game process through a pre-constructed multi-agent architecture, calculate the game value of each agent, and generate a corresponding game heatmap according to the game value. S104. In response to the preset valuation task, the target feature vector of the target product is analyzed, mapped to the game heat map, and a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor to obtain the dynamic valuation result, so as to dynamically control the valuation process.
[0020] This embodiment integrates multi-source environmental information, simulates multi-agent game processes, and dynamically adjusts valuation results in real time, thereby improving the accuracy, timeliness, and interpretability of the valuation results. In this embodiment, environmental data includes, but is not limited to, news data and industry report data. A pre-set event analysis model is used to perform in-depth analysis on the pre-acquired environmental data. This event analysis model includes, but is not limited to, a natural language processing (NLP) model. A pre-trained NLP model is obtained by training the NLP model with a large amount of historical environmental data. The environmental data is then input into the pre-trained NLP model, which calculates the association keys between information extracted from different data fragments. Information describing the same event is aggregated together to form independent event units, constructing an event sequence. By analyzing environmental data to construct event sequences, a large amount of environmental data is transformed into event chains with semantic and temporal relationships, reducing the complexity of data processing and analysis. Clustering information avoids interference from duplicate and redundant information, ensuring the simplicity and representativeness of the event sequence, and providing accurate environmental event data for product valuation.
[0021] Specifically, for each event in the event sequence, a pre-defined feature extraction model is used to parse each event into different event components. Features are extracted from these components to construct corresponding event feature vectors. The feature extraction model includes, but is not limited to, a convolutional neural network (CNN) model. This CNN model is trained using a large amount of event data to obtain a pre-trained model. Each event in the event sequence is then input into this pre-trained model. The model analyzes and extracts corresponding semantic features, quantitative indicators, etc., and outputs an event feature vector. By analyzing events and extracting features, representative feature data is obtained, providing a data basis for event characteristic analysis.
[0022] After calculating the event feature vectors, the impact of each event on the market is assigned to a corresponding agent. The game process between these agents is simulated and analyzed to quantify the dynamic impact of each event on the market state. Based on the event feature vectors, event types are analyzed, and a corresponding agent is constructed for each type of event. Each agent simulates the dynamic game process based on the current environmental state. During the simulation, the corresponding game value is calculated. The game values of each agent at different time points are integrated according to the time axis and agent type to generate a game heatmap. By simulating the dynamic game process of events through multiple agents, static event analysis is transformed into a dynamic and interactive game process simulation, which can more accurately reflect the impact of each event on value. By calculating the game value, the dynamic influence of different event types is quantified, accurately assessing the impact of each event and providing an accurate impact value reference for analyzing product value. By constructing a game heatmap, the dynamic changes in the market environment can be accurately reflected, improving the efficiency and accuracy of the valuation process.
[0023] Responding to a pre-defined valuation task, the system analyzes the target product's feature vector. Based on the time point or time period corresponding to the valuation task, it extracts heatmap slices for the corresponding time window from the game theory heatmap, extracts the corresponding features from these heatmap slices, constructs game theory vectors, configures a dynamic valuation mechanism, analyzes the target product's static attributes based on the target feature vectors, and calculates the corresponding basic valuation. It then analyzes the dynamic market game environment based on the game theory vectors and calculates the corresponding dynamic adjustment factor. Finally, it combines the basic valuation and the dynamic adjustment factor to obtain the dynamic valuation result. This dynamic valuation mechanism integrates the product's intrinsic value with the external market game environment, calculating a valuation result that reflects both the product's intrinsic value and the real-time market environment, thus improving the accuracy of the valuation result.
[0024] This application analyzes multi-source environmental data using an event analysis model to obtain event sequences, constructs a multi-agent architecture where each agent represents a type of event, and analyzes the contribution of each type of event to the environmental state through simulated decision-making and game theory processes, generating a game heatmap. The application then calculates the basic valuation and dynamic adjustment factor based on the game heatmap and a dynamic valuation mechanism, fusing them to obtain the dynamic valuation result. By combining real-time environmental events with multi-agent game theory, the application analyzes environmental changes, improving the accuracy of environmental state analysis results. Based on the game heatmap, the application calculates the basic valuation and dynamic adjustment factor, incorporating the impact of environmental changes on product value during the valuation process, thereby improving the accuracy and environmental adaptability of the valuation results.
[0025] Furthermore, based on the pre-acquired environmental data, the environmental changes are analyzed using a pre-defined event parsing model to extract the corresponding environmental events and construct an event sequence, including: S201. Preprocess the pre-acquired environmental data to obtain the first environmental data; S202. Based on the first environmental data, analyze the environmental changes through a preset event parsing model, obtain event information and cluster it to obtain an event set; S203. Arrange the events in the event set in chronological order to obtain an event sequence.
