Intelligent Strategy Deduction Method and Device for Machine Self-Game Theory

Abstract

The embodiments of this disclosure provide an intelligent strategy deduction method and apparatus for machine self-games. The method includes: constructing a game event agent and extracting deduction elements of the event to be deduced based on the game event agent; constructing a decision preference matrix with added interference coefficients; constructing an intelligent agent self-game model based on the deduction elements and the decision preference matrix, and generating multiple sets of game scenarios and game paths based on the intelligent agent self-game model, wherein the intelligent agent self-game model includes a large model based on the Transformer architecture and a small model based on a deep Q-network; and constructing a strategic evaluation model, and performing risk assessment and trend prediction on multiple sets of game scenarios and game paths based on the strategic evaluation model, and generating a deduction report.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of artificial intelligence technology, and more specifically, to a method and apparatus for intelligent strategy deduction in machine self-game. Background Technology

[0002] Strategic simulation is a crucial tool for supporting corporate development planning and strategic decision-making. By simulating and analyzing complex scenarios, it helps decision-makers evaluate and optimize strategic choices when facing uncertainty and complex situations. For example, in the business world, strategic simulation helps organizations improve the quality of their decisions, enhance their ability to respond to unforeseen circumstances, and continuously optimize strategic solutions in practice.

[0003] The shortcomings of traditional strategic simulation models are mainly reflected in the time-consuming preparation, high professional threshold, reliance on expert wisdom, and difficulty in guaranteeing the effectiveness of the results. Faced with these challenges, finding and implementing innovative simulation methods to improve simulation efficiency, reduce over-reliance on experts, and enhance simulation effectiveness has become an urgent problem for decision-makers and simulation organizers.

[0004] Currently, artificial intelligence has been integrated into strategic deduction design. For example, Large Language Models (LLMs) are being widely applied in multi-agent systems. LLMs can understand and generate natural language text, making communication between agents more natural and efficient. This capability is crucial for strategy formulation and execution in complex game environments. Through pre-training on massive amounts of data, LLMs possess extensive knowledge and reasoning capabilities, enabling them to quickly learn and generate effective strategies during game play. Furthermore, LLMs can interactively learn and adjust strategies in real time to respond to changes in the opponent's behavior. Therefore, the application of artificial intelligence technology provides new directions and approaches for improving the efficiency and optimizing the process of strategic deduction. Summary of the Invention

[0005] The embodiments described herein provide a method, apparatus, and computer-readable storage medium storing a computer program for intelligent strategy deduction in machine self-play.

[0006] According to a first aspect of this disclosure, an intelligent strategy deduction method for machine self-game is provided, comprising: constructing a game event agent and extracting deduction elements of the event to be deduced based on the game event agent; constructing a decision preference matrix with added interference coefficients; constructing an intelligent agent self-game model based on the deduction elements and the decision preference matrix, and generating multiple sets of game scenarios and game paths based on the intelligent agent self-game model, wherein the intelligent agent self-game model includes a large model based on the Transformer architecture and a small model based on a deep Q-network; and constructing a strategic evaluation model, and performing risk assessment and trend prediction on multiple sets of game scenarios and game paths based on the strategic evaluation model, and generating a deduction report.

[0007] In some embodiments of this disclosure, constructing a game event agent includes: collecting raw text data from multiple data sources; cleaning and preprocessing the collected raw text data; labeling key elements of the preprocessed data to obtain a training dataset, the key elements including event type, event background, participants, and decision-making style; training a large language model based on the training dataset and a policy optimization algorithm, introducing new training data into the large language model during the training process, adjusting the model parameters to obtain a trained large language model; and encapsulating the trained large language model into an agent, and integrating multiple algorithm processing logics into the agent.

[0008] In some embodiments of this disclosure, the data sources include news media, academic literature, policy documents, social media and historical cases, and the deduction elements include game name, theme, deduction background, deduction purpose, deduction parties, initial value settings of deduction party indicators and initial situation events.

[0009] In some embodiments of this disclosure, constructing a decision preference matrix with added interference coefficients includes: encoding the enterprise's characteristic vector and preference vector, wherein the k-th element of the preference vector is represented as:

[0010] In the formula, w ij It is the strategic weighting coefficient, P ij Let m represent the firm's preference for strategy j in market scenario i, m be the total number of strategies, and n be the number of market scenarios; an interference matrix R is constructed using the Monte Carlo method. ij R ij =Random(0.5,1); Before each game, according to P k ×R ij By determining the decision preference vectors of each party in this game, we obtain the decision preference matrix with interference coefficients added.

[0011] In some embodiments of this disclosure, the firm’s characteristic vector includes firm size, financial condition, technological level and market share, and the preference vector is the firm’s inclination toward different strategic directions, including expansion, contraction, innovation and cost control.

