AI-based cross-game experience migration system

Abstract

The invention discloses an AI-based cross-game experience migration system, which comprises a data acquisition module used for acquiring multi-modal data of a source game and a target game and performing normalization processing; the unified representation construction module is used for mapping various information of the source game and the target game to a unified feature space; the feature encoding module is used for inputting the extended unified representation into a feature encoder to obtain cross-game universal feature representation; the player portrait generation module is used for generating player portrait vectors based on the player historical behavior data; and the real-time reasoning module is used for outputting a target game control instruction through the target game action mapper. Compared with the prior art, the method combines comparative learning and domain adversarial learning, achieves unified modeling and general feature extraction of cross-game multi-modal data, has the advantages of being high in migration efficiency, high in generalization ability and good in personalized adaptability, and remarkably improves operation continuity and immersive experience consistency of players among different types of games.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and computer application, and particularly relates to a cross-game experience migration system based on AI. BACKGROUND

[0002] With the development of artificial intelligence technology and deep learning models, game intelligence and personalized experience have gradually become the focus of research and industry. Existing agent training methods mostly focus on a single game environment, using image frames, state variables and reward signals to train agents through supervised learning or reinforcement learning frameworks. These methods can achieve good performance in specific games.

[0003] Chinese patent application CN117258297A published on December 22, 2023 discloses a control method for AI agent behavior diversity in competitive games. The method abstracts and converts diverse behavior patterns in the game into a reward function, then combines these behavior patterns with existing observation information using OneHot encoding, and finally constructs a parallel diversity network outside the reinforcement learning backbone network. The combined observation is input into this network, and its output is added to the backbone network output to obtain the final network encoding.

[0004] In the above application, the trained model often lacks migration ability when facing different types or styles of games. Due to significant differences in visual style, state space, action set, and reward mechanism, traditional methods usually need to rebuild the model and conduct large-scale training for each new game, resulting in high computational cost and time overhead. At the same time, it also makes it difficult for players to continue their existing operation habits and experience when switching games. To address the lack of unified modeling, low migration efficiency and insufficient personalization in existing cross-game experience migration methods, the present application proposes a cross-game experience migration method based on multi-game transfer learning and domain adversarial learning. By collecting multi-modal data from source and target games, a unified representation is established, combining visual features, state variables and player input vectors for feature encoding. Contrastive learning is used to obtain semantic consistent cross-game general feature representation, and domain adversarial mechanism is used to achieve cross-domain distribution alignment. Meanwhile, personalized portrait vectors are generated based on player historical behavior and fused with general features. Finally, the target game control instructions are output through the action mapper, realizing the migration of cross-game experience. SUMMARY

[0005] An object of the present application is to propose an AI-based cross-game experience migration system. The present application builds a cross-game experience migration system composed of a data acquisition module, a unified representation construction module, a feature encoding module, a player portrait generation module, and a player portrait generation module. A multi-modal game data acquisition and normalization mechanism is adopted, combined with a feature encoding structure based on contrast learning and domain adversarial learning, to uniformly model visual embedding, state variables, player input, and reward signals, generate cross-game general feature representation, and fuse the player portrait vector of historical behavior with the general features. The system outputs the control instructions of the target game through the action mapper, realizing the continuous migration of cross-game experience. This method effectively solves the problems of insufficient data utilization, significant domain difference, low adaptation efficiency, and lack of personalization in existing cross-game migration technology, significantly improves the generalization ability, adaptability, and personalization effect of cross-game migration, and is suitable for intelligent migration applications in multiple types and styles of game environments.

[0006] The AI-based cross-game experience migration system according to an embodiment of the present application comprises: A data acquisition module for acquiring multi-modal data of source games and target games and performing normalization processing to build a cross-game training data set. A unified representation construction module for mapping each item of information of the source game and the target game in the cross-game training data set to a unified feature space and generating an extended unified representation. A feature encoding module for inputting the extended unified representation into a feature encoder to obtain a cross-game general feature representation. A player portrait generation module for calculating state mean vectors, input mean vectors, and reward means based on player historical behavior data to generate a player portrait vector. A real-time inference module for inputting the cross-game general feature representation and the player portrait vector into a target game action mapper and outputting target game control instructions through the target game action mapper.