[0026] In this embodiment, the pre-acquired environmental data is preprocessed. The environmental data includes, but is not limited to, news data and industry report data. The environmental data is preprocessed by data cleaning, noise removal, normalization and other processes to obtain the first environmental data. By preprocessing the environmental data, the quality of the data can be improved, the interference of noisy data on the model analysis can be eliminated, the misjudgment of events due to data errors can be avoided, and the accuracy of the event analysis results can be improved.
[0027] For the preprocessed first environment data, technical and semantic analysis are performed using an event parsing model. Technical analysis includes extracting core elements from each event, obtaining corresponding event information, calculating the correlation keys between different event information, aggregating information describing the same event, and constructing an event sequence. Parallel analysis of technical and semantic analysis improves the comprehensiveness and accuracy of event information extraction, revealing both surface factual elements and deeper logic and intent. Clustering event information merges scattered event data, avoids information fragmentation, ensures the integrity of event information, and improves the efficiency and accuracy of feature extraction and game simulation.
[0028] Specifically, the events in the event set are arranged in chronological order to obtain an event sequence. By arranging the events, the causal and sequential relationships between them are reflected, providing a temporal reference for analyzing the context and trends of environmental changes. In the process of simulating event game, the order in which the events occur is required to provide a sequential reference for the event game process, thereby improving the logic and reliability of the dynamic valuation results.
[0029] Furthermore, based on the first environmental data, the environmental changes are analyzed using a preset event analysis model to obtain event information and cluster it to obtain an event set, including: S301. Using a preset event parsing model, perform technical and semantic analysis on the first environmental data to obtain event information; S302. Calculate the association key between event information, and cluster the event information based on the association key, with each cluster as an event, to obtain an event set.
[0030] In this embodiment, a pre-defined event analysis model is used to perform technical and semantic analysis on the first environmental data to extract complete event information. The event analysis model includes, but is not limited to, a natural language processing (NLP) model. A pre-trained NLP model is obtained by training the NLP model with a large amount of historical environmental data. The environmental data is input into the pre-trained NLP model, which performs technical and semantic analysis on the data and extracts factual elements from the data, including but not limited to the event subject, event object, numerical indicators, and corresponding time points and time periods. The model also performs semantic understanding and analysis to identify the degree of event impact and causal relationship in the data and obtain the corresponding event information.
[0031] It should be noted that technical analysis ensures the accuracy and completeness of extracting the core elements of an event, providing a factual basis for accurate event analysis; semantic analysis delves into the potential information and causal relationships in the data, understanding the deeper meaning and causal impact of the event; combining technical and semantic analysis can avoid the one-sidedness of a single analytical method, ensuring the accurate capture and in-depth interpretation of event facts, and providing rich and high-quality event data for event analysis and clustering.
[0032] like Figure 2 As shown, the process involves calculating the association keys between event information, clustering the event information based on these association keys, and aggregating scattered event information according to their relevance. Each cluster is considered an event, resulting in an event set. The similarity between event information across various dimensions generates corresponding association keys, including but not limited to entity association keys, time association keys, semantic association keys, and action association keys. Entity association keys are obtained by calculating the similarity of entity names (e.g., subject or object), time association keys by calculating the proximity of time points or the overlap of time periods, semantic association keys by calculating the cosine similarity of text descriptions using semantic vectors, and action association keys by determining whether action types are the same or belong to the same category. Appropriate weights are assigned based on the influence of each dimension's association key on the relevance, and a weighted summation method is used to calculate the comprehensive relevance score between event information, resulting in association keys. Based on these association keys, the AP clustering algorithm is used to perform cluster analysis on all event information, with each cluster representing an event, thus obtaining an event set.
[0033] It is important to emphasize that the calculation of multi-dimensional association keys ensures the accuracy and comprehensiveness of event aggregation, combining surface feature matching and semantic-level association; the clustering algorithm can automatically determine the number of event categories, avoiding the limitations of preset category numbers and improving the system's scenario adaptability; by aggregating scattered information into complete event descriptions, the problem of information fragmentation can be effectively solved, providing an information foundation for constructing accurate event sequences and improving the efficiency and accuracy of event analysis and processing.
[0034] Furthermore, for each event in the event sequence, features are extracted using a preset feature extraction model to construct a corresponding event feature vector, including: S401. Parse each event in the event sequence into different event components to obtain a set of event components; S402. Based on the event component set, extract the corresponding event component features using a preset feature extraction model to obtain the event component feature set. S403. According to the corresponding event component dimensions, the event component feature sets are fused to construct the corresponding event feature vector.
