Enterprise cooperation and competition simulation method and system based on large language model

Abstract

The invention provides an enterprise cooperation and competition simulation method and system based on a large language model, and the method comprises the steps: 1, carrying out the modeling of a market environment, constructing a comprehensive model of supply and demand, consumer behaviors, market structure and policy intervention, and generating a dynamic and real market background; 2, performing multi-agent intervention, combining a large language model and reinforcement learning to realize dynamic strategy generation and natural language interaction of competitor agents, and adjusting pricing, advertisement and capacity layout; 3, dynamic game decision making, wherein non-cooperative, cooperative and mixed games are carried out between players and intelligent agents; 4, multi-agent debate and task disassembly are carried out, alliance negotiation and resource allocation are supported, task disassembly is realized through multi-round debate, and multi-party benefit balance is guaranteed; and 5, performing data analysis and feedback, recording market dynamics in real time, and providing visual analysis and strategy optimization suggestions. According to the invention, new technical support is provided for enterprise game simulation.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and in particular relates to a method and system for simulating enterprise cooperation and competition based on a large language model (LLM), which supports multi-agent task decomposition, multi-round debate, dynamic game, and market environment simulation. Background Art

[0002] The behaviors of enterprises in a dynamic market environment usually exhibit various complex forms of competition and cooperation. For example, enterprises may compete for market share through price wars, cooperate through resource sharing or technology alliances, or even form alliances in the fierce competition. However, traditional enterprise simulation systems are mostly based on static rules and preset logics, unable to reflect market dynamics and the true decision-making complexity among enterprises, which limits the flexibility, interactivity, and learning effect of the simulation.

[0003] In recent years, large language models (such as GPT-4, etc.) have made remarkable progress in natural language generation (NLG) and task planning. They can understand complex scenarios, generate dynamic strategies, and conduct multi-round conversations, thus providing a new technical basis for enterprise game simulation. At the same time, the combination of reinforcement learning (RL) and multi-agent system (MAS) enables agents to adaptively optimize strategies in complex environments.

[0004] Based on this, the present invention provides a brand-new method and system for simulating enterprise cooperation and competition, which realizes in-depth simulation of enterprise behaviors and enhanced interactivity through LLM-multi-agent-game engine-task decomposition-market dynamics modeling. Summary of the Invention

[0005] To solve the problems in the background art, the present invention provides a method and system for simulating enterprise cooperation and competition based on a large language model.

[0006] The system for simulating enterprise cooperation and competition based on a large language model is divided into six major modules. The functional implementation of each module is closely coupled, covering the entire process from market environment modeling, agent role design, decision-making and game engine to multi-round debate, task decomposition, and data feedback.

[0007] The present invention also provides a method for simulating enterprise cooperation and competition based on a large language model, including: Step 1. Implement market environment simulation, and provide a dynamic and realistic market background by comprehensively modeling supply and demand relationships, consumer behaviors, market structures, and policy interventions.

[0008] Step 2. Multi-agent intervention, by combining large language model and reinforcement learning technologies, realize dynamic strategy generation and natural language interaction, and be used for competitor agents to adjust pricing, advertising, and production capacity layout strategies in real time based on reinforcement learning to compete with players for market share.

[0009] Step 3. Decision-making and Game-playing, which is used to conduct dynamic games between players and agents, covering three modes: non-cooperative games, cooperative games, and mixed games.

[0010] Step 4. Multi-agent Debate and Task Decomposition, which is aimed at complex scenarios such as coalition negotiation and resource allocation, and is designed to support multi-round debates and collaborative task decomposition between agents and players, so as to achieve the balance and efficient execution of multi-party interests.

[0011] Step 5. Data Analysis and Feedback. The data analysis and feedback provide visual analysis and strategy optimization suggestions by recording the market dynamics of each round in real time.

[0012] Furthermore, Step 1 specifically includes: On the demand side, a price-demand function is adopted , where a represents the potential maximum demand, b is the price elasticity coefficient, and P is the price; on the supply side, it is characterized by the cost function C(Q), modeling the fixed cost and variable cost respectively. Consumer behavior is based on the discrete choice model, and through the utility function simulates consumers' preferences for price, quality, and brand factors. The higher the U, the higher the probability of consumers purchasing the product; Furthermore, Step 1 also includes: To make the simulation process closer to the real market, key parameters of the demand curve elasticity and the total demand quantity will be dynamically adjusted according to price fluctuations and the decisions of players or agents. For different market structures, it supports multiple scenarios. In a perfectly competitive scenario, a large number of small enterprises each set their own prices and jointly determine the market price; Furthermore, Step 1 also includes: In an oligopoly, a few enterprises dominate the market share and often form a price leadership mechanism; in a monopolistic competition state, enterprises compete in terms of price and quality through product differentiation.