[0007] Optionally, the operation steps of acquiring multi-modal data of source games and target games and performing normalization processing to build a cross-game training data set are as follows: S11, set the source game as , and the target game as , for each time step acquire multi-modal data to form an original sample five-tuple , wherein is an image frame, is the number of height pixels, is the number of width pixels, is the number of channels, is a state variable vector, In the state dimension, the state variable vector includes numerical information related to the game environment, such as player health, energy, position coordinates, orientation angle, and quest progress. Input vectors for players, As an input dimension, the player input vector includes data reflecting player actions such as keyboard key codes, mouse displacement, mouse click states, and gamepad control signals. The reward scalar includes numerical values ​​reflecting a player's gains or losses at that time step, such as score increments, mission rewards, changes in experience points, changes in resource collection, or penalty points. For timestamps; S12. Standardize and normalize the data in the original sample quintuples; S13. Assign a game identifier to each time step sample. , , This indicates that the sample originates from the source game. , This indicates that the sample originates from the target game. Normalized samples ,in, For the normalized image frame, This is the normalized state variable vector. The normalized player input vector, The normalized reward signal scalar; S14. Summarize the normalized samples from all time steps to form a cross-game training dataset. ,in, This represents the total number of time steps.

[0008] Optionally, the steps for mapping information from the source game and target game in the cross-game training dataset to a unified feature space and generating an extended unified representation are as follows: S21. Transfer image frames from the cross-game training dataset. The input is a visual encoder, which is a convolutional neural network, and the output is a visual embedding vector. S22, By analyzing the state variables in the cross-game training dataset and player input Perform a linear mapping to obtain the state embedding vector and the action embedding vector; S23. Weightedly fuse the visual embedding vector, state embedding vector, and action embedding vector to generate a unified representation vector. ; S24. The unified representation vector With game logo By concatenating the components, an extended unified representation is obtained. , ; S25. Summarize the extended unified representation of all time steps. To form a unified set of features , .

[0009] Optionally, the feature encoder consists of an input layer, a hidden layer, and an output layer.

[0010] Optionally, the feature encoder uses a contrastive learning method to train the feature representations of positive and negative samples, and trains the domain indiscriminateness of the feature encoder through adversarial training.

[0011] Optionally, the steps for training feature representations for positive and negative samples using the contrastive learning method are as follows: S31. Unify the representation of the extension at each time step. Input the feature encoder to obtain a cross-game universal feature representation. ; S32. Construct contrastive learning sample pairs, assuming the positive sample pair is... , indicating sample and Negative sample pairs are those from the same game domain and have adjacent or similar feature combinations. , indicating sample and Combinations of features from different or dissimilar fields; S33. Use cosine similarity to calculate sample pairs. Similarity between: ; in, Represents the Euclidean norm; S34. Construct a contrastive learning loss function based on the similarity, assuming the temperature parameter is... , Then the sample The contrast loss is: ; in, The total number of time steps. For temperature parameters, , It is a natural exponential function. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. S35. Summarize the contrastive losses of all samples to obtain the overall contrastive learning loss. : ; By minimizing Parameters of the feature encoder Optimize.

[0012] Optionally, the steps for training the domain indiscriminateness of the feature encoder through adversarial training are as follows: S41. Represent cross-game common features Input the gradient inversion layer, and then input the output of the gradient inversion layer into the neighborhood discriminator to obtain the neighborhood classification probability vector. , ,in, Indicates sample Game logo The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a constant during backward propagation to predict the probability. ,in, To provide the adversarial coefficient, the domain discriminator is a multilayer perceptron structure; S42. The parameters of the neighborhood discriminator are iteratively updated by minimizing the cross-entropy function, which is calculated as follows: ; in, The total number of time steps. For game identification, Indicates sample The probability of belonging to the source game. Indicates sample The probability of belonging to the target game is updated in the direction of minimizing the parameters of the neighborhood discriminator during backpropagation. The update direction of the feature encoder parameters is maximized due to the gradient reversal layer. This drives the feature encoder to generate domain-indistinguishable feature representations, and through iterative training, a set of encoded representations is obtained. In the feature space, the indistinguishability between the source game and the target game domain is satisfied, forming a cross-game universal feature representation.

