Generation model, training method, system and application based on horse race type multi-agent confrontation collaboration
By adopting a "racecourse"-style multi-agent adversarial collaborative generative model, the problems of pattern collapse, training instability, and convergence difficulty in GAN training are solved, achieving faster convergence speed and better generalization performance, which is applicable to fields such as image generation, data augmentation, and time series prediction.
Patent Information
- Application Number
- CN202511718687.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-21
- Publication Date
- 2026-03-03
AI Technical Summary
Existing generative adversarial networks (GANs) suffer from problems such as pattern collapse, training instability, convergence difficulties, and hyperparameter sensitivity. Furthermore, multi-agent methods lack effective cooperative competition mechanisms, resulting in slow training convergence speed and difficulty in forming synergistic effects.
We adopt a generative model based on multi-agent adversarial collaboration in a "racecourse" style. Through a dynamically evolving "teacher team" generator and student discriminator, combined with a periodic elite evaluation and last-place elimination mechanism, we ensure the dynamic balance and diversity between the generator and discriminator. We use a hybrid dataset for high-intensity adversarial training to optimize the training strategy.
It improves the training stability and convergence speed of the model, reduces the dependence on real data, enhances the generalization ability and scalability of the model, and supports dynamically adjusting the number of generators to adapt to tasks of different complexities.
Smart Images
Figure CN121599009A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and deep learning technology, specifically relating to a generative model, training method, system and application based on "racecourse"-style multi-agent adversarial collaboration. Background Technology
[0002] Generative Adversarial Networks (GANs), as one of the most important generative models today, learn data distributions through an adversarial game between a generator and a discriminator. However, traditional GAN training suffers from several inherent problems:
[0003] Pattern collapse problem: Generators often can only generate samples with limited diversity, failing to cover all patterns in the real data. This phenomenon is particularly evident when the training data distribution is complex, resulting in a lack of diversity in the generated samples.
[0004] Training instability: The abilities of the generator and discriminator are difficult to maintain a dynamic balance, which can easily lead to gradient vanishing or gradient exploding. Premature convergence of the discriminator can cause the generator to fail to obtain effective gradient signals.
[0005] Convergence difficulties: Under complex data distributions, models often struggle to converge to the ideal state, the training process is prone to oscillations, and the loss function is difficult to decrease steadily.
[0006] Hyperparameter sensitivity: Traditional GANs are extremely sensitive to hyperparameter settings, and even small changes can lead to huge differences in training results.
[0007] Existing improvement methods include Multi-Generator Adversarial Networks (MGAN) and Wasserstein GAN (WGAN). MGAN mitigates pattern collapse by introducing multiple generators, but lacks an effective competition mechanism to promote different generators learning different patterns of the data distribution. Furthermore, these methods cannot guarantee the diversity of the generator population, often resulting in multiple generators converging to the same pattern. While WGAN improves training stability to some extent, it still suffers from slow convergence speed when dealing with complex distributions.
[0008] Furthermore, most existing multi-agent methods employ simple parallel architectures, lacking effective collaborative and competitive mechanisms, and thus failing to fully leverage the advantages of swarm intelligence. During training, each agent often optimizes independently, making it difficult to achieve a synergistic effect. Summary of the Invention
[0009] The technical problem to be solved by this invention is to provide a generative model, training method and application based on "racecourse" style multi-agent adversarial network, which solves the problems of slow training convergence speed and difficulty in forming synergistic effects among agents in the prior art.
[0010] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0011] Based on a "racecourse"-style multi-agent adversarial collaborative generative model, it includes a dynamically evolving "teacher team" generator and a student discriminator guided by elite teachers. The dynamically evolving "teacher team" generator comprises an optimal generator module, a discriminator group module, and an auxiliary scorer module. The optimal generator module takes historical data as input and generates a data sequence that approximates the distribution of real data. The discriminator group module distinguishes whether the input data is real historical data or data generated by the generator and outputs predicted data. The auxiliary scorer module scores the data generated by the generator and the predicted data output by the discriminator. The student discriminator undergoes high-intensity adversarial training using a mixed dataset of real data and data generated by the optimal generator.
