Player emotion perception and experience optimization system based on artificial intelligence

Abstract

The invention discloses a player emotion perception and experience optimization system based on artificial intelligence, and the system comprises an emotion data collection module which is used for collecting multi-modal emotion data of a player in a game process and carrying out the preprocessing of the multi-modal emotion data; the emotion recognition module is used for acquiring the current emotion category and emotion intensity of the player according to the multi-modal data; the experience optimization decision module is used for generating game experience optimization actions matched with the current emotion category and emotion intensity of the player; and the feedback adjustment module is used for adjusting the game environment parameters according to the optimization action. According to the method, the multi-modal emotion recognition and reinforcement learning decision technology is combined, fusion perception of the face image and the voice signal of the player is achieved, the optimal experience adjustment strategy is generated, the method has the advantages of being high in accuracy and high in personalized adaptation ability, and the immersion, interactivity and stability of a game system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and player interaction, and in particular to a player emotion perception and experience optimization system based on artificial intelligence. Background Technology

[0002] Currently, with the continuous development of artificial intelligence technology, its application in the field of entertainment interaction, especially in video games, is becoming increasingly widespread. Existing game experience optimization mainly relies on player behavior data, such as completion time, operation frequency, and score, to make indirect inferences. Some studies have attempted to introduce emotion recognition, but most of them rely on a single modality of data source, such as facial expression recognition or voice timbre analysis to judge the player's emotions. However, the accuracy of single-modality emotion recognition is limited and is easily affected by external environmental factors such as lighting, noise, and occlusion, resulting in unstable results. In addition, existing emotion recognition technologies often lack real-time linkage with game mechanics. Even if a player's emotional state can be detected, it is difficult to immediately feed the results back into the game to adjust the experience, thus failing to truly achieve dynamic and personalized game adaptation.

[0003] Chinese invention patent application with publication number CN119541552A and publication date of February 28, 2025 discloses a method for complex emotion representation and perception based on emotion vectors. By combining multiple emotions into emotion vectors, each vector is composed of several meta-emotional elements, and the element values ​​are used to represent the intensity of the corresponding emotion, thereby realizing the joint expression of two or more emotions.

[0004] In the above invention applications, the emotion recognition model relies on single-modal data and lacks a multimodal fusion mechanism, which leads to a bias in the perception of the player's true emotional state. Secondly, there is a lack of dynamic decision-making methods based on reinforcement learning. Existing optimization measures often remain at the level of simple adjustments based on experience and fixed thresholds, and cannot form a closed-loop adaptive optimization process.

[0005] Therefore, there is an urgent need for a technical solution that can integrate multimodal data, achieve high-precision emotion recognition, and combine intelligent decision-making models to achieve dynamic experience optimization. Only by building an AI-based emotion perception and game feedback closed-loop system can we ensure recognition accuracy while achieving real-time perception and differentiated control of individual players, thereby significantly improving the immersion of the game and the overall experience of the player. Summary of the Invention

[0006] One objective of this invention is to propose an AI-based player emotion perception and experience optimization system. This invention constructs a closed-loop architecture consisting of an emotion data acquisition module, an emotion recognition module, an experience optimization decision-making module, and a feedback adjustment module. It employs a multimodal data acquisition and fusion mechanism, combined with the EmotionFormer model, to achieve high-precision emotion recognition from player facial images and voice signals. A deep Q-network is used for strategy decision-making, dynamically linking the player's emotional state with game environment parameters to output the optimal experience optimization actions. The feedback adjustment module then adjusts game difficulty, enemy numbers, prompt frequency, task pace, and story interaction methods in real time to achieve a personalized, immersive interactive experience. This method effectively solves the technical bottlenecks of existing technologies, such as emotion recognition relying on a single modality, fixed game optimization rules, and a lack of real-time and personalized adjustment. It significantly improves the system's emotion perception accuracy, experience adaptability, and player satisfaction, and is applicable to intelligent experience optimization in various game scenarios.

[0007] According to an embodiment of the present invention, an artificial intelligence-based player emotion perception and experience optimization system includes: The emotion data acquisition module is used to collect multimodal emotion data of players during the game and to preprocess the collected multimodal emotion data; The emotion recognition module is used to input preprocessed multimodal data into the pre-trained multimodal emotion recognition model EmotionFormer to obtain the player's current emotion category and emotion intensity. The experience optimization decision module is used to construct a state vector by combining the emotion category and emotion intensity with the current game environment state parameters, and input it into a strategy learning model based on a deep Q network to generate game experience optimization actions that match the player's current emotion category and emotion intensity. The feedback adjustment module is used to adjust game environment parameters based on the optimized actions, including game difficulty, number of enemies, prompt frequency, mission pace, and story interaction methods.