[0012] In some embodiments of this disclosure, constructing an intelligent agent self-game model based on inference elements and a decision preference matrix, and generating multiple game scenarios and paths based on the intelligent agent self-game model includes: inputting feature vectors and decision preference vectors into a large model based on a Transformer architecture, capturing the relationship between feature vectors and decision preference vectors through a self-attention mechanism, and generating a preliminary strategic framework for the enterprise. The self-attention mechanism is... Where Q, K, and V are the query, key, and value vectors, respectively, and d kThe key vector dimension is used; the initial strategic framework is input into a small model based on a deep Q-network, and the parameters of the deep Q-network model are fine-tuned through backpropagation algorithm, combining historical strategic data of the enterprise and current market sample data, in order to refine the initial strategic framework; the external environment of the enterprise is continuously monitored, and external environmental factors are input into a large model, so that the large model can predict market changes or policy trends based on the input external environmental factors, provide strategic adjustment suggestions, and the small model updates the Q value based on the strategic adjustment suggestions, combined with the current market status and competitive environment; an intelligent agent self-game model is constructed based on the large model of the Transformer architecture and the small model based on the deep Q-network; and the intelligent agent self-game model automatically performs multi-round game simulations to generate game scenarios and game paths covering a variety of strategy combinations.

[0013] In some embodiments of this disclosure, the deep Q-network includes multiple convolutional layers and fully connected layers. The convolutional layers are used to extract market competition features, and the fully connected layers are used to output strategic decisions. The loss function of the deep Q-network is: Where y i For the target Q value, s i ,a j Let Q be a state-action pair, and θ be the model parameters. After each strategic decision is implemented, the Q value is updated based on the new market state and the obtained reward, using the following formula:

[0014] Q(s,a)=Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)],

[0015] Where α is the learning rate, γ is the discount factor, r is the reward obtained, (s,a) is the current state-action pair, and (s′,a′) is the new state-action pair.

[0016] In some embodiments of this disclosure, a strategic evaluation model is constructed, and risk assessment and trend prediction are performed on multiple game scenarios and paths based on the strategic evaluation model to generate a simulation report. This includes: calculating the immediate payoff of each game path; assessing the execution difficulty, required resources, reputational risk, opponent's reaction strategy, and long-term effects after strategy execution for each strategy; selecting the strategy combination with the highest payoff under different decision-making styles; analyzing the historical behavior of the players and the current game path; predicting the evolution trend of the players' strategy direction, resource allocation preferences, and decision-making style; establishing a dynamic event model to simulate market changes or sudden events; predicting the players' behavior in different situations based on the current game path; and automatically generating a standardized game simulation report according to a predetermined format. The game simulation report includes: game path and decision analysis; assessment of the payoff, risk, and long-term impact of each strategy; analysis of the strategic tendencies of each player; trend prediction of the dynamic event model; and contingency plans and risk mitigation measures.

[0017] According to a second aspect of this disclosure, an intelligent strategy deduction device for machine self-game is provided. The device includes at least one processor and at least one memory storing a computer program. When the computer program is executed by the at least one processor, the device causes the following actions: constructing a game event agent and extracting deduction elements of the event to be deduced based on the game event agent; constructing a decision preference matrix with added interference coefficients; constructing an intelligent agent self-game model based on the deduction elements and the decision preference matrix, and generating multiple sets of game scenarios and game paths based on the intelligent agent self-game model, wherein the intelligent agent self-game model includes a large model based on a Transformer architecture and a small model based on a deep Q-network; and constructing a strategic evaluation model, and performing risk assessment and trend prediction on the multiple sets of game scenarios and game paths based on the strategic evaluation model, generating a deduction report.

[0018] In some embodiments of this disclosure, the computer program, when executed by at least one processor, causes the device to construct a game-theoretic agent by:

[0019] Raw text data is collected from multiple data sources, and the collected raw text data is cleaned and preprocessed. Key elements of the preprocessed data are labeled to obtain a training dataset. Key elements include event type, event background, participants, and decision-making style. A large language model is trained based on the training dataset and policy optimization algorithms. During the training process, new training data is introduced into the large language model, and the model parameters are adjusted to obtain a well-trained large language model. Finally, the well-trained large language model is encapsulated into an intelligent agent, and various algorithm processing logics are integrated into the intelligent agent.

[0020] In some embodiments of this disclosure, the computer program, when executed by at least one processor, causes the device to construct a decision preference matrix incorporating interference coefficients by:

[0021] The enterprise's characteristic vector and preference vector are encoded, and the k-th element of the preference vector is represented as:

[0022] In the formula, w ij It is the strategic weighting coefficient, P ij Let m represent the firm's preference for strategy j in market scenario i, m be the total number of strategies, and n be the number of market scenarios; an interference matrix R is constructed using the Monte Carlo method. ij R ij =Random(0.5,1); Before each game, according to P k ×R ij By determining the decision preference vectors of each party in this game, we obtain the decision preference matrix with interference coefficients added.

[0023] In some embodiments of this disclosure, when executed by at least one processor, the computer program causes the device to generate multiple sets of game scenarios and game paths by: inputting feature vectors and decision preference vectors into a large model based on a Transformer architecture; capturing the relationship between the feature vectors and decision preference vectors through a self-attention mechanism; and generating a preliminary strategic framework for the enterprise. The self-attention mechanism is... Where Q, K, and V are the query, key, and value vectors, respectively, and d k The dimension of the key vector;

[0024] The initial strategic framework is input into a small model based on a deep Q-network. Combined with historical strategic data and current market sample data, the parameters of the deep Q-network model are fine-tuned using a backpropagation algorithm to refine the initial strategic framework. The external environment is continuously monitored, and external environmental factors are input into a large model. This allows the large model to predict market changes or policy trends based on the input external environmental factors, providing strategic adjustment suggestions. The small model updates its Q-value based on these strategic adjustment suggestions, combined with the current market state and competitive environment. An intelligent agent self-game model is constructed using the large model based on the Transformer architecture and the small model based on the deep Q-network. This intelligent agent self-game model automatically performs multi-round game simulations, generating game scenarios and paths covering various strategy combinations.