[0012] The "teacher team" consists of high-level intelligent agents, and the agents are updated through regular evaluation and a bottom-ranking elimination mechanism.
[0013] The student discriminator includes a CNN discriminator and a parameter setting module.
[0014] The training method for a generative model based on "racecourse"-style multi-agent adversarial cooperation includes the following steps:
[0015] Step 1: Construct the initial models of the "teacher group" generator and the student discriminator, and initialize them;
[0016] Step 2: Based on the periodic elite evaluation and bottom-ranking elimination mechanism, dynamically update the agents of the "teacher team" generator model to maintain the elite teacher team;
[0017] Step 3: The student discriminator is guided and trained by a team of elite teachers. Under the joint guidance of the elite teachers, high-intensity adversarial training is carried out using a mixed dataset generated by real data and the optimal generator until the performance of the student discriminator reaches a stable state in the new evaluation period.
[0018] Step 4: Continuous evolution cycle and system optimization, construct a new intelligent agent monitoring system, analyze the monitoring data in real time, dynamically adjust the evaluation frequency according to the group evolution speed, optimize the overall training strategy, and continuously cycle the training to keep the generative model of "racecourse" type multi-agent adversarial in the optimal state.
[0019] In step 2, the specific process of the periodic elite evaluation and bottom-ranking elimination mechanism is as follows:
[0020] The evaluation and replacement process is executed at a preset cycle frequency. In each evaluation cycle, the performance of all discriminators in the teacher team is quantitatively scored using an auxiliary scorer. Based on the scoring results, the teacher team members are ranked, and a strict bottom-ranked elimination and replacement is implemented. The teacher discriminator with the lowest ranking is removed from the teacher team, and the resources it occupies are released. At the same time, a brand-new student discriminator with randomly initialized parameters is formally introduced into the teacher team as a new intelligent agent to replace the eliminated one, thus completing the dynamic update and maintenance of the overall capability of the teacher team within a cycle.
[0021] The initialization of the new intelligent agent adopts a differentiated strategy, prioritizing scarce architecture types to maintain group diversity. The parameter initialization adopts a specific distribution strategy to ensure that the new intelligent agent maintains its individuality and has development potential. The number of parameters of the new intelligent agent is maintained within a reasonable range of the average level of the teacher group.
[0022] In step 3, the specific process of providing guidance and training to students' discriminators through an elite team of teachers is as follows:
[0023] After the agent of the "teacher team" generator model is updated, the training process for the next-generation student discriminator is initiated. First, the elite teachers of the entire teacher team will jointly serve as mentors. Under the joint guidance of the elite teacher team, the student discriminator will undergo high-intensity adversarial training using a mixed dataset generated by real data and the optimal generator. During the training process, each elite teacher team provides the student with a stable gradient signal, and the average value is taken as the student gradient. When the prediction of the student discriminator is inconsistent with the collective prior judgment of the teacher team, the gradient of the student model is corrected according to the teacher's output until its performance reaches a stable state in the new evaluation period.
[0024] Step 4 specifically includes:
[0025] Step 4.1: Dynamically adjust the evaluation frequency according to the population evolution speed. When the population performance improves rapidly, shorten the evaluation cycle; when the population is stable, maintain the standard evaluation rhythm; when performance fluctuations occur, extend the evaluation cycle.
[0026] Step 4.2: Dynamically adjust the elimination ratio based on the population diversity index, and set upper and lower thresholds;
[0027] Step 4.3: Establish a diversity monitoring system to track architectural diversity, strategy diversity, and performance distribution uniformity, and analyze the monitoring data in real time;
[0028] Step 4.4: Collect training performance data for each agent, regularly introduce new architecture types, eliminate outdated architectures, optimize the overall training strategy, and dynamically adjust evaluation metrics and weight allocation based on feedback information during the evolution process.
[0029] Step 4.5: Establish a knowledge accumulation mechanism to systematically preserve successful training strategies and lessons learned.