[0013] Furthermore, Step 1 also includes: To prevent market failures caused by excessive concentration, when the market concentration exceeds 2500, anti-monopoly supervision will be automatically triggered; at the same time, the government agent dynamically adjusts the tax rate or issues subsidies according to the market performance to affect the production capacity and pricing strategy of enterprises. After each round combination, various indicators such as the current price level, market share evolution, and the behaviors of players and agents will be recorded, and a real-time feedback report will be generated to provide data support for subsequent decision-making.

[0014] Furthermore, Step 2 specifically includes: The partner agent proposes a reasonable income distribution or task division plan through coalition negotiation and resource sharing; the consumer agent selects products based on price and intelligent factors, thus affecting the sales performance and market trend of enterprises; the regulatory agent is responsible for implementing anti-monopoly policies or tax interventions, directly acting on market rules.

[0015] Furthermore, step 2 also includes: To address complex decision-making scenarios, a task decomposition method based on LLM is adopted to break down the macro task of "market competition" into several subtasks such as market demand and competition situation analysis, price or advertising budget optimization, and prediction of market share and profit at the first level, and gradually complete the strategy generation; During the multi-round interaction process, the agent outputs natural language decision-making suggestions with the help of LLM; Subsequently, the reinforcement learning mechanism further optimizes this strategy: When it is Q-learning, the value function of the competitor agent for state s and action a is iteratively updated according to the following formula: where a is the learning rate, is the discount factor, r is the reward value, s’ represents the subsequent state transferred to after performing action a in state s, and a’ is the action that maximizes in state s’. Then, the optimal strategy action is screened through multiple iterations and corresponding responses are made to external regulatory interventions, so as to achieve a high-fidelity simulation of multi-role decision-making and interaction in complex market games.

[0016] Furthermore, step 3 also includes: In a non-cooperative game model, each participant determines the optimal strategy through Nash equilibrium. In the scenario of a price war, the enterprise selects the price to maximize its own profit function , the profit function of enterprise i depends on the quotation combination of all enterprises including its own price , represents the price or quotation selected by enterprise i, n is the number of enterprises participating in the game in the market, and at the same time, it anticipates the reaction of competitors to its own decisions, so as to reach Nash equilibrium when it is optimal for all enterprises, that is, satisfy . Under the condition of Nash equilibrium, enterprise i cannot obtain higher profits by unilaterally adjusting ; that is, when the quotations of other enterprises are fixed, the partial derivative of the profit function of enterprise i with respect to is zero.

[0017] For the cooperative game model, the Shapley value or the nucleolus method is used to allocate the coalition payoff. In the task of resource sharing, if the value function of the coalition is , then the Shapley value for player i is expressed as: where N represents the set of all players participating in the game, S is an arbitrary subset that does not include the player, |S| is the number of elements in subset S, |N| is the total number of all players, and v(S) is the value or payoff that subset S can obtain, is the Shapley value of the player in the alliance, which is used to measure their "fair" income in cooperation. Thus, a fair distribution is made according to the marginal contribution of each enterprise to the overall income of the alliance.

[0018] In the mixed game model, the system dynamically switches between cooperation and competition: when the alliance negotiation fails, each enterprise will immediately enter the cooperation state and start a price war.

[0019] Furthermore, step 4 also includes: the agent decomposes the complex task into several subtasks according to the preset game goal, proposes an initial allocation plan in the alliance negotiation, then analyzes the advantages and disadvantages of the player's proposal, and finally continuously adjusts the income distribution using cooperative game theory.

[0020] In the multi-round debate process, the agent generates natural language descriptions through the large language model to refute the player's proposal. At the same time, to evaluate the feasibility of the negotiation result, the income and acceptance probability of each party's proposal are calculated by means of the game in step 3. If the utility function of player i is , it represents the utility obtained by player i under the current proposal , measuring the benefit level obtained by it from this proposal. When the reserved utility is , then when , player i tends to accept this proposal: further adopt the acceptance probability in the form of Logits regression: , where is the sensitivity constant. Through the dynamic calculation and feedback of multi-round proposals and debates, it can help multiple parties gradually approach the optimal or satisfactory solution in complex cooperative or adversarial scenarios, and finally complete task decomposition and game negotiation.

[0021] Furthermore, step 5 includes: at the end of each game or simulation, the core index profit of each enterprise or player will be collected and stored , which is described by the following formula: where is the product price of the i-th enterprise at time t, is its sales volume, is the corresponding cost function; Subsequently, the factual data will be visually displayed in the form of dynamic curves or bar charts to help players and agents more intuitively understand the market trend and competition pattern.