[0013] Optionally, the steps for calculating the state mean vector, input mean vector, and reward mean based on player historical behavior data to generate the player profile vector are as follows: S51. Let the set of player historical behavior data be... ,in, For historical state variables, For historical input vectors, As a signal of historical rewards, This represents the number of historical samples. S52. Aggregate the historical state variables, historical input vectors, and historical reward signals to obtain the state mean vector. Input mean vector and average reward ; State mean vector The calculation method is as follows: ; in, This is the state mean vector; Input mean vector The calculation method is as follows: ; Average reward The calculation method is as follows: ; S53, Generate player profile vector : ; in, , It is a mapping matrix.

[0014] Optionally, the steps of inputting the cross-game universal feature representation and player profile vector into the target game motion mapper, and outputting the target game control commands through the target game motion mapper are as follows: S61, Vectorize the player profile Cross-game universal feature representation The vector is obtained by concatenating the two vectors. ; S62, merge vectors Input the target game action mapper to obtain the action prediction probability distribution. , ,in, The target game action mapper uses a fully connected neural network to determine the number of target game action categories. The action categories include forward, backward, left, right, jump, attack, defend, interact, pause, etc. S63. By selecting the action category with the highest probability as the target game step... Control commands; S64. Input the control command into the target game interface for execution.

[0015] The beneficial effects of this invention are: 1. This invention establishes a multimodal data acquisition and unified representation mechanism, realizing comprehensive modeling of image frames, state variables, player inputs, and reward signals. This avoids the problem of incomplete single-modal feature representation in existing technologies. Through the fusion of visual encoding, state mapping, and action mapping, this invention can generate a unified cross-game universal feature representation, aligning the source game and the target game in the same feature space, thereby providing a reliable data foundation for subsequent transfer learning.

[0016] 2. This invention combines contrastive learning and domain adversarial learning, which not only ensures semantic similarity but also explicitly eliminates domain differences, overcoming the shortcomings of insufficient adaptability and unstable generalization performance during the transfer process.

[0017] 3. This invention constructs player profile vectors and incorporates players' historical operating habits and reward preferences into the fusion process of cross-game common features. This method enables the transfer results to retain cross-game commonalities while reflecting individual player characteristics, thereby achieving stable, fast, and personalized cross-game experience transfer in multi-type and multi-style game environments, significantly improving the continuity and immersion of the player experience. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of the AI-based cross-game experience migration system proposed in this invention. Figure 2 This is a flowchart illustrating the domain indiscriminateness of the feature encoder trained through adversarial training proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 An AI-based cross-game experience migration system, including: The data acquisition module is used to collect multimodal data from the source game and the target game, perform normalization processing, and build a cross-game training dataset; The unified representation building module is used to map various information from the source game and the target game in the cross-game training dataset to a unified feature space and generate an extended unified representation; The feature encoding module is used to input the extended unified representation into the feature encoder to obtain a cross-game universal feature representation; The player profile generation module is used to calculate the state mean vector, input mean vector, and reward mean vector based on the player's historical behavior data, and generate a player profile vector. The real-time inference module is used to input cross-game common feature representations and player profile vectors into the target game action mapper, and output target game control commands through the target game action mapper.

[0021] In this embodiment, the steps for collecting multimodal data from the source game and the target game, performing normalization processing, and constructing a cross-game training dataset are as follows: S11, Setting the source game as The target game is For each time step Collect multimodal data to form original sample quintuples ,in, For image frames, For height in pixels, For width in pixels, For the number of channels, For the state variable vector, In the state dimension, the state variable vector includes numerical information related to the game environment, such as player health, energy, position coordinates, orientation angle, and quest progress. Input vectors for players, As an input dimension, the player input vector includes data reflecting player actions such as keyboard key codes, mouse displacement, mouse click states, and gamepad control signals. The reward scalar includes numerical values ​​reflecting a player's gains or losses at that time step, such as score increments, mission rewards, changes in experience points, changes in resource collection, or penalty points. For timestamps; S12. Standardize and normalize the data in the original sample quintuples; S13. Assign a game identifier to each time step sample. , , This indicates that the sample originates from the source game. , This indicates that the sample originates from the target game. Normalized samples ,in, For the normalized image frame, This is the normalized state variable vector. The normalized player input vector, The normalized reward signal scalar; S14. Summarize the normalized samples from all time steps to form a cross-game training dataset. ,in, This represents the total number of time steps.