[0025] In some embodiments of this disclosure, the computer program, when executed by at least one processor, causes the device to generate a deduction report by:

[0026] The system calculates the immediate payoff for each game path, assesses the execution difficulty, required resources, reputational risk, opponent's response strategies, and long-term effects of each strategy, and selects the strategy combination with the highest payoff under different decision-making styles. It analyzes the historical behavior and current game path of the players, predicting the evolution trends of their strategy direction, resource allocation preferences, and decision-making styles. It establishes a dynamic event model to simulate market changes or unexpected events, predicting the players' behavior in different scenarios based on the current game path. Furthermore, it automatically generates standardized game simulation reports according to a predetermined format. These reports include: game path and decision analysis; assessment of the payoff, risk, and long-term impact of each strategy; analysis of each player's strategy tendencies; trend predictions from the dynamic event model; and contingency plans and risk mitigation measures.

[0027] According to a third aspect of this disclosure, a computer-readable storage medium storing a computer program is provided, wherein the computer program, when executed by a processor, implements the steps of the intelligent strategy deduction method for machine self-play according to a first aspect of this disclosure. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments will be briefly described below. It should be understood that the drawings described below only relate to some embodiments of this disclosure and are not intended to limit this disclosure, wherein:

[0029] Figure 1 A flowchart illustrating the overall process design of the intelligent strategy deduction method according to an embodiment of the present disclosure is shown.

[0030] Figure 2 An exemplary flowchart is shown for an intelligent strategy deduction method 200 for machine self-play according to an embodiment of the present disclosure;

[0031] Figure 3 A flowchart illustrating the construction process of an intelligent agent self-game model according to an embodiment of this disclosure is shown;

[0032] Figure 4 This is a schematic block diagram of an intelligent strategy deduction device 400 for machine self-play according to an embodiment of the present disclosure.

[0033] It should be noted that the elements in the attached diagram are schematic and not drawn to scale. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are also within the scope of protection of this disclosure.

[0035] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this subject matter pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having meanings consistent with their meanings in the context of the specification and in the relevant art, and shall not be interpreted in an idealized or overly formal form unless otherwise explicitly defined herein. Furthermore, terms such as “first” and “second” are used only to distinguish one component (or part of a component) from another component (or another part of a component).

[0036] Utilizing artificial intelligence technology for decision simulation and optimization, especially the innovation of generative artificial intelligence and the collaborative scheduling technology of large and small models, provides new directions and approaches for improving the efficiency of strategic simulation.

[0037] To address the problem of strategic deduction and decision optimization in complex environments, this disclosure proposes an intelligent strategic deduction method and apparatus based on machine self-game theory. Figure 1 A flowchart illustrating the overall process design of the intelligent strategy deduction method according to an embodiment of this disclosure is shown. (Refer to...) Figure 1 As shown, first, a game is created.

[0038] The game event intelligence agent performs multi-level and multi-angle in-depth analysis of hot events. Utilizing technologies such as natural language processing, semantic analysis, and information extraction, the agent extracts core elements of events from massive amounts of data, including the background, participants, involved fields, and themes, ensuring the accuracy and relevance of subsequent game simulations.

[0039] A decision preference matrix is ​​constructed. Through deep learning and historical data analysis, the decision preferences of each participant are quantified, thereby generating their strategic intentions and policy propositions under different scenarios, and constructing a decision preference matrix. The matrix mainly reflects the decision-making tendencies of the participants and provides personalized and diversified strategy inputs for subsequent game simulations.

[0040] Then, the intelligent strategic self-game model, based on the decision preference matrix and event background, generates game agents representing each participant (Company A, Company B, and Company C), and conducts automatic game playing using a collaborative scheduling method of large and small models. It simulates the dynamic development of events, explores the optimal strategy path, and realizes the self-iteration and optimization of the strategy.

[0041] Finally, the strategic evaluation model deeply integrates and analyzes the results of multiple rounds of game theory. Utilizing advanced data analysis, trend forecasting, and risk assessment techniques, the model extracts the development trends, potential risks, and corresponding strategies for hot-button events, automatically generating detailed analysis reports and trend analysis reports to provide decision-makers with rapid and accurate decision-making references.

[0042] Figure 2 An exemplary flowchart is shown for an intelligent strategy deduction method 200 for machine self-play according to an embodiment of the present disclosure.

[0043] exist Figure 2 At frame S202, construct the game event agent and extract the deduction elements of the event to be deduced based on the game event agent.

[0044] The process of building intelligent agents for game events is a systematic and data-driven endeavor, encompassing the entire process from data collection to model training and application. Its aim is to accurately analyze complex game events through the integration and intelligent analysis of multi-source data, providing a scientific basis for decision support and game simulation.

[0045] To enable the constructed intelligent agent to comprehensively understand game events, according to one embodiment of this disclosure, raw text data is collected from multiple data sources, and the collected raw text data undergoes data cleaning and preprocessing. This raw text data can cover different types of game events and their background information. For example, texts related to current events, hot topics, and economic competition are obtained from mainstream news websites, online news platforms, and various special reports; these texts typically possess strong timeliness and comprehensiveness. Content is collected from academic journals, research papers, conference proceedings, etc., especially research results in fields such as social sciences, game theory, economics, and international relations, to provide theoretical support and in-depth analysis for the model. Furthermore, policy documents, government reports, and various official announcements are also key sources for obtaining information on decision-making backgrounds, stakeholders, and policy analysis. Collecting user-generated content through social platforms such as Twitter and Weibo helps the model understand public sentiment, differing opinions, and their reactions to game events. Additionally, by analyzing historical case studies to understand the deductive patterns of similar events, the long-term evolution and impact of the behaviors of various parties in the events can be understood.