[0035] In this embodiment, each event in the event sequence is parsed into different event components, decomposed into basic constituent elements, and a set of event components is obtained. The event text is input into a preset semantic role labeling model, which includes, but is not limited to, a pre-trained deep neural network model. The deep neural network model is trained using a large amount of historical event data to obtain a pre-trained deep neural network model. The model analyzes the event text and identifies the corresponding event elements, including but not limited to event roles, time, and location, to obtain the corresponding set of event components. By parsing the events and extracting the corresponding event components, the events are transformed into corresponding discrete components, improving the accuracy and efficiency of feature processing of events.
[0036] Specifically, based on the event component set, corresponding event component features are extracted using a pre-defined feature extraction model, resulting in an event component feature set. The feature extraction model includes, but is not limited to, a pre-trained multi-channel feature learning network. Different feature extraction strategies are employed for different types of components. For textual components, a pre-trained word embedding model is used to extract fixed-dimensional vector representations, obtaining feature vectors reflecting the semantics, syntax, and contextual meaning of words. For categorical components, a pre-trained embedding model is used to map them into low-dimensional vectors, making semantically similar categories closer in the vector space. For numerical components, the model performs standardization and can directly use them as feature values. Combining the features of different components yields corresponding feature vectors, constituting the event component feature set. Feature extraction yields feature representations that reflect the core information of the event components, providing data for feature fusion and computation.
[0037] Based on the corresponding event component dimensions, the event component feature sets are fused, integrating the scattered component features into a feature vector that comprehensively represents the event, resulting in the corresponding event feature vector. All component feature vectors are aligned and organized according to their corresponding component types to obtain a feature matrix. The similarity between each component feature vector and its corresponding task context vector is calculated, and normalized using the Softmax function to obtain corresponding weight coefficients. All component feature vectors are then weighted and summed according to these weight coefficients to obtain a single, fixed-dimensional event feature vector. Feature fusion enhances key information, resulting in an accurate and information-rich overall event representation. While preserving the details of event components, it generates comprehensive features superior to those of individual components. The generated event feature vector has high task relevance, improving the accuracy and efficiency of product valuation tasks.
[0038] Furthermore, based on the event feature vector, the event game process is simulated through a pre-constructed multi-agent architecture, the game value of each agent is calculated, and a corresponding game heatmap is generated according to the game value, including: S501. Analyze event types based on event feature vectors and construct a multi-agent architecture, with each agent responsible for one event type; S502. In the multi-agent architecture, each agent simulates and analyzes the event state, simulates the event game process, calculates the game value corresponding to each agent, and obtains the game value sequence. S503. Map the game value sequence to a heatmap according to time order and agent type to construct a game heatmap.
[0039] In this embodiment, a clustering algorithm is used to perform pattern recognition on a large number of historical and real-time event feature vectors to identify stable event types. Based on the aggregation of vectors in the feature space, the clustering algorithm automatically groups events with similar features into the same cluster. Each cluster serves as an event type, including but not limited to macroeconomic policy announcements, industry technical standard updates, and major business decisions of specific companies. An independent agent is instantiated for each event type, constructing a multi-agent architecture. Each agent is not a simple event storage unit but a computational unit encapsulating specific behavioral logic and decision-making capabilities. During initialization, domain knowledge priors for the event type are set. For example, policy events are typically mandatory, while market events are diffusive. Each agent is equipped with an initialized policy network and an internal value function. The multi-agent architecture organizes the abstract, massive event stream into a finite number of predictable, interactive game participant models. By classifying events and constructing a multi-agent architecture, discrete events are divided into corresponding game participants according to their degree of impact. An agent is built for each type of event, and the game process between different types of events is simulated. This allows for accurate analysis of the impact of different types of events on the environment and product valuation, thereby improving the accuracy of product valuation results.
[0040] Specifically, each agent implements its decision-making logic through its policy network, which is typically a deep neural network. The input to this network is the current environmental state vector, which is encoded by historical event features and the influence of other agents' actions at the previous time step. The agent's behavioral logic is defined by its objective function. For example, the objective of a policy-oriented agent is to maintain market stability, while the objective of a competitive agent is to maximize its own market share. Guided by the objective function, the policy network, after training, learns the decision-making actions to take to achieve its objective.