[0022] In this embodiment, the steps for mapping various information from the source game and the target game in the cross-game training dataset to a unified feature space and generating an extended unified representation are as follows: S21. Transfer image frames from the cross-game training dataset. The input is a visual encoder, which is a convolutional neural network, and the output is a visual embedding vector. S22, By analyzing the state variables in the cross-game training dataset and player input Perform a linear mapping to obtain the state embedding vector and the action embedding vector; S23. Weightedly fuse the visual embedding vector, state embedding vector, and action embedding vector to generate a unified representation vector. ; S24. The unified representation vector With game logo By concatenating the components, an extended unified representation is obtained. , ; S25. Summarize the extended unified representation of all time steps. To form a unified set of features , .

[0023] In this embodiment, the feature encoder consists of an input layer, a hidden layer, and an output layer.

[0024] In this embodiment, the feature encoder uses a contrastive learning method to train the feature representation of positive and negative samples, and uses adversarial training to train the domain indiscriminateness of the feature encoder.

[0025] In this embodiment, the steps for training feature representations of positive and negative samples using a contrastive learning method are as follows: S31. Unify the representation of the extension at each time step. Input the feature encoder to obtain a cross-game universal feature representation. ; S32. Construct contrastive learning sample pairs, assuming the positive sample pair is... , indicating sample and Negative sample pairs are those from the same game domain and have adjacent or similar feature combinations. , indicating sample and Combinations of features from different or dissimilar fields; S33. Use cosine similarity to calculate sample pairs. Similarity between: ; in, Represents the Euclidean norm; S34. Construct a contrastive learning loss function based on the similarity, assuming the temperature parameter is... , Then the sample The contrast loss is: ; in, The total number of time steps. For temperature parameters, , It is a natural exponential function. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. S35. Summarize the contrastive losses of all samples to obtain the overall contrastive learning loss. : ; By minimizing Parameters of the feature encoder Optimize.

[0026] In this embodiment, the steps for training the domain indiscriminateness of the feature encoder through adversarial training are as follows: S41. Represent cross-game common features Input the gradient inversion layer, and then input the output of the gradient inversion layer into the neighborhood discriminator to obtain the neighborhood classification probability vector. , ,in, Indicates sample Game logo The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a constant during backward propagation to predict the probability. ,in, To provide the adversarial coefficient, the domain discriminator is a multilayer perceptron structure; S42. The parameters of the neighborhood discriminator are iteratively updated by minimizing the cross-entropy function, which is calculated as follows: ; in, The total number of time steps. For game identification, Indicates sample The probability of belonging to the source game. Indicates sample The probability of belonging to the target game is updated in the direction of minimizing the parameters of the neighborhood discriminator during backpropagation. The update direction of the feature encoder parameters is maximized due to the gradient reversal layer. This drives the feature encoder to generate domain-indistinguishable feature representations, and through iterative training, a set of encoded representations is obtained. In the feature space, the indistinguishability between the source game and the target game domain is satisfied, forming a cross-game universal feature representation.

[0027] In this embodiment, the steps for generating a player profile vector by calculating the state mean vector, input mean vector, and reward mean based on the player's historical behavior data are as follows: S51. Let the set of player historical behavior data be... ,in, For historical state variables, For historical input vectors, As a signal of historical rewards, This represents the number of historical samples. S52. Aggregate the historical state variables, historical input vectors, and historical reward signals to obtain the state mean vector. Input mean vector and average reward ; State mean vector The calculation method is as follows: ; in, This is the state mean vector; Input mean vector The calculation method is as follows: ; Average reward The calculation method is as follows: ; S53, Generate player profile vector : ; in, , It is a mapping matrix.