[0046] Next, data cleaning and preprocessing are performed on the multi-source data to ensure that the obtained text data is suitable for model training. For example, duplicate, irrelevant, and redundant text is removed, and the text format is standardized to resolve inconsistencies in formatting across different data sources. Word segmentation and part-of-speech tagging techniques are used to preprocess the text, preparing it for subsequent deep analysis and annotation. Key information in the text, such as the time, location, participants, and behaviors of events, is automatically identified. Natural language processing techniques (such as sentiment analysis and semantic role labeling) are used to extract deeper information from the text, understanding the underlying game patterns and decision-making logic involved.

[0047] Data annotation can identify the basic components of a game, such as event type, background, key participants, and decision-making styles, as well as initial conditions, based on the characteristics of the game event. This provides clear learning objectives and labels for large-scale models. For example, event types include economic games. Event background includes historical context and economic conditions. Key participants in the event, such as countries, companies, organizations, and individuals, are labeled, and their interests, positions, and resources are further analyzed. Decision-making styles of each party include cooperation, confrontation, compromise, and possible strategies (such as cooperation strategies and betrayal strategies in the game). Key event factors include the cause of conflict, resource allocation, and transaction terms. Initial conditions are set for each participant, including: resource quantity (such as fiscal funds and natural resources), economic strength (such as GDP and market share), and public support (such as public support and public opinion). The specific context at the start of the game is defined, including all observable variables and the initial state; these variables form the basis for game deduction.

[0048] To ensure the quality of the training corpus, a combination of expert review and automation can be used to guarantee the validity and accuracy of the data. After data integration and annotation, the large language model can be trained and optimized based on the training dataset and policy optimization algorithms, enabling the trained model to analyze the key elements of game events. For example, pre-trained language models such as Transformer and BERT can be used to train the model in game theory, context analysis, and strategy selection, helping the model better understand the deeper logic of the game. Key elements include the game name, theme, deduction background, deduction purpose, deducers, initial values ​​of deducer indicators, and initial situation events. As new data continuously emerges, the model is periodically updated through incremental learning; that is, new training data is introduced into the large language model during training to keep it sensitive to new game events, and the model parameters are adjusted online to obtain a well-trained large language model. Finally, the trained large model is encapsulated into an intelligent agent and integrated with multiple algorithmic processing logics, enabling the agent to comprehensively utilize different algorithmic processing logics when facing complex game situations.

[0049] Reference Figure 2 As shown, at box S204, a decision preference matrix with added interference coefficients is constructed.

[0050] When constructing the decision preference matrix, the decision support model is primarily formed by encoding the enterprise's characteristic vectors and preference vectors. The enterprise's characteristic vectors include enterprise size, financial condition, technological level, and market share. For example, enterprise size is typically quantified by the number of employees, revenue, or total assets. Financial condition can be represented by indicators such as profit margin, debt ratio, and capital structure. Technological level can be measured by R&D investment ratio and the number of technology patents. Market share reflects the enterprise's market share in the industry, usually expressed as a percentage.

[0051] These characteristic vectors are standardized to convert data from different dimensions into dimensionless data, which can then be input into the model for subsequent analysis. Commonly used standardization methods include Z-score standardization and Min-Max standardization. After standardization, each company's characteristic vectors can be directly compared with those of other companies.

[0052] The preference matrix is ​​used to represent a firm's inclination towards different strategic directions under different market scenarios, and it is converted into a preference vector. Suppose that a firm may face different market scenarios, such as market scenario 1 being a highly competitive market, market scenario 2 being a market with high growth potential, and market scenario 3 being a mature market. Under each market scenario, the firm may have different strategic choices, such as expansion, contraction, innovation, and cost control. Based on this, an n×m preference matrix P is constructed, where each element P... ijThis indicates the firm's preference for strategy j in market scenario i. The degree of preference can be determined from sources such as historical firm data, industry analysis, and strategic objectives. For example, P... 11 P represents the firm's preference for an "expansion" strategy in Scenario 1. 12 This indicates the degree of preference of enterprises for the "contraction" strategy under Scenario 1.

[0053] To more accurately quantify the preference for each strategy, the preference vector needs to be further refined and incorporated with the company's strategic priorities. Assume that a certain weighting coefficient is assigned to the strategic direction in each market scenario; these weights reflect the degree of importance the company places on each strategy in different scenarios. Preference Vector The k-th element can be calculated as w ij This represents the strategic weighting coefficient, reflecting the importance of each strategy. m is the total number of strategies, and n is the number of market scenarios. Using this calculation method, the preference vector for each market scenario can reflect the firm's strategic inclinations.

[0054] In multi-party games, besides the decision preference matrix, firms' strategic choices are also influenced by various external factors and uncertainties (such as market fluctuations, policy changes, and technological innovation). Therefore, complex scenario interference coefficients are incorporated into the decision-making model. To simulate these uncertainties, the Monte Carlo method can be used to construct an interference matrix, and interference coefficients can be added during the game process to enhance the model's realism and adaptability.