[0041] For example, a decision action is a behavioral tendency or influence pattern adopted by the agent in the current environment for this type of event; the action space is a predefined, discrete set of options. For example, for a regulatory policy agent, action types include issuing tightening policies, issuing easing policies, maintaining the status quo, and issuing warning signals; for a market panic agent, action types include sentiment spread, sentiment easing, and localized panic. The policy network outputs not a single action, but a probability distribution for each action type. The probability distribution reflects the agent's preference for different behavioral options in the current complex and uncertain environment. For example, the probability of issuing tightening policies is 0.7, and the probability of issuing warning signals is 0.3. Sampling is performed based on the probability distribution to determine the specific actions to be executed by the agent in this round, thereby incorporating uncertainty into the simulation.
[0042] The specific actions of the agents constitute joint actions, which are input into the environment simulator to update the environment state. Through Shapley value calculation and analysis based on cooperative game theory, the analysis examines how the environment state would differ if a particular agent's action were missing, quantifying the marginal contribution of that agent's current action to the final state change. This marginal contribution is combined with the agent's own value function evaluation and normalized to obtain the agent's game value at the current time step, quantifying its influence score in this round of the game. The game values of all agents are recorded chronologically to obtain a game value sequence. By simulating the event game process through multi-agent simulation, static event analysis is transformed into a dynamic event game process, simulating the impact of events on the environment and the mutual influence between events. The calculation of game values provides a quantitative analysis of the event impact, improving the accuracy and efficiency of the analysis results.
[0043] like Figure 3 As shown, for the calculated game value sequence, a two-dimensional matrix is constructed with time as the horizontal axis and agent type as the vertical axis. Each element in the matrix corresponds to a game value at a specific time point and for a specific agent. Each element in the matrix is mapped to the corresponding position in the heatmap, and the game values from low to high are mapped as a continuous gradient from cool to warm colors to construct the game heatmap. Figure 3 In this system, the thickness of lines represents the level of the game value, with thicker lines indicating a higher game value. By constructing a game heatmap and converting game values into corresponding heatmap images, the overall game situation can be directly observed, and the changing trends of the influence of different events can be identified. This provides a reference for the dynamic valuation process, enabling the valuation process to incorporate real-time environmental changes and improving the environmental adaptability and accuracy of the valuation results.
[0044] Furthermore, in the multi-agent architecture, each agent simulates and analyzes the event state, simulates the event game process, calculates the game value for each agent, and obtains a sequence of game values, including: S601. Based on each agent in the multi-agent architecture, simulate and analyze the event state to obtain the corresponding set of decision actions; S602. Simulate environmental changes based on the set of decision actions to obtain an environmental state vector; S603. Combining the environmental state vector and the set of decision actions, simulate the event game process, analyze the contribution of the corresponding agent to the environmental state, and calculate the game value corresponding to each agent. S604. Arrange the game values in chronological order to obtain a game value sequence.
[0045] In this embodiment, for each agent in the multi-agent architecture, the event state is simulated and analyzed to obtain a corresponding set of decision actions. Each agent includes a policy network, which includes, but is not limited to, a parameterized deep neural network whose parameters define the agent's decision logic. The input to the policy network is an environment state vector, which encodes the combined state of all historical event features up to the current time step, the previous action history of each agent, and other environmental indicators. The output layer of the policy network is activated by a Softmax activation function, outputting a probability distribution in a predefined action space for the agent. The action space is a discrete set of actions designed for the event type represented by the agent. For example, for an agent representing a monetary policy event, its action space includes quantitative easing, interest rate hikes, maintaining the benchmark interest rate, and window guidance. The probability distribution output by the network serves as the agent's set of decision actions, representing the probability of the agent taking each action under the current environmental state. This process simulates the possibility of an agent choosing different strategies based on the current situation in real-world decision-making, rather than a deterministic single output. By simulating the environment and event states through intelligent agents, corresponding behavioral actions can be calculated based on dynamically changing environmental states. Adjustments can be made in real time according to changes in environmental states, thereby improving the adaptability of the intelligent agent's decision-making process to the environment.
[0046] Specifically, based on a set of decision actions, an environment simulator is used to simulate environmental changes, resulting in an environment state vector. The environment simulator includes, but is not limited to, a recurrent neural network model trained using historical data. The model input includes the environment state vector from the previous time step and the joint action with the highest probability corresponding to the probability distribution of all agents' policy network outputs at the current time step. The environment simulator learns the dynamic relationship of how executing the joint action in a given state will lead to the evolution of the environment state. After processing the input, the simulator outputs a new environment state vector. This new vector not only contains the continuation of historical information but also encodes the net effect of all agents' game behavior in the current round, including but not limited to changes in overall market sentiment, potential changes in capital flows, or adjustments in policy expectations. The updated state will serve as the new input for all agents' decisions in the next round of simulation.