[0028] In this embodiment, the steps of inputting the cross-game universal feature representation and player profile vector into the target game motion mapper, and outputting the target game control commands through the target game motion mapper are as follows: S61, Vectorize the player profile Cross-game universal feature representation The vector is obtained by concatenating the two vectors. ; S62, merge vectors Input the target game action mapper to obtain the action prediction probability distribution. , ,in, The target game action mapper uses a fully connected neural network to determine the number of target game action categories. The action categories include forward, backward, left, right, jump, attack, defend, interact, pause, etc. S63. By selecting the action category with the highest probability as the target game step... Control commands; S64. Input the control command into the target game interface for execution.

[0029] Example In this embodiment, we take the example of a player who has been playing League of Legends for a long time migrating to Honor of Kings to illustrate the application process of the cross-game experience migration method of the present invention. The player has accumulated more than three years of game experience in League of Legends and has stable operating habits, skill release preferences and tactical selection methods. However, when he / she switches to Honor of Kings, due to the obvious differences in map design, operation interface and action commands between the two games, the player faces problems such as unfamiliarity with operation, low decision-making efficiency and fragmented game experience in the early stage. In the traditional way, the player needs to spend dozens of hours to familiarize himself / herself and train to reach the original level of operation, which is not only time-consuming and laborious, but also interrupts the continuity of the game experience.

[0030] When applying the cross-game experience migration method of this invention, the player's historical data in League of Legends was first collected and normalized through the data acquisition module. The collected data included game frames, hero position coordinates, skill release sequences, mouse clicks and keyboard input records for each game, as well as game reward signals (such as kills, assists, deaths, economic growth rate, etc.), which accumulated to form a dataset of more than 2,000 games and about 3TB. In the target game Honor of Kings, multimodal data of the player in 50 novice games were collected, and the data dimensions included movement trajectory, skill release rhythm, click frequency, gold coin increment, etc.

[0031] In the unified representation construction phase, the system uses a visual encoder to perform convolutional embedding on image frames to extract map environment and battlefield information; it uses a state mapping function to transform variables such as health, mana, economy, and position into state vectors; and it uses an action mapping function to transform keyboard and mouse input sequences into action vectors. Subsequently, these representations are mapped to a unified state-action feature space and concatenated with game domain labels to form a training dataset.

[0032] In the feature encoding stage, the system uses contrastive learning to compare the player's similar actions in different games, such as retreating with low health. The model is required to extract similar behavioral features, while differentiating dissimilar behaviors in different contexts. After about 48 hours of GPU training, the model obtains stable game-independent feature representations. In the domain adversarial training stage, the system introduces a domain discriminator and performs adversarial optimization through a gradient inversion layer to gradually eliminate the domain differences between League of Legends and Honor of Kings. This makes it impossible for the encoder to distinguish the source game when outputting features, resulting in truly cross-game universal features.

[0033] In actual testing, the method of this invention was applied to the player's migration scenario. The results showed that after only 5 games of adaptive training, the player's win rate in "Honor of Kings" increased from the initial 32.5% to 56.8%, an improvement of 21.3% compared with the control group players who did not use the method of this invention. At the same time, the player's average reaction time decreased from the initial 540 milliseconds to 412 milliseconds, and the success rate of skill combos increased from 68.2% to 85.6%, resulting in a significant improvement in the overall gaming experience.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-based cross-game experience migration system, characterized in that, include: The data acquisition module is used to collect multimodal data from the source game and the target game, perform normalization processing, and build a cross-game training dataset; The unified representation building module is used to map various information from the source game and the target game in the cross-game training dataset to a unified feature space and generate an extended unified representation; The feature encoding module is used to input the extended unified representation into the feature encoder to obtain a cross-game universal feature representation; The player profile generation module is used to calculate the state mean vector, input mean vector, and reward mean vector based on the player's historical behavior data, and generate a player profile vector. The real-time inference module is used to input cross-game common feature representations and player profile vectors into the target game action mapper, and output target game control commands through the target game action mapper.

2. The AI-based cross-game experience migration system according to claim 1, characterized in that, The steps for collecting multimodal data from the source game and the target game, performing normalization processing, and constructing a cross-game training dataset are as follows: S11, Setting the source game as The target game is For each time step Collect multimodal data to form original sample quintuples ,in, For image frames, For the state variable vector, Input vectors for players, To reward the signal scalar, For timestamps; S12. Standardize and normalize the data in the original sample quintuples; S13. Assign a game identifier to each time step sample. , , This indicates that the sample originates from the source game. , This indicates that the sample originates from the target game. Normalized samples ,in, For the normalized image frame, This is the normalized state variable vector. The normalized player input vector, The normalized reward signal scalar; S14. Summarize the normalized samples from all time steps to form a cross-game training dataset. ,in, This represents the total number of time steps.