[0022] On this basis, combining historical data and game results, generate strategy suggestions for players: by tracing back the profit and market share changes in the previous few rounds, based on multi-objective optimization or regression analysis, calculate a comprehensive score function , which is used to evaluate the potential income and risk of different strategy combinations. If If the corresponding strategy combination is better than the current plan, suggestions such as "reduce the price by 10%" or "increase the advertising budget" will be output to the player to achieve more efficient competition and resource allocation in subsequent rounds. The present invention also provides an enterprise cooperation and competition simulation system based on a large language model, which is implemented by the above-mentioned enterprise cooperation and competition simulation method based on a large language model, including: A market environment simulation module that simulates the dynamic market environment in the enterprise competition and cooperation scenario, including consumer behavior, supply and demand relationship, market structure and policy intervention. Construct market parameters and various market structures that support dynamic adjustment.

[0023] A multi-agent module that, based on a large language model, the agents simulate different roles in the market. Support task decomposition, dynamic strategy generation, natural language interaction and reinforcement learning optimization.

[0024] A decision-making and game engine: Provide dynamic game support, including non-cooperative games, cooperative games and mixed games (dynamically switching between cooperation and competition). Use game theory to optimize the strategies of agents and players.

[0025] A multi-agent debate and task decomposition module: Support multi-round debates between agents and players, collaborative task decomposition and dynamic negotiation. Provide misaligned debate, equal debate and round-table debate modes, and generate complex decision-making logic in combination with the task decomposition framework.

[0026] A data analysis and feedback module: Record market data in real time, generate visual feedback, and provide suggestions for strategy improvement.

[0027] A learning and optimization module that, based on reinforcement learning and evolutionary game theory, optimizes the long-term behavior and strategies of agents. Players learn decision-making improvement through feedback. Agents adjust their behavior through historical game data.

[0028] Advantages of the present invention: 1. More realistic market environment: By combining supply and demand relationships, policy interventions, and market structures, it supports dynamic adjustment of market parameters, enhancing the realism of simulations. 2. More optimized agent decisions: By integrating LLM and reinforcement learning, agents can dynamically generate strategies and interact through natural language, improving the intelligence level of decision-making. 3. More comprehensive game strategies: It supports non-cooperative games, cooperative games, and mixed games, and can simulate real commercial competition and cooperation scenarios. 4. More refined task decomposition: Through multi-agent debate and task decomposition, negotiation efficiency is improved, making strategy execution more hierarchical and operable. 5. More accurate data feedback: By adopting real-time data visualization and historical analysis, precise strategy optimization suggestions are provided to players, enhancing competitive advantages. 6. Stronger long-term learning ability: By combining reinforcement learning and evolutionary games, agent behavior is optimized, enabling continuous improvement of strategy levels in multiple rounds of competition. The present invention can be widely applied to fields such as enterprise strategy simulation, market competition research, policy-making simulation, and business negotiation training. It can not only help enterprises optimize their business strategies, but also be used in teaching and experimental research in disciplines such as economics and management, greatly enhancing the intelligence level and decision-making accuracy of market simulations. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the system flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be described clearly and completely below. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0031] As Figure 1 shown, the enterprise cooperation and competition simulation system based on the large language model of the present invention is divided into six major modules. The functional implementation of each module is closely coupled, covering the entire process from market environment modeling, agent role design, decision-making and game engines to multi-round debate, task decomposition, and data feedback.

[0032] The present invention also provides a method for simulating enterprise cooperation and competition based on the large language model, including: Step 1. Implement market environment simulation by comprehensively modeling supply and demand relationships, consumer behavior, market structure, and policy interventions to provide a dynamic and real market background.

[0033] The demand side adopts a price-demand function , where a represents the potential maximum demand, b is the price elasticity coefficient, and P is the price; the supply side is characterized by the cost function C(Q), modeling fixed costs and variable costs separately. Consumer behavior is based on the Discrete Choice Model, and through the utility function simulates consumers' preferences for price, quality, and brand factors. The higher the value of U, the higher the probability that consumers will purchase the product. To make the simulation process closer to the real market, key parameters such as the elasticity of the demand curve and the total demand will be dynamically adjusted according to price fluctuations and the decisions of players or agents. For different market structures, multiple scenarios are supported; in a perfectly competitive scenario, numerous small enterprises set their own prices and jointly determine the market price; in an oligopoly, a few enterprises dominate the market share and often form a price leadership mechanism; in a monopolistic competition state, enterprises can compete in dimensions such as price and quality through product differentiation.

[0034] To prevent market failures caused by excessive concentration, when the market concentration (measured by the Herfindahl-Hirschman Index HHI) exceeds 2500, antitrust supervision will be automatically triggered, such as splitting or fining; at the same time, the government agent can dynamically adjust tax rates or provide subsidies according to market performance to affect the production capacity and pricing strategies of enterprises. After each round combination, i will record various indicators such as the current price level, the evolution of market share, the behavior of players and agents, and generate a real-time feedback report to provide data support for subsequent decisions.