[0047] It should be noted that by using an environment simulator to dynamically analyze the impact of the agent game process, the combined effect of multi-agent decisions on the environment is simulated, improving the rationality and consistency of environmental state changes. It can reflect the real-time state of the environment, and the updated environmental state vector reflects the current state of the environment in real time and provides input for the next state update process. It can continuously simulate and analyze the environmental state, improving the accuracy and real-time performance of the environmental state analysis results.
[0048] By combining the environmental state vector and the set of decision actions, the event game process is simulated. The contribution of each agent to the environmental state is analyzed, the contribution of each agent's behavior to the change of the environmental state is evaluated, and the game value corresponding to each agent is calculated. The event game process is simulated through a pre-set contribution evaluation network model, which includes, but is not limited to, a convolutional neural network model. The convolutional neural network model is trained using a large number of historical events to obtain a pre-trained convolutional neural network model. The initial environmental state vector, the set of decision actions of all agents, and the updated environmental state vector are input into the pre-trained convolutional neural network model. The model decomposes the effect of the agents' joint actions into the individual actions of each agent. The impact of the absence of an agent's action on the final environmental state is analyzed to evaluate the contribution of the corresponding agent to the environmental state. The role of the agent in promoting the environmental state towards the current outcome during the game process is quantified, and the game value corresponding to each agent is obtained.
[0049] It is important to emphasize that by quantifying the influence of intelligent agents through contribution analysis, the calculated game value enables the quantitative analysis of the influence of different types of events. This provides accurate data support for the impact of events on changes in the environmental state, and provides data support for constructing heat maps and conducting dynamic valuation, thereby improving the reliability and effectiveness of the valuation results.
[0050] Specifically, the game values are arranged in chronological order to obtain a game value sequence. By constructing the game value sequence, the instantaneous and static game values are transformed into a continuous and dynamic environmental evolution process, providing data support for analyzing the trend of environmental state changes, periodic fluctuations, and key turning points.
[0051] Furthermore, in response to the preset valuation task, the target feature vector of the target product is analyzed, mapped onto the game heatmap, and a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor to obtain the dynamic valuation result, including: S701. In response to the preset valuation task, the valuation information of the target product is analyzed through the preset task analysis model to obtain the target feature vector. S702. According to the time correspondence, map the target feature vector onto the game heat map and calculate the game vector; S703. Combining the game vector, configure a dynamic valuation mechanism to calculate the corresponding basic valuation and dynamic adjustment factor, and obtain the dynamic valuation result.
[0052] In this embodiment, in response to a preset valuation task, the valuation information of the target product is analyzed using a preset task analysis model to obtain a target feature vector. When a valuation task is received, the valuation task includes, but is not limited to, descriptive information of the target product. The descriptive information is analyzed using a preset task analysis model, which includes, but is not limited to, a deep neural network model. The deep neural network model is trained using a large amount of historical valuation task data to obtain a pre-trained deep neural network model. The valuation task is input into the pre-trained deep neural network model, and the model analyzes the task data and extracts features. According to the dimension of the time feature vector, it outputs the target feature vector. By extracting features from the valuation task, product information of different forms and sources is transformed into a unified vector format, which can eliminate the impact of data structure differences. By maintaining the consistency of the vector dimension with the event feature extraction model, it is ensured that the product vector and the event game environment can be calculated. The extracted target feature vector reflects the core value of the target product, improving the accuracy of the product valuation process.
[0053] Specifically, according to the time correspondence, the target feature vector is mapped onto the game heatmap, and the game vector is calculated. Through spatiotemporal mapping and correlation analysis, the target product is associated with the environment, and the product valuation is mapped to the corresponding spatiotemporal environment. This ensures that the valuation results can be analyzed in conjunction with the time environment, and the extracted game vectors can reflect the dynamic real-time environment, providing accurate data support for the dynamic evaluation and adjustment of product value.
[0054] The dynamic valuation mechanism includes calculating the corresponding basic valuation and dynamic adjustment factors based on the calculated game vectors. It integrates the product's intrinsic static value with the impact of the external dynamic environment, and weights the basic valuation to obtain the dynamic valuation result. Through the dynamic valuation mechanism, combining the product's static value and environmental impact, the dynamic evaluation of the valuation result is unified with the actual value. The basic valuation ensures that the valuation process is based on the product's actual value, while the combination of dynamic environmental impact can quantify the real-time impact of external environmental changes on the product value, improving the timeliness and accuracy of the product valuation process. The resulting dynamic valuation result can reflect the product value and respond to environmental changes, thereby improving the real-time performance and accuracy of the product valuation result.