[0008] Optionally, the steps for collecting multimodal emotional data of players during the game and preprocessing the collected multimodal emotional data are as follows: S11. Configure camera and microphone acquisition parameters to capture the player's facial video through the camera and simultaneously capture the player's voice through the microphone; S12. Denoise the face video frame sequence to generate a standardized video block sequence; and simultaneously segment the discrete speech sequence according to timestamps to generate a standardized speech feature sequence.

[0009] Optionally, the EmotionFormer includes an input layer, a single-modal feature extraction layer, a cross-modal fusion layer, a global average pooling layer, and an output layer; The input layer receives visual modality token sequences and audio modality token sequences; The single-modal feature extraction layers extract single-modal features through independent multi-layer Transformer encoders. Each layer includes a multi-head self-attention mechanism, a feedforward fully connected network, residual connections, and layer normalization. The visual encoder has the following number of layers: The number of audio encoder layers is ; The cross-modal fusion layer concatenates the feature sequences encoded by a single modality in the modal dimension and uses a multi-head interactive attention mechanism to calculate the attention weight matrix from vision to audio and the attention weight matrix from audio to vision. Global average pooling layers are used to compress the fused multimodal features into vector representations. ; The output layer includes classification and regression branches, representing vectors. The data are input to the classification branch and the regression branch, respectively. The classification branch outputs the probability distribution of the emotion category through a fully connected layer and a softmax function, while the regression branch outputs the predicted value of the emotion intensity through a fully connected layer and a sigmoid function.

[0010] Optionally, the EmotionFormer is pre-trained before performing emotion recognition.

[0011] Optionally, the pre-training steps for EmotionFormer before emotion recognition are as follows: S21. Obtain the manually annotated dataset and divide it into a training set and a validation set; S22. Using a loss function Training is performed, and the loss function is... The calculation method is as follows: ; in, For classification loss function, For regression loss function, As weight; The classification loss function is calculated as follows: ; in, For the sample size, No. The true emotion category of each sample For the first The emotion category vector corresponding to each sample For normalization function, The calculation method is as follows: ; in, It is a natural exponential function. For the first Each sample in the true emotion category The logit component is on the , where C is the total number of all categories. For the first One sample in the emotion category The logit component on; The regression loss function is calculated as follows: ; in, For the sample size, To predict the intensity of emotions, This represents the true intensity of emotion.

[0012] S23, Using a learning rate of Iterative training of the Adam optimization algorithm In each training round, the validation loss is calculated on the validation set at the end of each round, and the set of parameters that minimizes the validation loss is obtained.

[0013] Optionally, the steps for inputting the preprocessed multimodal data into the pre-trained multimodal emotion recognition model EmotionFormer to obtain the player's current emotion category and intensity are as follows: S31. Transform the standardized video block sequence into a visual token sequence through a linear mapping. Simultaneously, the standardized speech feature sequence is transformed into an audio token sequence through a linear mapping. ; S32, which are visual token sequences respectively. With audio token sequence Add intramodal positional encoding to generate the visual input matrix. and audio input matrix ; S33, Input visual matrix With audio input matrix The input is fed into the EmotionFormer's single-modal feature extraction layer. Each single-modal feature extraction layer extracts features through independent multi-layer Transformer encoders. The visual encoder has [number] layers. The number of audio encoder layers is Each layer includes a multi-head self-attention mechanism, a feedforward fully connected network, residual connections, and layer normalization. Visual and audio features are extracted through each Transformer encoder layer in the single-modal feature extraction layer to generate a visual feature matrix. With audio feature matrix ; S34, Visual feature matrix With audio feature matrix The input is fed into the cross-modal fusion layer of EmotionFormer, which calculates the attention weight matrix from visual to audio through an interactive attention mechanism. And the attention weight matrix from audio to vision : ; ; in, This is the weight matrix. The dimension of the key vector. This is the normalization function; S35, Utilizing the visual-to-audio attention weight matrix And the attention weight matrix from audio to vision Calculate the visual feature matrix after cross-modal interaction With audio feature matrix : ; ; in, For visual feature matrix, For audio feature matrix, This is the weight matrix; S36. The visual feature matrix after cross-modal interaction With audio feature matrix The modality dimension is concatenated, and then global average pooling is used to obtain the sentiment feature vector. ; S37. Transfer the emotion feature vector Input the classification branch and the regression branch respectively; The classification branch outputs the probability distribution of the emotion category. : ; in, This is the weight matrix. For bias, This is the normalization function; The regression branch outputs the predicted value of sentiment intensity. : ; in, For normalization function, This is the weight matrix. For bias.