[0030] The diversity monitoring system includes evolution efficiency monitoring and resource utilization efficiency monitoring. Evolution efficiency monitoring focuses on the growth rate of new intelligent agents and the overall improvement rate of the group, while resource utilization efficiency monitoring is used for the rational allocation and use of computing resources.
[0031] The system is based on a "racecourse"-style multi-agent adversarial cooperative generative system. The system includes a processor, which applies the aforementioned "racecourse"-style multi-agent adversarial cooperative generative model to construct an agent training environment with a competitive ranking mechanism.
[0032] The application of a generative model based on a "racecourse"-style multi-agent adversarial collaboration is used in the fields of image generation, data augmentation, time series synthesis and prediction, natural language processing, and drug molecule design; among which, the task of time series data synthesis and prediction includes stocks, futures, and cryptocurrencies.
[0033] Compared with the prior art, the present invention has the following beneficial effects:
[0034] 1. This method uses less real-world data for training, which can reduce overfitting.
[0035] 2. Improve the generalization of the model by using a competitive ranking mechanism to allow discriminators of different categories to learn from each other, that is, to learn the advantages of other categories.
[0036] 3. Significantly improved training stability: The "racecourse"-like competitive environment naturally maintains a dynamic balance between the generator and the discriminator. Excellent scalability: The system supports dynamically increasing or decreasing the number of generators, adapting to tasks of varying complexity.
[0037] 4. Significantly faster convergence speed: The differentiated training strategy ensures the optimal allocation of computing resources. Attached Figure Description
[0038] Figure 1 This is a diagram showing the overall architecture of the racetrack-style training system of the present invention. Detailed Implementation
[0039] The structure and working process of the present invention will be further described below with reference to the accompanying drawings.
[0040] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a generative adversarial network training method and system based on a "racecourse"-style elite teacher group and a bottom-ranking elimination mechanism. By establishing a continuously evolving agent ecosystem, the continuous improvement of group performance is ensured. A dynamically evolving agent training system is constructed by simulating the survival-of-the-fittest mechanism of a "racecourse." The system maintains a "teacher group" composed of high-level agents, and continuously optimizes the group composition through regular evaluation and a bottom-ranking elimination mechanism. Specifically, the system continuously eliminates the weakest agents and introduces new individuals to be guided by the best agents, forming a virtuous cycle of continuous evolution.
[0041] The generative model based on "racecourse"-style multi-agent adversarial training includes a dynamically evolving "teacher team" generator and a student discriminator guided by elite teachers. The dynamically evolving "teacher team" generator comprises an optimal generator module, a discriminator group module, and an auxiliary scorer module. The optimal generator module takes historical data as input and generates a data sequence that approximates the distribution of real data. The discriminator group module distinguishes whether the input data is real historical data or data generated by the generator and outputs predicted data. The auxiliary scorer module scores the data generated by the generator and the predicted data output by the discriminator. The student discriminator undergoes high-intensity adversarial training using a mixed dataset of real data and data generated by the optimal generator.
[0042] The training method for generative models based on "racecourse"-style multi-agent adversarial models includes the following steps:
[0043] Step 1: Construct the initial models of the "teacher group" generator and the student discriminator, and initialize them;
[0044] Step 2: Based on the periodic elite evaluation and bottom-ranking elimination mechanism, dynamically update the agents of the "teacher team" generator model to maintain the elite teacher team;
[0045] Step 3: The student discriminator is guided and trained by a team of elite teachers. Under the joint guidance of the elite teachers, high-intensity adversarial training is carried out using a mixed dataset generated by real data and the optimal generator until the performance of the student discriminator reaches a stable state in the new evaluation period.
[0046] Step 4: Continuous evolution cycle and system optimization, construct a new intelligent agent monitoring system, analyze the monitoring data in real time, dynamically adjust the evaluation frequency according to the group evolution speed, optimize the overall training strategy, and continuously cycle the training to keep the generative model of "racecourse" type multi-agent adversarial in the optimal state.
[0047] Specific embodiments, such as Figure 1 As shown:
[0048] This embodiment is implemented using the PyTorch 2.0 deep learning framework and Python 3.9 as the development language. The hardware environment includes an NVIDIA A100 GPU, CUDA 11.7, and 64GB of memory.