[0035] Step 2. Multi-agent intervention, through the combination of large language models (LLMs) and reinforcement learning techniques, realizes dynamic policy generation and natural language interaction. The competitor agent adjusts strategies such as pricing, advertising, and production capacity layout in real time based on reinforcement learning to compete with players for market share; the partner agent proposes reasonable revenue distribution or task division plans through alliance negotiation and resource sharing; the consumer agent selects products based on factors such as price and intelligence, thus affecting the sales performance and market trend of enterprises; the regulatory agent is responsible for implementing antitrust policies or tax interventions, directly acting on market rules.

[0036] To address complex decision-making scenarios, this step adopts a task decomposition method based on LLM, breaking down macro-tasks such as "market competition" into several sub-tasks, including market demand and competition situation analysis, price or advertising budget optimization, and prediction of market share and profit at the first level, and gradually completing strategy generation. During the multi-round interaction process, the agent uses LLM to output natural language decision-making suggestions. For example, in the scenario of a price war, it may give a strategy description like "According to the current market demand prediction, reducing the price by 10% will effectively increase the market share but will affect the profit margin". Subsequently, the reinforcement learning mechanism (such as DQN or PPO) further optimizes this strategy: taking Q-learning as an example, the value function of the competitor agent for state s and action a is iteratively updated according to the following formula: where a is the learning rate, is the discount factor, r is the reward value, s’ represents the subsequent state transferred after performing action a in state s, and a’ is the action that maximizes in state s’. In this way, the optimal strategy actions (such as pricing or advertising strategies) can be screened through multiple iterations and corresponding responses can be made to external regulatory interventions (such as taxation or merger restrictions), thus achieving a high-fidelity simulation of multi-role decision-making and interaction in complex market games.

[0037] Step 3. Decision-making and game-playing. Decision-making and game-playing are the core of the entire system, used for dynamic game-playing between players and agents, covering three modes: non-cooperative games, cooperative games, and mixed games.

[0038] In the non-cooperative game model, each participant determines the optimal strategy through Nash equilibrium. For example, in the scenario of a price war, an enterprise chooses a price to maximize its own profit function , the profit function of enterprise i, which depends on the combination of quotes of all enterprises including its own price , represents the price or quote chosen by enterprise i, n is the number of enterprises participating in the game in the market, and at the same time anticipates the reaction of competitors to its own decision, so as to reach Nash equilibrium when is optimal for all enterprises, that is, satisfying . Under the Nash equilibrium condition, enterprise i cannot obtain higher profits by unilaterally adjusting ; that is, when the quotes of other enterprises are fixed, the partial derivative of the profit function of enterprise i with respect to is zero.

[0039] For the cooperative game model, the Shapley value or the nucleolus method is used to allocate the coalition payoff. For example, in the task of resource sharing, if the value function of the coalition is , then the Shapley value for player i is expressed as: Among them, \(N\) represents the set of all players participating in the game, \(S\) is an arbitrary subset that does not contain players, \(|S|\) is the number of elements in subset \(S\), \(|N|\) is the total number of all players, and \(v(S)\) is the value or benefit that subset \(S\) can obtain. is the Shapley value of the player in the coalition, which is used to measure its "fair" benefit in cooperation.

[0040] Thus, a fair distribution is carried out according to the marginal contribution of each enterprise to the overall benefit of the coalition. In the mixed game model, the system can dynamically switch between cooperation and competition: when the coalition negotiation fails, each enterprise will immediately enter the cooperation state and start a price war.

[0041] Since this step 3 is linked in real time with the market environment simulation in step 1, the costs and benefits of each action will be synchronously calculated and updated, thus more realistically reflecting market changes and providing accurate benefit references for the game decisions of players and agents.

[0042] Step 4. Multi-agent Debate and Task Decomposition Multi-agent and task decomposition mainly targets complex scenarios such as coalition negotiation and resource allocation, aiming to support multi-round debates and collaborative task decomposition between agents and players, and achieve the balance and efficient execution of multi-party interests. In terms of technical implementation, first, the agent decomposes the complex task into several subtasks according to the preset game goal. For example, in coalition negotiation, it first proposes an initial allocation plan, then analyzes the advantages and disadvantages of the player's proposal, and finally continuously adjusts the benefit distribution using cooperative game theory (such as the Shapley value or the nucleolus method). During the multi-round debate process, the agent generates natural language descriptions through a large language model (LLM). For example, in response to the player's proposal, it refutes: "Your plan ignores our marginal contribution and needs to redistribute the benefit ratio." At the same time, to evaluate the feasibility of the negotiation result, this step uses the game in step 3 to calculate the benefits and acceptance probabilities of each party's proposal. If the utility function of player \(i\) is and the reservation utility is , then when , player \(i\) tends to accept this proposal: further adopt the acceptance probability in the form of Logits regression: , where is the sensitivity constant. Through the dynamic calculation and feedback of multi-round proposals and debates, it can help multiple parties gradually approach the optimal or satisfactory solution in complex cooperative or adversarial scenarios, and finally complete task decomposition and game negotiation.