3. The AI-based cross-game experience migration system according to claim 1, characterized in that, The steps for mapping information from the source game and target game in the cross-game training dataset to a unified feature space and generating an extended unified representation are as follows: S21. Transfer image frames from the cross-game training dataset. The input is a visual encoder, which is a convolutional neural network, and the output is a visual embedding vector. S22, By analyzing the state variables in the cross-game training dataset and player input Perform a linear mapping to obtain the state embedding vector and the action embedding vector; S23. Weightedly fuse the visual embedding vector, state embedding vector, and action embedding vector to generate a unified representation vector. ; S24. The unified representation vector With game logo By concatenating the components, an extended unified representation is obtained. , ; S25. Summarize the extended unified representation of all time steps. To form a unified set of features , .

4. The AI-based cross-game experience migration system according to claim 1, characterized in that, The feature encoder consists of an input layer, a hidden layer, and an output layer.

5. The AI-based cross-game experience migration system according to claim 1, characterized in that, The feature encoder uses a contrastive learning method to train the feature representation of positive and negative samples, and uses adversarial training to train the domain indiscriminateness of the feature encoder.

6. The AI-based cross-game experience migration system according to claim 5, characterized in that, The steps for training feature representations of positive and negative samples using the contrastive learning method are as follows: S31. Unify the representation of the extension at each time step. Input the feature encoder to obtain a cross-game universal feature representation. ; S32. Construct contrastive learning sample pairs, assuming the positive sample pair is... Negative sample pairs are ; S33. Use cosine similarity to calculate sample pairs. Similarity between: ； in, Represents the Euclidean norm; S34. Construct a contrastive learning loss function based on the similarity, assuming the temperature parameter is... , Then the sample The contrast loss is: ； in, The total number of time steps. For temperature parameters, , It is a natural exponential function. For indicator functions, when The value is 1 if the condition is met, and 0 otherwise. S35. Summarize the contrastive losses of all samples to obtain the overall contrastive learning loss. : ； By minimizing Parameters of the feature encoder Optimize.

7. The AI-based cross-game experience migration system according to claim 5, characterized in that, The steps for training the domain indiscriminateness of the feature encoder through adversarial training are as follows: S41. Represent cross-game common features Input the gradient inversion layer, and then input the output of the gradient inversion layer into the neighborhood discriminator to obtain the neighborhood classification probability vector. , ,in, Indicates sample Game logo The predicted probability; S42. The parameters of the neighborhood discriminator are iteratively updated by minimizing the cross-entropy function, which is calculated as follows: ； in, The total number of time steps. For game identification, Indicates sample The probability of belonging to the source game. Indicates sample The probability of belonging to the target game.

8. The AI-based cross-game experience migration system according to claim 1, characterized in that, The steps for generating a player profile vector by calculating the state mean vector, input mean vector, and reward mean based on player historical behavior data are as follows: S51. Let the set of player historical behavior data be... ,in, For historical state variables, For historical input vectors, As a signal of historical rewards, This represents the number of historical samples. S52. Aggregate the historical state variables, historical input vectors, and historical reward signals to obtain the state mean vector. Input mean vector and average reward ; S53, Generate player profile vector : ； in, , It is a mapping matrix.

9. The AI-based cross-game experience migration system according to claim 1, characterized in that, The steps for inputting cross-game universal feature representations and player profile vectors into the target game motion mapper, and then outputting target game control commands through the target game motion mapper, are as follows: S61, Vectorize the player profile Cross-game universal feature representation The vector is obtained by concatenating the two vectors. ; S62, merge vectors Input the target game action mapper to obtain the action prediction probability distribution. , ,in, The target game action mapper uses a fully connected neural network to represent the number of target game action categories. S63. By selecting the action category with the highest probability as the target game step... Control commands; S64. Input the control command into the target game interface for execution.

Images