[0055] Furthermore, according to the time correspondence, the target feature vector is mapped onto the game heatmap, and the game vector is calculated, including: S801. Analyze the valuation time point from the target feature vector; S802. Extract heatmap slices of the game heatmap with preset time windows according to the estimated time points; S803. Analyze the agent game values in the heat map slices, extract the corresponding game features, and construct game vectors.
[0056] In this embodiment, the valuation time point is parsed from the target feature vector. When constructing the target feature vector, the product data contains corresponding time information. The corresponding specific time is extracted from the product feature representation to obtain the corresponding valuation time point. By parsing the valuation time, the accuracy of the valuation process in the time dimension is ensured. The product value assessment is associated with the time point to avoid valuation deviations caused by time mismatch and improve the accuracy of the valuation results.
[0057] Based on the extracted valuation time points, heatmap slices within preset time windows are extracted from the game theory heatmap. After extracting the valuation time points, time windows are set according to the accuracy requirements of the valuation process. Heatmap data corresponding to the valuation time points are then extracted forward from the time windows to obtain the game value data for all agent types within the time window, which are used as heatmap slices. The heatmap slices retain the structure of the heatmap along the agent type dimension. By extracting heatmap slices corresponding to the time windows, the calculation of all historical game data is avoided, improving computational efficiency. Focusing on the most relevant historical stages near the valuation time points ensures the timeliness and relevance of the environmental impact factor analysis. The extracted slice data provides accurate data input for the product valuation process.
[0058] Specifically, the game values of agents in the heatmap slices are analyzed to extract corresponding game features and construct game vectors. A pre-set heatmap analysis model, including but not limited to a neural network model, is used to analyze the heatmap slices. This model is trained using a large amount of historical heatmap data to obtain a pre-trained neural network model. The heatmap slices and target feature vectors are input into the pre-trained neural network model. The model extracts features from the time-series game values of each agent type, analyzes their changing trends within the time window, generates time-series features for each agent, calculates the correlation and importance weights between the target product features and the event types represented by each agent, and concatenates and fuses the weighted agent time-series features to obtain the game vector. Through feature extraction and fusion, the efficiency and accuracy of the heatmap slice analysis process are improved, ensuring that the extracted game features are highly correlated with the target product, and that the game vectors are strongly correlated with the current valuation task, thereby achieving dynamic valuation of the target product.
[0059] Furthermore, by combining the game vectors, a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor, resulting in a dynamic valuation result, including: S901. Based on the target feature vector, analyze the basic value of the target product through a preset basic valuation model, and calculate the corresponding basic valuation. S902. Based on the game vector, analyze the impact of the agent's game state on value through a preset dynamic adjustment model, and calculate the corresponding dynamic adjustment factor. S903. The basic valuation and the valuation adjusted by the dynamic adjustment factor are weighted and fused to obtain the dynamic valuation result. The weights in the weighted fusion process are calculated based on the sensitivity information in the game vector.
[0060] In this embodiment, based on the target feature vector, the fundamental value of the target product is analyzed using a pre-defined fundamental valuation model, and the corresponding fundamental valuation is calculated. The fundamental valuation model includes, but is not limited to, a deep learning regression model. A pre-trained deep learning regression model is trained using a large amount of historical target feature vector data. The target feature vector is input into the pre-trained deep learning regression model, and the model analyzes the value of the target product based on the target feature vector, obtaining the fundamental value of the target product and outputting the corresponding fundamental valuation. By analyzing the fundamental valuation of the target product, a stable value benchmark unaffected by market fluctuations is provided for the valuation process, ensuring that the valuation result is based on the product's intrinsic value. Using the fundamental valuation as a benchmark for the dynamic valuation adjustment process, the final dynamic valuation result can distinguish between the portion of value change originating from the product itself and the portion originating from the external environment, improving the interpretability of the valuation result.
[0061] Specifically, based on game vectors, a pre-defined dynamic adjustment model analyzes the impact of the agent's game state on value and calculates the corresponding dynamic adjustment factor. The dynamic adjustment model includes, but is not limited to, a recurrent neural network (RNN) model. A pre-trained RNN model is obtained by training the RNN model with a large amount of historical game vector data. The model learns the relationship between different game environment patterns and the degree of deviation of the product's value from its basic value within the corresponding period, establishing a mapping from game characteristics to value adjustment amounts. Game vectors are input into the pre-trained RNN model to analyze the game situation and the basic value adjustment amount, outputting the dynamic adjustment factor. By calculating the dynamic adjustment factor, external value components caused by multi-party market games that are not included in the product's static attributes can be identified, enabling the valuation results to reflect the dynamic changes in the market environment in a timely manner. The model transforms the features extracted from the game environment into specific value adjustment amounts, quantifying the adjustment situation, thereby improving the accuracy and timeliness of the valuation results.