[0043] Step 5. Data Analysis and Feedback. Data analysis and feedback provide visual analysis and strategy optimization suggestions by recording the market dynamics (such as price, output, profit, and market share) of each round in real time.

[0044] Specifically, at the end of each game or simulation, the core metrics of each enterprise or player, such as profit, are collected and stored. , which is described by the following formula: Where is the product price of the i-th enterprise at time t, is its sales volume, is the corresponding cost function. Subsequently, these factual data are visually presented in the form of dynamic curves or bar charts, etc., to help players and agents more intuitively understand the market trends and competition pattern. On this basis, combining historical data with game results, strategic suggestions are generated for players: for example, by looking back at the changes in profit and market share in the previous few rounds, and based on multi-objective optimization or regression analysis, a comprehensive scoring function is calculated to evaluate the potential benefits and risks of different strategy combinations. If the corresponding strategy combination is better than the current plan, specific suggestions such as "reduce the price by 10%" or "increase the advertising budget" are output to the player, so as to achieve more efficient competition and resource allocation in subsequent rounds. Through continuous iterative data recording, visualization, and decision-making suggestions, it can help players and agents continuously adjust and improve their respective strategies, so as to maintain a competitive advantage in a complex market environment.

[0045] In addition, the present invention also provides an enterprise cooperation and competition simulation system based on a large language model, which is implemented by the above-mentioned enterprise cooperation and competition simulation method based on a large language model, including: A market environment simulation module that simulates the dynamic market environment in enterprise competition and cooperation scenarios, including consumer behavior, supply and demand relationships, market structure, and policy intervention. Construct market parameters (price elasticity, total demand) that support dynamic adjustment and various market structures (such as perfect competition, oligopoly).

[0046] A multi-agent module (LLM Agents) that, based on a large language model (LLM), simulates different roles in the market (such as competitors, partners, consumers, governments). Supports task decomposition, dynamic strategy generation, natural language interaction, and reinforcement learning optimization.

[0047] A decision-making and game engine Provides dynamic game support, including non-cooperative games (such as price wars), cooperative games (such as resource sharing), and mixed games (dynamic switching between cooperation and competition). Uses game theory (such as Nash equilibrium and Shapley value allocation) to optimize the strategies of agents and players.

[0048] A multi-agent debate and task decomposition module Supports multi-round debates between agents and players, collaborative task decomposition, and dynamic negotiation. Provides misaligned debate, equal debate, and round-table debate modes, and generates complex decision-making logic in combination with the task decomposition framework.

[0049] Data Analysis and Feedback Module Record market data (such as price, theory, market share) in real time, generate visual feedback, and provide suggestions for strategy improvement.

[0050] Learning and Optimization Module, based on reinforcement learning (RL) and evolutionary game theory, optimizes the long-term behavior and strategies of agents. Players learn to improve decisions through feedback. Agents adjust their behavior based on historical game data. Embodiment

[0051] Suppose there are 5 enterprises participating in market competition and cooperation.

[0052] This embodiment is based on the operation process of an enterprise cooperation and competition simulation system driven by a large language model. Suppose there are 5 enterprises in the market (1 player enterprise and 4 agent enterprises), and there are also government supervision agents, consumer agents and market environment simulation. The case covers dynamic processes such as market competition, alliance cooperation, resource allocation, negotiation and dissolution, and price war.

[0053] Market Environment: Industry: New energy vehicle market Market characteristics: Annual market demand is 1 million vehicles Product type: Mid-range electric vehicles (price range: $20,000 - $35,000) Competition characteristics: High price elasticity: Consumers are more sensitive to price.

[0054] Technology is the key point: Battery technology determines vehicle performance.

[0055] Government policies: The government provides subsidies, but strictly supervises market monopoly behavior.

[0056] Participating enterprises Player enterprise (E1): A mid-sized new energy vehicle enterprise, focusing on the mid-range market, with certain R & D capabilities and a market share of 20%.

[0057] Agent enterprises: E2 (Agent enterprise A): The market leader, with a 30% share, having strong brand and technological advantages.

[0058] E3 (Agent enterprise B): A medium-sized enterprise, with a 20% market share, focusing on low-cost production.