[0014] Optionally, the single-modal feature extraction layer extracts single-modal features from the visual input matrix and the audio input matrix using two independent encoders, respectively. The visual encoder is composed of... The audio encoder consists of layered Transformer encoders. It consists of a layer Transformer encoder.

[0015] Optionally, the steps for constructing a state vector by combining the emotion category and intensity with the current game environment state parameters, inputting it into a deep Q-network-based policy learning model, and generating game experience optimization actions that match the player's current emotion category and intensity are as follows: S41, Combining the probability distribution of emotion categories Compared with the predicted value of emotion intensity Constructing emotional state vectors , ; S42. Obtain the current game environment state vector. , ,in, Indicates the game's... One adjustable parameter value This represents the total number of parameters, and the parameter types include game difficulty level, number of enemies, prompt frequency, mission pace, and story interaction methods. S43. Convert the emotion state vector With the game environment state vector The vectors are concatenated to form the state vector for reinforcement learning. , ,in, Index for the current time step; S44. Transfer the state vector The input is fed into a policy learning model based on a deep Q-network, wherein the policy learning model has a parameter of A multilayer perceptron consists of an input layer, several hidden layers, and an output layer. The number of units in the output layer is equal to the size of the action space. The output layer outputs the Q-value vectors corresponding to each action. ,in, , For the set of all adjustable movements, Indicates an action; S45, Based on the current state Select the optimal action : ; in, , Optimize the action set for a predefined game experience. Optimize action output to match the player's current emotional state in the game experience.

[0016] Optionally, the steps for adjusting game environment parameters based on optimized actions, including game difficulty, number of enemies, prompt frequency, mission pace, and story interaction methods, are as follows: S51, Based on the optimal action Obtain a set of game parameter adjustment vectors, and add the current game parameter vector to the adjustment vectors element by element to obtain the updated parameter vector; S52. Write the updated parameter vector into the real-time configuration interface of the game engine and make real-time adjustments within the current game cycle.

[0017] The beneficial effects of this invention are: 1. This invention introduces multimodal data acquisition and deep learning models into the process of player emotion perception and experience optimization, and combines them with the dynamic decision-making mechanism of reinforcement learning to form a complete closed-loop system of emotion perception-decision optimization-feedback adjustment. Compared with the prior art, this invention can significantly improve the accuracy and robustness of emotion recognition and overcome the defects of single-modal methods that are susceptible to external environmental interference.

[0018] 2. This invention combines the emotional state output by the EmotionFormer model with a deep Q-network to achieve real-time strategy generation based on individual player emotions. It no longer relies on preset fixed rules, but adjusts actions by adaptively learning to output the optimal game experience, thereby enhancing the system's flexibility and personalized adaptability.

[0019] 3. This invention applies optimized actions to the game engine in real time through a feedback adjustment module, enabling dynamic control over game difficulty, enemy numbers, prompt frequency, task pace, and story interaction methods. This allows emotion recognition results to be truly transformed into experience optimization measures, enhancing the player's immersive interaction with the game. Due to the system's continuous learning and real-time closed-loop characteristics, this invention can maintain high efficiency and adaptability to different players and game scenarios, significantly improving the problems of poor real-time performance, lack of personalization, and unstable optimization effects of existing methods. Attached Figure Description

[0020] 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 player emotion perception and experience optimization system based on artificial intelligence proposed in this invention; Figure 2This is a structural diagram of EmotionFormer, a multimodal emotion recognition model for the AI-based player emotion perception and experience optimization system proposed in this invention. Detailed Implementation

[0021] 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.

[0022] refer to Figure 1 and Figure 2 An AI-based player emotion perception and experience optimization system includes: The emotion data acquisition module is used to collect multimodal emotion data of players during the game and to preprocess the collected multimodal emotion data; The emotion recognition module is used to input preprocessed multimodal data into the pre-trained multimodal emotion recognition model EmotionFormer to obtain the player's current emotion category and emotion intensity. The experience optimization decision module is used to construct a state vector by combining the emotion category and emotion intensity with the current game environment state parameters, and input it into a strategy learning model based on a deep Q network to generate game experience optimization actions that match the player's current emotion category and emotion intensity. The feedback adjustment module is used to adjust game environment parameters based on the optimized actions, including game difficulty, number of enemies, prompt frequency, mission pace, and story interaction methods.