[0055] According to one embodiment of this disclosure, the interference coefficient typically fluctuates within a specific range, such as 0.5 to 1, meaning that the interference coefficient will either reduce the original decision preference (less than 1) or remain unchanged (equal to 1). The interference matrix R is constructed using the Monte Carlo method. n×m R ij R represents a random number in the range (0.5 to 1). ij =Random(0.5,1), this random number can be generated using a uniform distribution or by fitting a distribution to historical data. Before each game, according to P... k ×R ij By determining the decision preference vectors of each party in this game, we obtain the decision preference matrix with added interference coefficients. These adjusted preference vectors serve as preconditions for the game theory model. During the game, each party selects the most suitable strategy based on these adjusted preference vectors and conducts game analysis. In this way, agents can make more accurate and realistic strategic choices based on actual market uncertainties.

[0056] In various market scenarios, intelligent agents can provide companies with reasonable strategic choices based on their different strategic preferences and company characteristics. For example, when market conditions change, companies can adjust their strategic choices according to a preference matrix, and the company's decision preferences will be adjusted according to the uncertainty before each game, thereby improving the actual effectiveness and accuracy of decision-making.

[0057] Subsequently, at box S206, an intelligent agent self-game model is constructed based on the inference elements and decision preference matrix. Multiple game scenarios and game paths are generated based on the intelligent agent self-game model. The intelligent agent self-game model includes a large model based on the Transformer architecture and a small model based on a deep Q-network.

[0058] Enterprise-level strategic simulations typically involve multiple factors, such as the market environment, competitor dynamics, and the company's own resources and capabilities. To address these complex influencing factors, an intelligent agent self-game model can be constructed to help companies dynamically adjust and make decisions based on actual situations. This process involves the collaborative action of a large model and smaller models to jointly build a framework for enterprise strategic simulation and optimization.

[0059] Figure 3 A flowchart illustrating the construction process of an intelligent agent self-game model according to an embodiment of this disclosure is shown. Figure 3 As shown, firstly, a preliminary strategic framework is generated based on the large model.

[0060] In some embodiments of this disclosure, feature vectors and decision preference vectors are input into a large model based on the Transformer architecture. A self-attention mechanism is used to capture the relationship between the feature vectors and decision preference vectors, generating a preliminary strategic framework for the enterprise. The self-attention mechanism is... Where Q, K, and V are the query, key, and value vectors, respectively, and d k The dimension of the key vector.

[0061] Specifically, the large model receives the enterprise's characteristic vectors. (e.g., market size, competitive landscape, corporate resources, etc.) and preference vector (such as corporate strategic preferences, market objectives, etc.). Through self-attention mechanisms ( Where Q, K, and V are the query, key, and value vectors, respectively, and d kThe model captures the relationships between input characteristics and decision preference vectors (using the key vector dimension) to identify the importance and interrelationships of various factors. Specifically, the query vector (Q) represents the needs or problems of the current strategy. The key vector (K) represents potential strategic directions or market information. The value vector (V) is the specific content extracted from market and strategic information. The model generates a preliminary strategic framework through these relationships, outputting key market areas that the company should focus on, directions for cultivating core competencies, etc. For example, deciding whether to increase R&D investment or expand into emerging markets.

[0062] Subsequently, the initial strategic framework is refined based on the small model. At this point, the initial strategic framework is input into the small model based on a deep Q-network, and combined with the company's historical strategic data and current market sample data, the parameters of the deep Q-network model are fine-tuned using the backpropagation algorithm to refine the initial strategic framework. The specific steps are as follows:

[0063] The DQN model comprises a multi-layered neural network: convolutional layers extract features such as market competition dynamics and identify key factors influencing market changes. Fully connected layers output specific strategic decisions based on the extracted features. The strategic framework generated by the large model serves as the initial input to the DQN model. Historical and sample data are used to adjust the DQN model's parameters through backpropagation, thereby optimizing the strategic decisions. The loss function of DQN is:

[0064] Where y i For the target Q value, s i ,a j For state-action pairs (such as market state and corresponding strategic action), θ represents the DQN parameters, and the loss function is optimized using stochastic gradient descent (SGD). After multiple training and adjustments, the DQN model can refine a firm's strategic decisions, making them more aligned with market demands and firm objectives.

[0065] During the collaborative operation of the large and small models, strategy parameters are adjusted based on environmental monitoring. The large model not only provides a preliminary strategic framework but also continuously monitors the external environment of the enterprise, including market trends, policy and regulatory changes, and technological innovations, providing timely macro-level suggestions for strategic adjustments to the DQN small model. The large model encodes these external environmental factors as vector inputs through analysis of industry news, government policies, competitor announcements, and other information. Based on the input environmental factors, the large model predicts potential market changes or policy trends. For example, when a market transformation is predicted, the large model suggests that the enterprise adjust its strategy, such as shifting to new technologies or market sectors. The large model provides macro-level strategic adjustment suggestions to the DQN model, such as reallocating resources when the market shifts. Based on these suggestions, the DQN small model, combined with its current strategic state and perception of the competitive landscape in local markets (such as recent competitor actions and the performance of its own products in specific market segments), further optimizes strategic decisions through SGD. After each strategic decision is implemented, the Q-value is updated based on the new market state s′ and the obtained reward r (such as market share growth, profit improvement, etc.), using the formula:

[0066] Q(s,a)=Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)], where α is the learning rate, γ is the discount factor, r is the reward obtained, (s,a) is the current state-action pair, and (s′,a′) is the new state-action pair.