[0059] E4 (Agent enterprise C): A new emerging enterprise, with a 10% market share, having strong technological innovation capabilities.

[0060] E5 (Agent Enterprise D): Medium-sized enterprise, with a 20% market share, sound financial condition, and expertise in the supply chain.

[0061] Management: Government Agent: Responsible for supervising the market, preventing monopolistic behavior, and providing subsidies.

[0062] Consumer Agent: Simulates consumer behavior, including sensitivity to price, quality, brand, and technology.

[0063] Process and Scenario Phase 1: Market Competition (Initial Price War) Scenario Description: After the system initialization, the market is in a fierce competition state. The player enterprise (E1) needs to formulate strategies on price, quality, and advertising budget to compete for the market share with the other 4 enterprises. The Consumer Agent selects products based on price, quality, and brand preferences.

[0064] Simulation Process Market Initialization Data: Annual Market Demand: 1 million vehicles Initial Prices of Each Enterprise: E1 (Player Enterprise): $25,000 (mid-range pricing) E2 (Agent A): $28,000 (high-end pricing) E3: (Agent B): $22,000 (low-end pricing) E4 (Agent C): $26,000 (innovative product, slightly higher than the mid-range market) E5 (Agent D): $24,000 (stable mid-range pricing) Player Decision The player chooses to lower the price by 5% (to $23,750) and increase the advertising budget by 20% to attract more consumers.

[0065] Agent Strategies: E2 (Agent A): Maintains high-end pricing, focuses on brand premium, does not lower the price, but increases technology promotion.

[0066] E3 (Agent B): Lowers the price to $21,000, trying to capture the market share through extremely low prices.

[0067] E4 (Agent C): Maintains the price, but attracts consumers' interest by releasing new battery technology.

[0068] E5 (Agent D): Slightly lowers the price by 2% and optimizes the advertising placement.

[0069] Market Feedback (First Round): The consumer agent redistributes demand based on price and brand weights: E1 (Player Enterprise): Market share increases from 20% to 25%.

[0070] E2 (Agent A): Market share decreases from 30% to 28%.

[0071] E3 (Agent B): Market share increases from 20% to 22%.

[0072] E4 (Agent C): Market share remains at 10%.

[0073] E5 (Agent D): Market share decreases from 20% to 15%.

[0074] Result Analysis The player enterprise effectively increases its market share through price cuts and advertising investment, but the profit margin decreases.

[0075] The market share of the high-end brand (E2) slightly declines but still remains leading.

[0076] The low-end enterprise (E3) expands its market through price war but has a greater impact on profit compression.

[0077] Phase 2: Alliance Cooperation (Forming a R & D Alliance) Scenario Description Market competition leads to increased technological pressure, and enterprises start to consider cooperating in R & D of new energy battery technology. The player enterprise (E1) initiates alliance negotiations with E4 (Agent C) and E5 (Agent D), attempting to reduce R & D costs and share technological achievements through cooperation.

[0078] Simulation Process Player's Proposal The player (E1) proposes to form a R & D alliance to jointly develop a new generation of battery technology and suggests distributing technological achievements according to the proportion of R & D investment.

[0079] Agent's Proposal: E4 (Agent C): Agrees to the alliance but proposes to increase cooperation transparency and requests more patent usage rights.

[0080] E5 (Agent D): Suggests distributing technological achievements according to market share rather than the proportion of R & D investment.

[0081] Multiple Rounds of Debates and Task Disassembly (Supported by LLM): Player (E1): Generates a rebuttal debate through the large language model: "If distributed according to market share, it will lead to unfairness because we bear the main R & D costs." E4 (Agent C): Proposes a compromise solution, suggesting that the R & D cost distribution be 60% (E1), 30% (E4), and 10% (E5), but the technology sharing should be implemented according to the equal principle.

[0082] E5 (Agent D): Accepts the compromise solution but requests that the alliance term be 3 years.

[0083] Alliance cost: Finally, the alliance reaches an agreement according to E4's plan, and the three enterprises sign a cooperation agreement.

[0084] Result analysis The alliance reduces the R & D cost and allows the three enterprises to share new technologies, enhancing the overall competitiveness. E2 and E3 that do not participate in the alliance need to face the pressure of technological competition and may be forced to adjust their strategies.

[0085] Stage 3: Alliance dissolution and intensified competition Scenario description After two years of operation, due to uneven distribution of interests and differences in market strategies, E4 proposes to dissolve the alliance. The player enterprise (E1) needs to decide whether to continue to maintain the alliance or withdraw and enter direct competition with E4 and E5.

[0086] Simulation process Alliance dissolution negotiation: Player (E1): Hopes to continue to maintain the alliance but requests to re-adjust the technology distribution rights.