[0023] In this embodiment, the steps for collecting multimodal emotional data of players during the game and preprocessing the collected multimodal emotional data are as follows: S11. Configure camera and microphone acquisition parameters to capture the player's facial video through the camera and simultaneously capture the player's voice through the microphone; S12. Denoise the face video frame sequence to generate a standardized video block sequence; and simultaneously segment the discrete speech sequence according to timestamps to generate a standardized speech feature sequence.

[0024] In this embodiment, the EmotionFormer includes an input layer, a single-modal feature extraction layer, a cross-modal fusion layer, a global average pooling layer, and an output layer. The input layer receives visual modality token sequences and audio modality token sequences; The single-modal feature extraction layers extract single-modal features through independent multi-layer Transformer encoders. Each layer includes a multi-head self-attention mechanism, a feedforward fully connected network, residual connections, and layer normalization. The visual encoder has the following number of layers: The number of audio encoder layers is ; The cross-modal fusion layer concatenates the feature sequences encoded by a single modality in the modal dimension and uses a multi-head interactive attention mechanism to calculate the attention weight matrix from vision to audio and the attention weight matrix from audio to vision. Global average pooling layers are used to compress the fused multimodal features into vector representations. ; The output layer includes classification and regression branches, representing vectors. The data are input to the classification branch and the regression branch, respectively. The classification branch outputs the probability distribution of the emotion category through a fully connected layer and a softmax function, while the regression branch outputs the predicted value of the emotion intensity through a fully connected layer and a sigmoid function.

[0025] In this embodiment, EmotionFormer is pre-trained before performing emotion recognition.

[0026] In this embodiment, the pre-training steps of EmotionFormer before emotion recognition are as follows: S21. Obtain the manually annotated dataset and divide it into a training set and a validation set; S22. Using a loss function Training is performed, and the loss function is... The calculation method is as follows: ; in, For classification loss function, For regression loss function, As weight; The classification loss function is calculated as follows: ; in, For the sample size, No. The true emotion category of each sample For the first The emotion category vector corresponding to each sample For normalization function, The calculation method is as follows: ; in, It is a natural exponential function. For the first Each sample in the true emotion category The logit component is on the , where C is the total number of all categories. For the first One sample in the emotion category The logit component on; The regression loss function is calculated as follows: ; in, For the sample size, To predict the intensity of emotions, This represents the true intensity of emotion.

[0027] S23, Using a learning rate of Iterative training of the Adam optimization algorithm In each training round, the validation loss is calculated on the validation set at the end of each round, and the set of parameters that minimizes the validation loss is obtained.

[0028] In this embodiment, the steps for inputting the preprocessed multimodal data into the pre-trained multimodal emotion recognition model EmotionFormer to obtain the player's current emotion category and intensity are as follows: S31. Transform the standardized video block sequence into a visual token sequence through a linear mapping. Simultaneously, the standardized speech feature sequence is transformed into an audio token sequence through a linear mapping. ; S32, which are visual token sequences respectively. With audio token sequence Add intramodal positional encoding to generate the visual input matrix. and audio input matrix ; S33, Input visual matrix With audio input matrix The input is fed into the EmotionFormer's single-modal feature extraction layer. Each single-modal feature extraction layer extracts features through independent multi-layer Transformer encoders. The visual encoder has [number] layers. The number of audio encoder layers is Each layer includes a multi-head self-attention mechanism, a feedforward fully connected network, residual connections, and layer normalization. Visual and audio features are extracted through each Transformer encoder layer in the single-modal feature extraction layer to generate a visual feature matrix. With audio feature matrix ; S34, Visual feature matrix With audio feature matrix The input is fed into the cross-modal fusion layer of EmotionFormer, which calculates the attention weight matrix from visual to audio through an interactive attention mechanism. And the attention weight matrix from audio to vision : ; ; in, This is the weight matrix. The dimension of the key vector. This is the normalization function; S35, Utilizing the visual-to-audio attention weight matrix And the attention weight matrix from audio to vision Calculate the visual feature matrix after cross-modal interaction With audio feature matrix : ; ; in, For visual feature matrix, For audio feature matrix, This is the weight matrix; S36. The visual feature matrix after cross-modal interaction With audio feature matrix The modality dimension is concatenated, and then global average pooling is used to obtain the sentiment feature vector. ; S37. Transfer the emotion feature vector Input the classification branch and the regression branch respectively; The classification branch outputs the probability distribution of the emotion category. : ; in, This is the weight matrix. For bias, This is the normalization function; The regression branch outputs the predicted value of sentiment intensity. : ; in, For normalization function, This is the weight matrix. For bias.