[0049] Detailed network architecture configuration:
[0050] The generator cluster contains 1 member:
[0051] Transformer generator: two-layer encoder and decoder, 128-dimensional hidden layer, four attention heads.
[0052] The discriminator cluster contains two heterogeneous members:
[0053] CNN discriminator: two 1D convolutional layers and three mapping layers, with a kernel size of 1x3 and padding of 1.
[0054] Transformer discriminator, single-layer encoder, hidden layer dimension 256, four attention heads.
[0055] The training parameters are set as follows:
[0056] Optimizer: Adam (β1=0.5, β2=0.999)
[0057] Learning rate: 2e-5 for generator, 2e-5 for discriminator
[0058] Batch size: 64
[0059] Training rounds: 100 rounds for the generator's initial learning phase, 50 rounds for the discriminator's initial learning phase, and 50 rounds for the racetrack training phase.
[0060] Based on the above model structure, a novel paradigm for generative model training is provided. By simulating the competitive evolutionary mechanism in nature, it effectively solves several fundamental problems in deep learning. Specifically, as shown in Algorithm 1 below:
[0061]
[0062] This system is not only applicable to generative adversarial networks, but can also be extended to other types of generative models and machine learning tasks. Specifically, it includes the following steps:
[0063] 1. System Initialization and Module Construction
[0064] Construct a reverse teacher system comprising three core modules, including:
[0065] Optimal Generator (G): Pre-training is performed using the MSE (Mean Sequence Equation) of the generator G and the real data as the loss, ultimately yielding a relatively good generator G. The goal of the generator is to mimic the distribution of the real data to synthesize data sequences.
[0066] Discriminator group (D): a group of m discriminators (D1, D2, ..., Dm) with different structures and initial parameters. m The discriminator group is pre-trained using the MSE of the next numerical prediction and the actual data as the loss, and finally obtained. The task of each discriminator is to directly predict the subsequent numerical values from the input data sequence.
[0067] Auxiliary Scorer (S): An independently trained performance evaluation module. It does not participate in direct adversarial training, but rather objectively and quantitatively scores the generator's data generation quality and the discriminator group's price prediction ability based on a set of predefined quantitative metrics (such as the root mean square error between predicted and true values), providing a basis for subsequent identification of strong and weak parties.
[0068] 2. Periodic elite evaluation and bottom-ranking elimination mechanism
[0069] The system executes the evaluation and replacement process at a preset cycle frequency. In each evaluation cycle, the auxiliary scorer S is used to quantitatively score the performance of all discriminators in the teacher team; based on the scoring results, the teacher team members are ranked, and a strict bottom-ranked elimination and replacement is implemented. The rules are as follows: the teacher discriminator with the lowest ranking is removed from the teacher team, and the resources it occupies are released; at the same time, a brand-new student discriminator with randomly initialized parameters is formally introduced into the teacher team to replace the eliminated one, thereby ensuring the dynamic updating and maintenance of the overall ability of the teacher team.
[0070] The initialization of the new agent employs a differentiated strategy. The architecture selection is based on the diversity analysis of the current teacher group, prioritizing scarce architecture types to maintain group diversity. Parameter initialization uses a specific distribution strategy to ensure that the new agent retains its individuality while possessing development potential.
[0071] The size of the new intelligent agent is precisely controlled, and the number of parameters is maintained within a reasonable range of the average level of the teaching team, ensuring sufficient expressive power while avoiding overly complex model structures.
[0072] 3. Elite teacher-led training for student discriminators
[0073] After the replacement operation is completed, the system immediately initiates the training process for the new generation student discriminator. First, the entire team of elite teachers jointly provides guidance. Under the joint guidance of this team, the student discriminator undergoes intensive adversarial training using a mixed dataset consisting of a small amount of real data and data generated by the optimal generator (the proportion of real data depends on the task difficulty; see the experimental results comparison table for details). During training, each team of elite teachers provides a stable gradient signal to the student, and the average value is taken as the student's gradient. When the student discriminator's predictions are inconsistent with the collective prior judgments of the teachers, the system performs gradient correction on the student model based on the teachers' outputs, thereby accelerating its learning process until its performance reaches a stable state within the new evaluation period.