[0067] Through the above steps, a large model based on the Transformer architecture and a small model based on a deep Q-network are constructed to create an intelligent agent self-game model for strategic game simulation. By leveraging the synergy between the large and small models, the potential behaviors of competitors can be simulated, and corresponding strategic adjustments can be made. The large model analyzes textual information such as industry news, competitor announcements, and financial reports to uncover the underlying intentions behind competitors' market actions. The small model (DQN) predicts the possible actions of competitors based on historical data and current market dynamics. For example, if it predicts that a competitor will lower prices in a certain product area, the DQN model can decide whether to follow suit or respond by using differentiation strategies such as increasing product added value.

[0068] The intelligent agent automatically conducts multi-round game simulations using a self-playing model, generating game scenarios and paths encompassing various strategy combinations. During the game, the intelligent agent makes multiple decisions, considering not only its own strategy but also anticipating the opponent's possible actions and adjusting its strategic responses. For example, the intelligent agent can adjust its resource allocation and marketing strategies in a timely manner based on the opponent's pricing strategies and marketing activities. Through the self-playing mechanism, the intelligent agent simulates game paths under different scenarios, exploring optimal response strategies. This process generates multiple strategy combinations and provides comprehensive analytical results for subsequent strategic evaluation models. Through multi-round game simulations, the intelligent agent generates various possible game paths, providing diversified and comprehensive evaluation results for corporate strategic decision-making, helping companies make optimal decisions in complex and uncertain market environments.

[0069] Finally, at box S208, a strategic assessment model is constructed, and based on the strategic assessment model, risk assessment and trend prediction are performed on multiple game scenarios and game paths to generate a simulation report.

[0070] According to one embodiment of this disclosure, after multiple rounds of machine game simulation, a series of strategic results are derived by deeply analyzing the game path, assessing risks and benefits, and identifying the strategic tendencies of each party, providing decision-makers with detailed contingency plans and forecast reports.

[0071] The specific analytical elements for multiple game paths include: evaluating the direct economic benefits of each option based on factors such as current market conditions, product pricing, and marketing strategies; considering the long-term effects after the strategy is implemented, such as the enhancement of brand value and the increase in market share; and including the synergistic effects in terms of market competition, resource consumption, and technological innovation.

[0072] It can calculate the immediate payoff of each game path, assess the execution difficulty, required resources, reputational risk, opponent's response strategies, and long-term effects of each strategy, and screen out the strategy combinations with the highest payoff under different decision-making styles. It analyzes the historical behavior of the players and the current game path to predict the evolution trends of their strategy direction, resource allocation preferences, and decision-making styles. It establishes dynamic event models to simulate market changes or unexpected events, predicting the players' behavior in different scenarios based on the current game path. Finally, it automatically generates standardized game simulation reports according to a predetermined format. The game simulation reports include: game path and decision analysis; assessment of the payoff, risk, and long-term impact of each strategy; analysis of the strategic tendencies of each player; trend predictions from the dynamic event model; and contingency plans and risk mitigation measures.

[0073] Figure 4 This is a schematic block diagram of a machine self-playing intelligent strategy deduction device 400 according to an embodiment of the present disclosure. Figure 4As shown, the device 400 may include a processor 410 and a memory 420 storing a computer program. When the computer program is executed by the processor 410, the device 400 is made capable of performing actions such as... Figure 2 The steps of method 200 are shown below. In one example, device 400 may be a computer device or a cloud computing node. Device 400 can construct a game event agent and extract inference elements of the event to be inferred based on the game event agent. Device 400 can construct a decision preference matrix with added interference coefficients. In response to constructing the game event agent and the decision preference matrix, device 400 can construct an intelligent agent self-game model based on the inference elements and the decision preference matrix, and generate multiple sets of game scenarios and game paths based on the intelligent agent self-game model. The intelligent agent self-game model includes a large model based on the Transformer architecture and a small model based on a deep Q-network. In response to the intelligent agent self-game model generating multiple sets of game scenarios and game paths, device 400 can construct a strategic evaluation model and perform risk assessment and trend prediction on the multiple sets of game scenarios and game paths based on the strategic evaluation model, generating an inference report.

[0074] In some embodiments of this disclosure, the device 400 can collect raw text data from multiple data sources, perform data cleaning and preprocessing on the collected raw text data; label the key elements of the preprocessed data to obtain a training dataset, the key elements including event type, event background, participants, and decision-making style; train a large language model based on the training dataset and policy optimization algorithm, introduce new training data into the large language model during the training process, adjust the model parameters to obtain a trained large language model; and encapsulate the trained large language model into an intelligent agent, and integrate multiple algorithm processing logics in the intelligent agent.

[0075] In some embodiments of this disclosure, the device 400 can encode the enterprise's characteristic vector and preference vector, wherein the k-th element of the preference vector is represented as:

[0076] In the formula, w ij It is the strategic weighting coefficient, P ij This represents the degree of preference of a firm for strategy j in market scenario i, where m is the total number of strategies and n is the number of market scenarios.

[0077] The interference matrix R is constructed using the Monte Carlo method. ij R ij =Random(0.5,1); Before each game, according to P k ×R ij By determining the decision preference vectors of each party in this game, we obtain the decision preference matrix with interference coefficients added.