[0029] In this embodiment, the single-modal feature extraction layer extracts single-modal features from the visual input matrix and the audio input matrix using two independent encoders, respectively. The visual encoder is composed of... The audio encoder consists of layered Transformer encoders. It consists of a layer Transformer encoder.

[0030] In this embodiment, the steps for constructing a state vector by combining the emotion category and emotion intensity with the current game environment state parameters, inputting it into a deep Q-network-based policy learning model, and generating game experience optimization actions that match the player's current emotion category and emotion intensity are as follows: S41, Combining the probability distribution of emotion categories Compared with the predicted value of emotion intensity Constructing emotional state vectors , ; S42. Obtain the current game environment state vector. , ,in, Indicates the game's... One adjustable parameter value This represents the total number of parameters, and the parameter types include game difficulty level, number of enemies, prompt frequency, mission pace, and story interaction methods. S43. Convert the emotion state vector With the game environment state vector The vectors are concatenated to form the state vector for reinforcement learning. , ,in, Index for the current time step; S44. Transfer the state vector The input is fed into a policy learning model based on a deep Q-network, wherein the policy learning model has a parameter of A multilayer perceptron consists of an input layer, several hidden layers, and an output layer. The number of units in the output layer is equal to the size of the action space. The output layer outputs the Q-value vectors corresponding to each action. ,in, , For the set of all adjustable movements, Indicates an action; S45, Based on the current state Select the optimal action : ; in, , Optimize the action set for a predefined game experience. Optimize action output to match the player's current emotional state in the game experience.

[0031] In this embodiment, the steps for adjusting game environment parameters based on optimized actions, including game difficulty, number of enemies, prompt frequency, mission pace, and story interaction methods, are as follows: S51, Based on the optimal action Obtain a set of game parameter adjustment vectors, and add the current game parameter vector to the adjustment vectors element by element to obtain the updated parameter vector; S52. Write the updated parameter vector into the real-time configuration interface of the game engine and make real-time adjustments within the current game cycle.

[0032] 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 player emotion perception and experience optimization system, characterized in that, include: The emotion data acquisition module is used to collect multimodal emotion data of players during the game and to preprocess the collected multimodal emotion data; The emotion recognition module is used to input preprocessed multimodal data into the pre-trained multimodal emotion recognition model EmotionFormer to obtain the player's current emotion category and emotion intensity. The experience optimization decision module is used to construct a state vector by combining the emotion category and emotion intensity with the current game environment state parameters, and input it into a strategy learning model based on a deep Q network to generate game experience optimization actions that match the player's current emotion category and emotion intensity. The feedback adjustment module is used to adjust game environment parameters based on the optimized actions, including game difficulty, number of enemies, prompt frequency, mission pace, and story interaction methods.

2. The AI-based player emotion perception and experience optimization system according to claim 1, characterized in that, The steps for collecting multimodal emotional data from players during gameplay and preprocessing the collected multimodal emotional data are as follows: S11. Configure camera and microphone acquisition parameters to capture the player's facial video through the camera and simultaneously capture the player's voice through the microphone; S12. Denoise the face video frame sequence to generate a standardized video block sequence; and simultaneously segment the discrete speech sequence according to timestamps to generate a standardized speech feature sequence.

3. The AI-based player emotion perception and experience optimization system according to claim 1, characterized in that, The EmotionFormer includes an input layer, a single-modal feature extraction layer, a cross-modal fusion layer, a global average pooling layer, and an output layer.

4. The AI-based player emotion perception and experience optimization system according to claim 1, characterized in that, The EmotionFormer is pre-trained before performing emotion recognition.