[0062] The basic valuation and the valuation adjusted by the dynamic adjustment factor are weighted and fused to obtain the dynamic valuation result. Based on game vector analysis, which characterizes the intensity of environmental uncertainty and the target product's sensitivity to volatility (including but not limited to the variance of game values and the agent's game strength), the weights of the basic valuation and the adjusted valuation are calculated based on sensitivity information. The basic valuation and the valuation adjusted by the dynamic adjustment factor are then weighted and summed according to their respective weights to obtain the dynamic valuation result. This weighted fusion, dynamically allocating weights based on environmental conditions and product characteristics, effectively suppresses excessive deviation of the valuation result from the actual value in extremely volatile market environments, improves the valuation result's resistance to interference and responsiveness, and thus enhances the accuracy and timeliness of the target product's valuation result.
[0063] like Figure 4 As shown, a large-scale model dynamic valuation control system is used to implement a large-scale model dynamic valuation control method, including: The environmental analysis module analyzes environmental changes based on pre-acquired environmental data using a preset event parsing model, identifies corresponding environmental events, and constructs an event sequence. The event analysis module extracts features for each event in the event sequence using a preset feature extraction model and constructs a corresponding event feature vector. The game heatmap construction module, based on the event feature vector, simulates the event game process through a pre-built multi-agent architecture, calculates the game value of each agent, and generates the corresponding game heatmap according to the game value; The dynamic valuation module, in response to a preset valuation task, analyzes the target feature vector of the target product, maps it to the game heatmap, configures a dynamic valuation mechanism to calculate the corresponding basic valuation and dynamic adjustment factor, and obtains the dynamic valuation result to dynamically control the valuation process.
[0064] In this embodiment, the environmental analysis module receives multi-source heterogeneous environmental data and analyzes the data using a preset event parsing model. It identifies, extracts, and clusters the data to obtain environmental events, constructing an event sequence in chronological order. By parsing the event sequence from the environmental data, it can provide data support for the valuation analysis process. For each independent event in the sequence, the event analysis module decomposes and extracts the core attributes of the event using a preset feature extraction model, constructing a high-dimensional event feature vector. This model quickly and accurately extracts feature information, providing standardized, high-quality input for the game simulation process. This enables the system to understand and process event patterns, enhancing the system's intelligence and representation capabilities.
[0065] Specifically, the game heatmap construction module constructs a multi-agent architecture based on the input event feature vectors. Each agent represents a type of event. By simulating their decision-making process and mutual game interaction, the module calculates the game value of each agent at different time points. Based on the calculated game values, a game heatmap is constructed. By introducing a dynamic game simulation mechanism, static event analysis is elevated to dynamic game theory. This allows for the analysis of the strength and changing trends of the influence of different event types over time, providing a dynamic environment input for the dynamic valuation process and improving the accuracy of the valuation results. The dynamic valuation module responds to valuation requests, analyzes the characteristics of the target product, maps them to the game heatmap, configures the dynamic valuation mechanism, and outputs dynamic valuation results. This integrates the product's intrinsic static value with the influence of the external dynamic environment, improving the accuracy and environmental adaptability of the valuation results.
[0066] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.
Claims
1. A dynamic estimation control method for large-scale models, characterized in that, include: Based on the pre-acquired environmental data, the environmental changes are analyzed using a pre-set event parsing model to extract the corresponding environmental events and construct an event sequence. For each event in the event sequence, features are extracted using a preset feature extraction model to construct a corresponding event feature vector. Based on the event feature vector, the event game process is simulated through a pre-constructed multi-agent architecture, the game value of each agent is calculated, and the corresponding game heatmap is generated according to the game value. In response to the preset valuation task, the target feature vector of the target product is analyzed and mapped onto the game heat map. A dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor to obtain the dynamic valuation result, so as to dynamically control the valuation process.
2. The large-scale model dynamic estimation control method according to claim 1, characterized in that, The step involves analyzing environmental changes based on pre-acquired environmental data using a preset event parsing model, extracting corresponding environmental events, and constructing an event sequence, including: The pre-acquired environmental data is preprocessed to obtain the first environmental data; Based on the first environmental data, the environmental changes are analyzed through a preset event parsing model to obtain event information and cluster it to obtain an event set; Arrange the events in the event set in chronological order to obtain the event sequence.