[0074] 4. Continuous Evolution Cycle and System Optimization
[0075] The evaluation frequency is dynamically adjusted according to the population's evolution speed. When the population's performance improves rapidly, the evaluation cycle is shortened to accelerate evolution; when the population is stable, the standard evaluation rhythm is maintained; and when performance fluctuations occur, the evaluation cycle is extended to ensure stability.
[0076] The elimination ratio is dynamically adjusted based on the population diversity indicator, with reasonable upper and lower limits set to avoid excessive or insufficient elimination.
[0077] Establish a diversity monitoring system to track architectural diversity, strategy diversity, and performance distribution uniformity. Evolutionary efficiency monitoring focuses on the growth rate of new intelligent agents and the overall improvement rate of the group. Resource utilization efficiency monitoring ensures the rational allocation and use of computing resources.
[0078] Real-time analysis of monitoring data provides a basis for system optimization decisions.
[0079] Regularly introduce new architecture types and replace outdated ones. Collect training performance data for each agent to optimize the overall training strategy. Dynamically adjust evaluation metrics and weight allocation based on feedback during the evolution process.
[0080] Establish a knowledge accumulation mechanism to systematically preserve successful training strategies and lessons learned for future training reference.
[0081] Through the above technical solutions, this invention establishes a self-evolving and continuously optimized intelligent agent training ecosystem, which can effectively improve the training effect and final performance of generative adversarial networks.
[0082] The aforementioned generative model based on "racecourse"-style multi-agent adversarial interaction can be applied to fields including but not limited to image generation, data augmentation, time series synthesis and prediction, natural language processing, and drug molecule design; among them, the task of time series data synthesis and prediction includes stocks, futures, and cryptocurrencies.
[0083] Example 1: The above model is applied to the field of "a certain currency" prediction. Specifically,
[0084] Optimal Generator (G): n heterogeneous generators take historical daily candlestick data of a currency (such as opening price, highest price, lowest price, trading volume, etc.) as input for each generator, and their goal is to generate a synthetic candlestick data sequence that is highly consistent with the distribution of real data.
[0085] Discriminator Group (D): m discriminators with the same structure but different initial parameters. The task of each discriminator is to distinguish whether the input data is a real historical candlestick sequence or a sequence synthesized by the generator, and at the same time directly predict the value of the closing price of the next day.
[0086] This model can be widely applied to the synthesis and forecasting of financial time series data, and it has significant application value, particularly in the futures market, the US stock market, and the cryptocurrency market. Specifically, it can be used for:
[0087] 1. Synthesize a statistically accurate price series for a specific currency for strategy backtesting or data augmentation.
[0088] 2. Predict the closing price trend of the next day to assist in quantitative trading decisions.
[0089] 3. In extreme market conditions (such as sharp rises and falls), the robustness and generalization ability of the model can be enhanced through the inverse teacher mechanism.
[0090] 4. Multi-period and multi-factor fusion time series modeling enhances the model's ability to capture market structure.
[0091] The system generates synthetic candlestick data through a generator, predicts price trends through a discriminator, and assists a scorer in evaluating the quality of generation and prediction, thus achieving closed-loop optimization from data generation to price prediction.
[0092] Example 2: To verify the effectiveness of the algorithm presented in this example, four representative stock index datasets were selected, including two domestic indices and two international indices: processed_CSI300_day.csv (CSI 300 Index), processed_SSE50_day.csv (SSE 50 Index), processed_SP500_day.csv (S&P 500 Index), and processed_DowJones_day.csv (Dow Jones Industrial Average). Each dataset was divided into training, testing, and validation sets in a ratio of 7.5:1.25:1.25. All models were tested on the validation set, and the evaluation metrics used included mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). Detailed experimental results for the four datasets are shown in Tables 1 to 4, with a random seed of 5 used in the experiments.
[0093] Normal training requires 75% real data, while the algorithm in this paper only needs to use 37.5% real data and 37.5% virtual data (data generated by the generator) for training in the racetrack stage, and its performance can even exceed the results of normal training.