5. The AI-based player emotion perception and experience optimization system according to claim 4, characterized in that, The pre-training steps for EmotionFormer before emotion recognition are as follows: S21. Obtain the manually annotated dataset and divide it into a training set and a validation set; S22. Using a loss function Training is performed, and the loss function is... The calculation method is as follows: ； in, For classification loss function, For regression loss function, As weight; The classification loss function is calculated as follows: ； in, For the sample size, No. The true emotion category of each sample For the first The emotion category vector corresponding to each sample This is the normalization function; The regression loss function is calculated as follows: ； in, For the sample size, To predict the intensity of emotions, This represents the true intensity of emotion. S23, Using a learning rate of Iterative training of the Adam optimization algorithm In each training round, the validation loss is calculated on the validation set at the end of each round, and the set of parameters that minimizes the validation loss is obtained.

6. The AI-based player emotion perception and experience optimization system according to claim 1, characterized in that, The steps for inputting the preprocessed multimodal data into the pre-trained multimodal emotion recognition model EmotionFormer to obtain the player's current emotion category and intensity are as follows: S31. Transform the standardized video block sequence into a visual token sequence through a linear mapping. Simultaneously, the standardized speech feature sequence is transformed into an audio token sequence through a linear mapping. ; S32, which are visual token sequences respectively. With audio token sequence Add intramodal positional encoding to generate the visual input matrix. and audio input matrix ; S33, Input visual matrix With audio input matrix The input is fed into the EmotionFormer's unimodal feature extraction layer. Visual and audio features are extracted through each Transformer encoder layer in the unimodal feature extraction layer, generating a visual feature matrix. With audio feature matrix ; S34, Visual feature matrix With audio feature matrix The input is fed into the cross-modal fusion layer of EmotionFormer, which calculates the attention weight matrix from visual to audio through an interactive attention mechanism. And the attention weight matrix from audio to vision : ； ； in, This is the weight matrix. The dimension of the key vector. This is the normalization function; S35, Utilizing the visual-to-audio attention weight matrix And the attention weight matrix from audio to vision Calculate the visual feature matrix after cross-modal interaction With audio feature matrix : ； ； in, For visual feature matrix, For audio feature matrix, This is the weight matrix; S36. The visual feature matrix after cross-modal interaction With audio feature matrix The modality dimension is concatenated, and then global average pooling is used to obtain the sentiment feature vector. ; S37. Transfer the emotion feature vector Input the classification branch and the regression branch respectively; The classification branch outputs the probability distribution of the emotion category. : ； in, This is the weight matrix. For bias, This is the normalization function; The regression branch outputs the predicted value of sentiment intensity. : ； in, For normalization function, This is the weight matrix. For bias.

7. The AI-based player emotion perception and experience optimization system according to claim 6, characterized in that, The single-modal feature extraction layer extracts single-modal features from the visual input matrix and the audio input matrix using two independent encoders. The visual encoder is composed of... The audio encoder consists of layered Transformer encoders. It consists of a layer Transformer encoder.

8. The AI-based player emotion perception and experience optimization system according to claim 1, characterized in that, The steps for constructing a state vector by combining the emotion category and intensity with the current game environment state parameters, and inputting it into a deep Q-network-based policy learning model to generate game experience optimization actions that match the player's current emotion category and intensity are as follows: S41, Combining the probability distribution of emotion categories Compared with the predicted value of emotion intensity Constructing emotional state vectors ; S42. Obtain the current game environment state vector. , ,in, Indicates the game's... One adjustable parameter value This represents the total number of parameters. S43. Convert the emotion state vector With the game environment state vector The vectors are concatenated to form the state vector for reinforcement learning. , ,in, Index for the current time step; S44. Transfer the state vector The input is fed into a policy learning model based on a deep Q-network, wherein the policy learning model has a parameter of A multilayer perceptron consists of an input layer, several hidden layers, and an output layer. The number of units in the output layer is equal to the size of the action space. The output layer outputs the Q-value vectors corresponding to each action. ,in, , For the set of all adjustable movements, Indicates an action; S45, Based on the current state Select the optimal action : ； in, , Optimize the action set for a predefined game experience. Optimize action output to match the player's current emotional state in the game experience.

9. The AI-based player emotion perception and experience optimization system according to claim 1, characterized in that, The steps for adjusting game environment parameters based on optimized actions, including game difficulty, number of enemies, prompt frequency, mission pace, and story interaction methods, are as follows: S51, Based on the optimal action Obtain a set of game parameter adjustment vectors, and add the current game parameter vector to the adjustment vectors element by element to obtain the updated parameter vector; S52. Write the updated parameter vector into the real-time configuration interface of the game engine and make real-time adjustments within the current game cycle.

Images