[0078] In some embodiments of this disclosure, device 400 can input feature vectors and decision preference vectors into a large model based on the Transformer architecture, capture the relationship between feature vectors and decision preference vectors through a self-attention mechanism, and generate a preliminary strategic framework for the enterprise. The self-attention mechanism is... Where Q, K, and V are the query, key, and value vectors, respectively, and d k The key vector dimension is used; the initial strategic framework is input into a small model based on a deep Q-network, and the parameters of the deep Q-network model are fine-tuned through backpropagation algorithm, combining historical strategic data of the enterprise and current market sample data, in order to refine the initial strategic framework; the external environment of the enterprise is continuously monitored, and external environmental factors are input into a large model, so that the large model can predict market changes or policy trends based on the input external environmental factors, provide strategic adjustment suggestions, and the small model updates the Q value based on the strategic adjustment suggestions, combined with the current market status and competitive environment; an intelligent agent self-game model is constructed based on the large model of the Transformer architecture and the small model based on the deep Q-network; and the intelligent agent self-game model automatically performs multi-round game simulations to generate game scenarios and game paths covering a variety of strategy combinations.

[0079] In some embodiments of this disclosure, the device 400 can calculate the immediate payoff of each game path, assess the execution difficulty, required resources, reputational risk, opponent's response strategy, and long-term effects after strategy execution for each strategy, and screen out the strategy combination with the highest payoff under different decision-making styles; analyze the historical behavior of the players and the current game path, and predict the evolution trend of the players' strategy direction, resource allocation preferences, and decision-making style; establish a dynamic event model to simulate market changes or sudden events, and predict the behavior of the players in different situations based on the current game path; and automatically generate a standardized game simulation report according to a predetermined format. The game simulation report includes: game path and decision analysis; assessment of the payoff, risk, and long-term impact of each strategy; analysis of the strategic tendencies of each player; trend prediction of the dynamic event model; and contingency plans and risk mitigation measures.

[0080] In some embodiments of this disclosure, the device 400 may be, in some embodiments of this disclosure, a processor 410 may be, for example, a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a processor based on a multi-core processor architecture, etc. The memory 420 may be any type of memory implemented using data storage technology, including but not limited to random access memory, read-only memory, semiconductor-based memory, flash memory, disk storage, etc.

[0081] Furthermore, in embodiments of this disclosure, device 400 may also include an input device 430, such as a keyboard or mouse, for inputting hot events, training datasets from multiple data sources, etc. Additionally, device 400 may also include an output device 440, such as a display, for outputting inference reports, etc.

[0082] In other embodiments of this disclosure, a computer-readable storage medium storing a computer program is also provided, wherein the computer program, when executed by a processor, is capable of performing the following functions: Figure 2 The steps of the intelligent strategy deduction method for machine self-game are shown.

[0083] In summary, the intelligent strategy deduction method and apparatus for machine self-play according to embodiments of this disclosure, by constructing an intelligent agent for game events, can perform comprehensive simulation and deduction based on the deduction elements of actual events. The intelligent agent can extract multi-dimensional elements related to the game according to specific scenarios, forming multiple game paths, and can simulate dynamic changes in the game, ensuring accurate reproduction of complex game scenarios. The decision preference matrix with added interference coefficients can dynamically adjust the decision preferences of the decision-makers, reflecting uncertainties or external interference factors in the game environment. This mechanism makes the deduction results more closely resemble the actual dynamic environment, improving the accuracy and practicality of decision deduction.

[0084] The large model, based on the Transformer architecture, can handle high-dimensional and long-step game paths, optimizing long-term dependencies in game inference. The small model, based on a deep Q-network, can efficiently perform reinforcement learning and quickly adapt to the players' immediate decisions. This collaborative optimization scheme between the large and small models makes the inference results more accurate and flexible.

[0085] By analyzing multiple game paths, the strategic assessment model can identify the potential strategic tendencies of different players, predict the next development of the game, and generate contingency plans based on the predictions. This helps decision-makers prepare in advance for unexpected events or complex market changes, reducing the risks of strategy implementation.

[0086] Therefore, through machine game simulation and risk assessment, this solution can quantify decision-making risks and provide specific optimization suggestions, helping decision-makers make more informed decisions and reduce risks caused by incomplete or uncertain information. It can be applied to various complex game scenarios.

[0087] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatuses and methods according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0088] Unless otherwise expressly indicated by the context, the singular form of words used herein and in the appended claims includes the plural form, and vice versa. Thus, when referring to the singular, the plural form of the corresponding term is generally included. Similarly, the terms “comprising” and “including” shall be interpreted as including rather than exclusively. Likewise, the terms “including” and “or” shall be interpreted as including unless such interpretation is expressly prohibited herein. Where the term “example” is used herein, particularly when it follows a set of terms, the “example” is merely exemplary and illustrative and should not be considered exclusive or extensive.

[0089] Further aspects and scope of adaptation become apparent from the description provided herein. It should be understood that various aspects of this application may be implemented individually or in combination with one or more other aspects. It should also be understood that the descriptions and specific embodiments herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0090] Several embodiments of this disclosure have been described in detail above. However, it is obvious that those skilled in the art can make various modifications and variations to the embodiments of this disclosure without departing from the spirit and scope of this disclosure. The scope of protection of this disclosure is defined by the appended claims.