[0094] Table 1. Results of Racecourse and Regular Training in the S&P 500 Index (USA)
[0095]
[0096] Table 2 Results of Racecourse and Regular Training in the Dow Jones Index (USA)
[0097]
[0098] Table 3 Results of Racecourse and Regular Training in the CSI 300 Index (China)
[0099]
[0100] Table 4 Results of Racecourse and Regular Training in the SSE 50 Index (China)
[0101]
[0102] The data from the four tables above demonstrate that this algorithm achieves superior generalization performance with less data. We use five metrics—mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and goodness of fit (R²)—to comprehensively evaluate the model's generalization ability. As shown in Tables 1 and 2, using only half of the normal training data, the model based on the racetrack algorithm outperforms the model trained on the full dataset across all five metrics. Although its performance is slightly inferior to the latter in more challenging tasks (Tables 3 and 4), considering that it only consumes half the training data, this result sufficiently demonstrates the algorithm's excellent data efficiency.
[0103] This solution also provides a generative system based on "racecourse"-style multi-agent adversarial collaboration, characterized in that: the system includes a processor, which applies the generative model based on "racecourse"-style multi-agent adversarial collaboration to construct an agent training environment with a competitive ranking mechanism.
[0104] In summary, this invention, through a competitive ranking mechanism and diverse rewards, forces different generators to learn different patterns of data distribution, thus completely solving the pattern collapse problem. At the same time, the ranking mechanism and multiple monitoring metrics provide a complete interpretable framework for model training, making it easier to understand the learning state and evolution process of each generator.
[0105] Those skilled in the art can make various modifications and alterations to this invention based on the above teachings. For example, the group size can be adjusted, the network architecture can be changed, and different distance metric functions can be used. These equivalent forms also fall within the scope defined by the appended claims.
[0106] It should be understood that this solution is not limited to the specific embodiments described above. Devices and structures not described in detail herein should be understood as being implemented in a manner common to the art. Any person skilled in the art can make many possible variations and modifications to this solution, or modify it into equivalent embodiments, without departing from the scope of this solution, using the methods and techniques disclosed above. This does not affect the substantive content of this solution. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this solution, without departing from its scope, still fall within the protection scope of this solution.
Claims
1. A generative model based on "racecourse"-style multi-agent adversarial cooperation, characterized by: It includes a dynamically evolving "teacher team" generator and a student discriminator guided by elite teachers. The dynamically evolving "teacher team" generator consists of an optimal generator module, a discriminator group module, and an auxiliary scorer module. The optimal generator module takes historical data as input and generates a data sequence that approximates the distribution of real data. The discriminator group module distinguishes whether the input data is real historical data or data generated by the generator and outputs predicted data. The auxiliary scorer module scores the data generated by the generator and the predicted data output by the discriminator. The student discriminator undergoes high-intensity adversarial training using a mixed dataset of real data and data generated by the optimal generator.
2. The generative model based on "racecourse"-style multi-agent adversarial cooperation as described in claim 1, characterized in that: The "teacher team" consists of high-level intelligent agents, which are updated through regular evaluation and a bottom-ranking elimination mechanism; the student discriminator includes a CNN discriminator and a parameter setting module.
3. A training method for a generative model based on "racecourse"-style multi-agent adversarial cooperation, characterized by: Includes the following steps: Step 1: Construct the initial models of the "teacher group" generator and student discriminator as described in claim 1, and initialize them; Step 2: Based on the periodic elite evaluation and bottom-ranking elimination mechanism, dynamically update the agents of the "teacher team" generator model to maintain the elite teacher team; Step 3: The student discriminator is guided and trained by a team of elite teachers. Under the joint guidance of the elite teachers, high-intensity adversarial training is carried out using a mixed dataset generated by real data and the optimal generator until the performance of the student discriminator reaches a stable state in the new evaluation period. Step 4: Continuous evolution cycle and system optimization, construct a new intelligent agent monitoring system, analyze the monitoring data in real time, dynamically adjust the evaluation frequency according to the group evolution speed, optimize the overall training strategy, and continuously cycle the training to keep the generative model of "racecourse" type multi-agent adversarial in the optimal state.