[0087] E4 (Agent C): Proposes to withdraw from the alliance and at the same time requests to restrict the patent usage rights of E1.

[0088] E5 (Agent D): Neutral, inclined to maintain the alliance but unwilling to invest more resources.

[0089] Negotiation breakdown: Finally, E4 insists on withdrawing from the alliance, and the alliance is dissolved.

[0090] Intensified competition: The player enterprise (E1) quickly adjusts its strategy, uses the existing technology to reduce the product cost, and seizes the market share of E4.

[0091] E4 focuses on high-end technology and tries to attract consumers through technological leadership.

[0092] E5 maintains a steady strategy and keeps the cost advantage of the supply chain.

[0093] Result analysis The player enterprise seizes a part of the market share by reducing prices in the short term, but needs to face technological pressure in the long term. E4 gradually stabilizes in the high-end market and forms competition with E2 (the high-end leader).

[0094] Phase 4: Government Intervention (Antitrust Investigation) Scenario Description Since the player enterprises (E1) and E2 occupy more than 60% of the market share in the market, the government agent initiates an antitrust investigation and may take measures (such as restricting mergers and acquisitions or splitting enterprises).

[0095] Simulation Process Actions of the government agent: Warn the player enterprise (E1), require it to disclose the alliance technology use agreement, and restrict the growth of its market share.

[0096] Player's coping strategies: Propose a cooperation plan with small enterprises (non-agents) to avoid more stringent supervision.

[0097] Summary and Analysis: This embodiment demonstrates the competition, cooperation and government intervention scenarios of 5 enterprises in a dynamic market environment, specifically embodying the following technical characteristics: Market dynamic modeling: The complex interaction of supply and demand relationship, price elasticity, consumer behavior and policy intervention. Multi-agent task decomposition and debate: The agent generates multi-round negotiation and strategy logic through LLM. Dynamic game and alliance cooperation: Simulate the dynamic decision-making switch of enterprises in competition and cooperation. Government intervention model: Introduce policy constraints through the antitrust mechanism to enhance the authenticity of the simulation.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for simulating enterprise cooperation and competition based on large language models, characterized in that, Including: Step 1. Implement market environment simulation. By comprehensively modeling supply and demand relationships, consumer behavior, market structure, and policy intervention, a dynamic and realistic market background is provided. Step 2. Multi-agent intervention. By combining large language models and reinforcement learning techniques, dynamic policy generation and natural language interaction are achieved. This is used for competitor agents to adjust pricing, advertising, and production capacity layout strategies in real time based on reinforcement learning to compete with players for market share. Step 3. Decision-making and gaming, which is used for dynamic gaming between players and agents, covering three modes: non-cooperative gaming, cooperative gaming, and mixed gaming. Step 4. Multi-agent debate and task decomposition. For complex scenarios such as alliance negotiation and resource allocation, it aims to support multi-round debates and collaborative task decomposition between agents and players to achieve the balance and efficient execution of multi-party interests. Step 5. Data analysis and feedback. Data analysis and feedback record the market dynamics of each round in real time and provide visual analysis and strategic optimization suggestions.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: on the demand side, a price-demand function is adopted , where a represents the potential maximum demand, b is the price elasticity coefficient, and P is the price; on the supply side, it is characterized by the cost function C(Q), and the fixed cost and variable cost are modeled respectively. Consumer behavior is based on the discrete choice model, and through the utility function simulates consumers' preferences for price, quality, and brand factors. The higher the U, the higher the probability of consumers purchasing the product; to be closer to the real market, the key parameters of the demand curve elasticity and the total demand volume will be dynamically adjusted according to price fluctuations and the decisions of players or agents. For different market structures, multiple scenarios are supported. In the case of perfect competition, numerous small enterprises each set their own prices and jointly determine the market price.

3. The method according to claim 1, characterized in that, Step 1 also includes: In oligopoly, a few enterprises dominate the market share and often form a price leadership mechanism. In monopolistic competition, enterprises compete in terms of price and quality through product differentiation.

4. The method according to claim 1, wherein Step 1 also includes: In oligopoly, a few enterprises dominate the market share and often form a price leadership mechanism. In monopolistic competition, enterprises compete in terms of price and quality through product differentiation.

5. The method according to claim 1, characterized in that Step 1 also includes: In oligopoly, a few enterprises dominate the market share and often form a price leadership mechanism. In monopolistic competition, enterprises compete in terms of price and quality through product differentiation.