3. The large-scale model dynamic estimation control method according to claim 2, characterized in that, Based on the first environmental data, the environmental changes are analyzed using a preset event parsing model to obtain event information and cluster it to obtain an event set, including: Using a pre-defined event parsing model, technical and semantic analyses are performed on the first environmental data to obtain event information. Calculate the association keys between event information, cluster the event information based on the association keys, and treat each cluster as an event to obtain an event set.
4. The large-scale model dynamic valuation control method according to claim 1, characterized in that, For each event in the event sequence, features are extracted using a preset feature extraction model to construct a corresponding event feature vector, including: Each event in the event sequence is parsed into different event components, resulting in a set of event components; Based on the event component set, the corresponding event component features are extracted using a preset feature extraction model to obtain the event component feature set; According to the corresponding event component dimensions, the event component feature sets are fused to construct the corresponding event feature vector.
5. The large-scale model dynamic estimation control method according to claim 1, characterized in that, Based on the event feature vector, the event game process is simulated through a pre-constructed multi-agent architecture. The game value of each agent is calculated, and a corresponding game heatmap is generated according to the game value, including: Based on the analysis of event feature vectors, an event type is constructed, and a multi-agent architecture is built, with each agent responsible for one event type. In the multi-agent architecture, each agent simulates and analyzes the event state, simulates the event game process, calculates the game value corresponding to each agent, and obtains the game value sequence. The game value sequence is mapped to a heatmap according to time order and agent type to construct a game heatmap.
6. The large-scale model dynamic valuation control method according to claim 5, characterized in that, In the multi-agent architecture, each agent simulates and analyzes the event state, simulates the event game process, calculates the game value corresponding to each agent, and obtains a sequence of game values, including: Based on each agent in the multi-agent architecture, the event state is simulated and analyzed to obtain the corresponding set of decision actions; Based on the set of decision actions, simulate environmental changes to obtain an environmental state vector; By combining the environmental state vector and the set of decision actions, the event game process is simulated, the contribution of the corresponding agent to the environmental state is analyzed, and the game value corresponding to each agent is calculated. The game values are arranged in chronological order to obtain a game value sequence.
7. The large-scale model dynamic valuation control method according to claim 1, characterized in that, In response to a preset valuation task, the target feature vector of the target product is parsed, mapped onto the game heatmap, and a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor to obtain the dynamic valuation result, including: In response to a pre-defined valuation task, the valuation information of the target product is analyzed using a pre-defined task analysis model to obtain the target feature vector. According to the time correspondence, the target feature vector is mapped onto the game heatmap, and the game vector is calculated; By combining the game vectors, a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor, and the dynamic valuation result is obtained.
8. The large-scale model dynamic valuation control method according to claim 7, characterized in that, The step of mapping the target feature vector to the game heatmap according to the time correspondence and calculating the game vector includes: Analyze the valuation time points from the target feature vector; Based on the valuation time points, heatmap slices with preset time windows are extracted from the game heatmap. The agent game values in the heatmap slices are analyzed, the corresponding game features are extracted, and game vectors are constructed.
9. The large-scale model dynamic valuation control method according to claim 7, characterized in that, By combining the aforementioned game vectors, a dynamic valuation mechanism is configured to calculate the corresponding basic valuation and dynamic adjustment factor, resulting in a dynamic valuation result, including: Based on the target feature vector, the basic value of the target product is analyzed through a preset basic valuation model, and the corresponding basic valuation is calculated. Based on the game vector, the impact of the agent's game state on value is analyzed through a preset dynamic adjustment model, and the corresponding dynamic adjustment factor is calculated. The basic valuation and the valuation adjusted by the dynamic adjustment factor are weighted and fused to obtain the dynamic valuation result. The weights in the weighting and fusion process are calculated based on the sensitivity information in the game vector.
10. A large-scale model dynamic valuation control system, characterized in that, A method for implementing a large model dynamic estimation control as described in any one of claims 1 to 9, comprising: The environmental analysis module analyzes environmental changes based on pre-acquired environmental data using a preset event parsing model, identifies corresponding environmental events, and constructs an event sequence. The event analysis module extracts features for each event in the event sequence using a preset feature extraction model and constructs a corresponding event feature vector. The game heatmap construction module, based on the event feature vector, simulates the event game process through a pre-built multi-agent architecture, calculates the game value of each agent, and generates the corresponding game heatmap according to the game value; The dynamic valuation module, in response to a preset valuation task, analyzes the target feature vector of the target product, maps it to the game heatmap, configures a dynamic valuation mechanism to calculate the corresponding basic valuation and dynamic adjustment factor, and obtains the dynamic valuation result to dynamically control the valuation process.