Claims

1. A method for intelligent strategy deduction in machine self-game, characterized in that, include: Construct a game event intelligent agent, and extract the deduction elements of the event to be deduced based on the game event intelligent agent; Construct a decision preference matrix with added interference coefficients, which includes encoding the enterprise's characteristic vector and preference vector, wherein the k-th element of the preference vector is represented as: In the formula, w ij It is the strategic weighting coefficient, P ij Let m represent the firm's preference for strategy j in market scenario i, m be the total number of strategies, and n be the number of market scenarios; an interference matrix R is constructed using the Monte Carlo method. ij R ij =Random(0.5,1); Before each game, according to P k ×R ij Determine the decision preference vectors of each party in this game, and obtain the decision preference matrix with interference coefficients added; Based on the aforementioned inference elements and decision preference matrix, an intelligent agent self-game model is constructed. Multiple game scenarios and paths are generated based on this model. The intelligent agent self-game model includes a large model based on a Transformer architecture and a small model based on a deep Q-network. The process includes: inputting the feature vector and decision preference vector into the large model based on the Transformer architecture; capturing the relationship between the feature vector and decision preference vector through a self-attention mechanism to generate a preliminary strategic framework for the enterprise. The self-attention mechanism is... Where Q, K, and V are the query, key, and value vectors, respectively, and d k The key vector dimension is used; the initial strategic framework is input into a small model based on a deep Q-network, and combined with historical strategic data of the enterprise and current market sample data, the parameters of the deep Q-network model are fine-tuned through backpropagation algorithm to refine the initial strategic framework; the external environment of the enterprise is continuously monitored, and external environmental factors are input into the large model so that the large model can predict market changes or policy trends based on the input external environmental factors, provide strategic adjustment suggestions, and the small model updates the Q value based on the strategic adjustment suggestions, combined with the current market state and competitive environment; an intelligent agent self-game model is constructed based on the large model based on the Transformer architecture and the small model based on the deep Q-network; the intelligent agent self-game model automatically performs multi-round game simulations to generate game scenarios and game paths covering multiple strategy combinations; and A strategic assessment model is constructed, and based on the strategic assessment model, risk assessment and trend prediction are performed on the multiple sets of game scenarios and game paths, generating a simulation report.

2. The intelligent strategy deduction method for machine self-game as described in claim 1, characterized in that, The construction of the game-theoretic event intelligent agent includes: Raw text data is collected from multiple data sources, and the collected raw text data is then cleaned and preprocessed. The key elements of the preprocessed data are labeled to obtain the training dataset. The key elements include event type, event background, participants, and decision-making style. A large language model is trained based on the aforementioned training dataset and policy optimization algorithm. During training, new training data is introduced into the large language model, and the model parameters are adjusted to obtain a well-trained large language model. The trained large language model is encapsulated into an intelligent agent, and multiple algorithmic processing logics are integrated into the intelligent agent.

3. The intelligent strategy deduction method for machine self-game according to claim 1 or 2, characterized in that, The data sources include news media, academic literature, policy documents, social media, and historical cases. The simulation elements include the game name, theme, simulation background, simulation purpose, simulation parties, initial value settings of simulation party indicators, and initial situation events.

4. The intelligent strategy deduction method for machine self-game according to claim 1, characterized in that, The enterprise's characteristic vector includes enterprise size, financial condition, technological level, and market share. The preference vector represents the enterprise's inclination towards different strategic directions, including expansion, contraction, innovation, and cost control.

5. The intelligent strategy deduction method for machine self-game according to claim 1, characterized in that, The deep Q-network comprises multiple convolutional layers and fully connected layers. The convolutional layers are used to extract market competition features, and the fully connected layers are used to output strategic decisions. The loss function of the deep Q-network is: Where y i For the target Q value, s i ,a j Let Q be a state-action pair, and θ be the model parameters. After each strategic decision is implemented, the Q value is updated based on the new market state and the obtained reward, using the following formula: Q(s,a)=Q(s,a)+α[r+γmax a′ Q(s′,a′)-Q(s,a)], Where α is the learning rate, γ is the discount factor, r is the reward obtained, (s,a) is the current state-action pair, and (s′,a′) is the new state-action pair.

6. The intelligent strategy deduction method for machine self-game according to claim 1, characterized in that, The construction of a strategic assessment model, and the risk assessment and trend prediction of the multiple game scenarios and paths based on the strategic assessment model, generating a simulation report, includes: Calculate the immediate payoff for each game path, assess the execution difficulty, required resources, reputational risk, opponent's response strategy, and long-term effects of each strategy, and select the strategy combination with the highest payoff under different decision-making styles. Analyze the historical behavior and current game path of the players to predict the evolution trends of their strategic direction, resource allocation preferences, and decision-making styles. Establish dynamic event models to simulate market changes or unexpected events, and predict the behavior of players in different scenarios based on the current game path; and According to a predetermined format, a standardized game simulation report is automatically generated. The game simulation report includes: game path and decision analysis; assessment of the benefits, risks and long-term impacts of each strategy; analysis of the strategy tendencies of each player; trend prediction of the dynamic event model; and contingency plans and risk mitigation measures.

7. A machine-based self-game intelligent strategy deduction device, characterized in that, The device includes: At least one processor; and At least one memory storing a computer program; Wherein, when the computer program is executed by the at least one processor, the device performs the steps of the intelligent strategy deduction method for machine self-play according to any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent strategy deduction method for machine self-play according to any one of claims 1 to 6.

Images