6. The method according to claim 1, characterized in that Step 2 also includes: To solve complex decision-making scenarios, a task decomposition method based on LLM is adopted. The macro task of "market competition" is broken down into several subtasks such as market demand and competition situation analysis, price or advertising budget optimization, and prediction of market share and profit, and the strategy generation is gradually completed. During the multi-round interaction process, the agent outputs natural language decision-making suggestions with the help of LLM. Subsequently, the reinforcement learning mechanism further optimizes this policy: when it is Q-learning, the value function of the competitor agent for state s and action a is iteratively updated according to the following formula: where a is the learning rate, is the discount factor, r is the reward value, s’ represents the subsequent state transferred to after executing action a in state s, and a’ is the action that maximizes in multiple iterations, the optimal policy action is screened and corresponding responses are made to external regulatory interventions, so as to achieve a high-fidelity simulation of multi-role decision-making and interaction in complex market games.

7. The method according to claim 1, characterized in that, Step 3 also includes: In the non - cooperative game model, each participant determines the optimal strategy through Nash equilibrium. In the price war scenario, the enterprise selects the price to maximize its own profit function while anticipating the reaction of competitors to its own decisions, thus reaching Nash equilibrium when it is optimal for all enterprises, that is, satisfying .; For the cooperative game model, the Shapley value or the nucleolus method is used to allocate the coalition payoff. In the resource sharing task, if the value function of the coalition is , then the Shapley value of the player is expressed as: where N represents the set of all players participating in the game, S is an arbitrary subset that does not include the player, |S| is the number of elements in the subset S, |N| is the total number of all players, v(S) is the value or payoff that the subset S can obtain, is the Shapley value of the player in the coalition; thus, a fair distribution is made according to the marginal contribution of each enterprise to the overall coalition payoff; In the mixed game model, the system dynamically switches between cooperation and competition. When the coalition negotiation fails, each enterprise will immediately enter the cooperation state and start a price war.

8. The method according to claim 1, characterized in that, Step 4 also includes: The agent decomposes complex tasks into several subtasks according to the preset gaming goal, first proposes an initial allocation plan in alliance negotiation, then analyzes the advantages and disadvantages of the player's proposal, and finally continuously adjusts the income distribution using cooperative gaming theory. During the multi-round debate process, the agent generates natural language descriptions through a large language model to refute the player's proposal. At the same time, to evaluate the feasibility of the negotiation result, the payoff and acceptance probability of each party's proposal are calculated through the game in step 3. If the player's utility function is and the reservation utility is , then when , the player tends to accept the proposal: further adopt the acceptance probability in the form of Logits regression: , where is the sensitivity constant. Through the dynamic calculation and feedback of multi-round proposals and debates, it can help multiple parties gradually approach the optimal or satisfactory solution in complex cooperative or adversarial scenarios, and finally complete task decomposition and game negotiation.

9. The method according to claim 1, characterized in that, Step 5 includes: at the end of each game or simulation, the core metric profit of each enterprise or player is collected and stored , described by the following formula: where is the product price of the i-th enterprise at time t, is its sales volume, is the corresponding cost function. Subsequently, the factual data is visualized in the form of dynamic curves or bar charts to help players and agents more intuitively understand the market trends and competition patterns; On this basis, by combining historical data with game results, generate strategic suggestions for players: By tracing back the changes in profit and market share in the previous few rounds, and based on multi-objective optimization or regression analysis, calculate a comprehensive scoring function , which is used to evaluate the potential benefits and risks of different strategy combinations. If the corresponding strategy combination is better than the current plan, then output suggestions such as "reduce price by 10%" or "increase advertising budget" to the player, so as to achieve more efficient competition and resource allocation in subsequent rounds.

10. An enterprise cooperation and competition simulation system based on a large language model, which is implemented by the enterprise cooperation and competition simulation method based on a large language model according to any one of claims 1-10, characterized in that, Including: Market environment simulation module, which simulates the dynamic market environment in enterprise competition and cooperation scenarios, including consumer behavior, supply and demand relationships, market structure, and policy intervention, and constructs market parameters and various market structures that support dynamic adjustment. Multi-agent module, based on large language models, agents simulate different roles in the market, supporting task decomposition, dynamic policy generation, natural language interaction, and reinforcement learning optimization. Decision-making and gaming engine, which provides dynamic gaming support, including non-cooperative gaming, cooperative gaming, and mixed gaming, and uses gaming theory to optimize the strategies of agents and players. Multi-agent debate and task decomposition module, which supports multi-round debates between agents and players, collaborative task decomposition, and dynamic negotiation, provides misaligned debate, equal debate, and round-table debate modes, and generates complex decision-making logic in combination with the task decomposition framework. Data analysis and feedback module, which records market data in real time, generates visual feedback, and provides suggestions for strategy improvement; Learning and optimization module, which optimizes the long-term behavior and strategies of the agent based on reinforcement learning and evolutionary game theory. Players learn to improve their decisions through feedback, and the agent adjusts its behavior based on historical game data.

Images