4. The training method for the generative model based on "racecourse"-style multi-agent adversarial cooperation according to claim 3, characterized in that: In step 2, the specific process of the periodic elite evaluation and bottom-ranking elimination mechanism is as follows: The evaluation and replacement process is executed at a preset cycle frequency. In each evaluation cycle, the performance of all discriminators in the teacher team is quantitatively scored using an auxiliary scorer. Based on the scoring results, the teacher team members are ranked, and a strict bottom-ranked elimination and replacement is implemented. The teacher discriminator with the lowest ranking is removed from the teacher team, and the resources it occupies are released. At the same time, a brand-new student discriminator with randomly initialized parameters is formally introduced into the teacher team as a new intelligent agent to replace the eliminated one, thus completing the dynamic update and maintenance of the overall capability of the teacher team within a cycle.
5. The training method for the generative model based on "racecourse"-style multi-agent adversarial cooperation according to claim 4, characterized in that: The initialization of the new intelligent agent adopts a differentiated strategy, prioritizing scarce architecture types to maintain group diversity. The parameter initialization adopts a specific distribution strategy to ensure that the new intelligent agent maintains its individuality while possessing development potential. The number of parameters of the new intelligent agent is maintained within a reasonable range of the average level of the teacher group.
6. The training method for the generative model based on "racecourse"-style multi-agent adversarial cooperation according to claim 3, characterized in that: In step 3, the specific process of providing guidance and training to students' discriminators through an elite team of teachers is as follows: After the agent of the "teacher team" generator model is updated, the training process of the next generation student discriminator is launched. First, the elite teachers of the entire teacher team will jointly serve as the guides. Under the joint guidance of the elite teacher team, the student discriminator will undergo high-intensity adversarial training using a mixed dataset generated by real data and the optimal generator. During training, each elite teacher team provides students with stable gradient signals, and the average value is taken as the student gradient. When the prediction of the student discriminator is inconsistent with the collective prior judgment of the teacher team, the gradient of the student model is corrected according to the teacher's output until its performance reaches a stable state in the new evaluation period.
7. The training method for the generative model based on "racecourse"-style multi-agent adversarial cooperation according to claim 3, characterized in that: Step 4 specifically includes: Step 4.1: Dynamically adjust the evaluation frequency according to the population evolution speed. When the population performance improves rapidly, shorten the evaluation cycle; when the population is stable, maintain the standard evaluation rhythm; when performance fluctuations occur, extend the evaluation cycle. Step 4.2: Dynamically adjust the elimination ratio based on the population diversity index, and set upper and lower thresholds; Step 4.3: Establish a diversity monitoring system to track architectural diversity, strategy diversity, and performance distribution uniformity, and analyze the monitoring data in real time; Step 4.4: Collect training performance data for each agent, regularly introduce new architecture types, eliminate outdated architectures, optimize the overall training strategy, and dynamically adjust evaluation metrics and weight allocation based on feedback information during the evolution process. Step 4.5: Establish a knowledge accumulation mechanism to systematically preserve successful training strategies and lessons learned.
8. The training method for the generative model based on "racecourse"-style multi-agent adversarial cooperation according to claim 7, characterized in that: The diversity monitoring system includes evolution efficiency monitoring and resource utilization efficiency monitoring. Evolution efficiency monitoring focuses on the growth rate of new intelligent agents and the overall improvement rate of the group, while resource utilization efficiency monitoring is used for the rational allocation and use of computing resources.
9. A generative system based on "racecourse"-style multi-agent adversarial cooperation, characterized in that: The system includes a processor that applies the generative model based on "racecourse" style multi-agent adversarial cooperation as described in claim 1 to construct an agent training environment with a competitive ranking mechanism.
10. The application of a generative model based on "racecourse"-style multi-agent adversarial cooperation, characterized by: The generative model is used in image generation, data augmentation, time series synthesis and prediction, natural language processing, and drug molecule design; among them, the task of time series data synthesis and prediction includes stocks, futures, and cryptocurrencies.