Mobile game complex scene self-adaptive operation method based on multi-task learning

Abstract

The invention discloses a mobile game complex scene self-adaptive operation method based on multi-task learning, which comprises the following steps: S1, collecting and preprocessing data in a mobile game complex scene to generate an original multi-modal data set; s2, constructing a multi-modal feature extraction network, extracting features and generating a task graph; s3, performing enhancement processing on the features, and generating multi-modal features through feature fusion; s4, constructing a multi-task learning model, optimizing task features and generating task priorities; s5, performing unified optimization by using a recursive optimization method to generate a global task optimization result; and S6, in combination with parameters in a simulation environment, optimizing a global task strategy and a local scene strategy through a reinforcement learning model based on an Actor-Critic architecture. According to the method, reinforcement learning and multi-task learning are combined, the operation strategy of the complex scene of the mobile game is optimized, and the method has the advantages of being high in intelligent adaptability, efficient in task coordination and capable of saving computing resources.

Description

Technical Field

[0001] The present invention relates to the field of complex scene operation of mobile games, and in particular to a complex scene adaptive operation method of mobile games based on multi-task learning. Background Art

[0002] With the rapid development of mobile Internet, mobile games have become an important part of modern people's daily entertainment life. Especially today, as the hardware performance of smartphones continues to improve, the types and complexity of mobile games are becoming more diverse. Many mobile games not only rely on the player's operating skills, but also involve complex game scenes, task allocation, real-time feedback and strategy adjustment. In order to provide a more intelligent and personalized gaming experience, in recent years, more and more research has begun to explore how to use technologies such as artificial intelligence and deep learning to enhance the adaptive operation capabilities in games.

[0003] Traditional mobile game control methods often rely on players' intuitive operations, and complex scenes and tasks in games are usually handled by fixed rules and manually set control methods. For example, the actions of characters or objects in the game are often predetermined or implemented through simple rules. Although this method is simple and effective, it has poor adaptability and flexibility for complex and changing game environments. In addition, as the complexity of mobile game scenes continues to increase, players not only need to consider real-time visual information when operating, but also need to make decisions based on the ever-changing task requirements and game rules, which makes traditional operation methods gradually appear inadequate.

[0004] In order to solve this problem, some researchers have begun to try to introduce advanced technologies such as deep learning and reinforcement learning, and use big data analysis and intelligent algorithms to automatically optimize operational behaviors. Deep learning can imitate the human decision-making process to a certain extent through training on massive data, and make autonomous adjustments according to changes in the environment. Reinforcement learning is particularly suitable for dealing with decision-making problems. By exploring the state space and action space, strategies can be continuously optimized. However, existing deep learning and reinforcement learning applications often have some limitations. First, many applications focus on the optimization of a single task, and the coordinated optimization of multiple tasks and multiple goals in complex scenarios is insufficient. Second, existing methods usually rely on single-modal data for learning and decision-making. This method often ignores the synergy of multimodal information in game operations (such as visual information, operation records, and game rules, etc.), and it is difficult to comprehensively improve the intelligence and adaptability of game operations.

[0005] At present, most of the research on adaptive operation methods of mobile games based on deep learning and reinforcement learning focuses on a specific task or a specific type of operation scenario. For example, some studies use image processing technology to analyze game screens and perform automatic aiming or obstacle avoidance operations by detecting the position of enemies or target objects. Other studies focus on training agents through reinforcement learning to optimize strategies in a single task or a relatively simple environment. Although these methods can improve the intelligence level of game operation to a certain extent, they often face the following problems: First, complex mobile game scenarios not only contain multiple target tasks, but also involve dependencies and mutual constraints between tasks. Therefore, how to effectively optimize tasks in complex task dependency graphs is an urgent problem to be solved. Second, existing technologies usually only learn and predict through a single modality (such as visual information or operation records), lack of effective fusion and optimization of multimodal information, resulting in poor adaptability of the system in complex environments. Third, although reinforcement learning can help agents adjust strategies based on feedback, the existing methods are still too simple in resource allocation and task priority adjustment in multi-task scenarios, and cannot dynamically respond to real-time changes in complex scenarios.

[0006] In addition, most traditional deep learning and reinforcement learning methods use fixed loss functions and optimization strategies, lacking a dynamic adjustment mechanism for weights in different tasks and scenarios. In complex mobile game scenarios, there is often dependency, competition, or even mutual exclusion between tasks. How to reasonably adjust the loss weights between tasks, optimize resource allocation, and dynamically adjust task priorities is the key to improving game operation performance and intelligence. However, existing technologies have not fully addressed this problem. Most existing reinforcement learning methods ignore this complex relationship between tasks, resulting in the system being unable to dynamically adjust resource allocation and operation strategies according to the urgency, importance, or executability of the task, making it difficult to achieve flexible and effective adaptive operations in complex multi-task environments.

[0007] Therefore, in view of the limitations of current technology, a new method is urgently needed that can comprehensively consider multiple information sources such as the relationship between tasks, visual information, operation records and game rules under the framework of multi-task learning, and optimize the operation strategy and task execution in the game. Especially when there are complex dependencies, mutual exclusion and parallel relationships between tasks, how to dynamically adjust task priorities, resource allocation and operation strategies through intelligent algorithms is the key to improving the level of intelligent operation of mobile games. Summary of the invention

[0008] One purpose of the present invention is to propose an adaptive operation method for complex scenes of mobile games based on multi-task learning. The present invention combines multimodal data fusion, multi-task learning and reinforcement learning technology to optimize the operation strategy in complex mobile game scenes. By accurately extracting operation records, visual information and game rules, a task graph is generated and task priorities are dynamically adjusted, thereby improving the efficiency of task coordination and resource allocation.

[0009] According to an embodiment of the present invention, a method for adaptively operating complex scenes of mobile games based on multi-task learning comprises the following steps:

[0010] S1. Collect data from complex scenes of mobile games, extract operation records, visual information and game rules, and preprocess the data to generate original multimodal data sets;

[0011] S2. Based on the original multimodal dataset, a multimodal feature extraction network is constructed to extract visual features, operation behavior features and rule features, and a task graph is generated through the spatial relationship in the visual features, the task constraints in the rule features and the operation logic in the operation behavior features;

[0012] S3. Enhance the visual features, operation behavior features and rule features, use the improved contrast learning method to optimize the expression of operation behavior features, and combine the multi-view fusion technology and embedded coding method to optimize the visual features and rule features respectively, and perform feature fusion on the optimized features to generate multimodal features;

[0013] S4, taking multimodal features and task graphs as input, building a multi-task learning model, optimizing task features, adjusting task loss weights in combination with a dynamic weight allocation mechanism, and generating task priorities;

[0014] S5. Based on the task priority and the optimized task characteristics, the tasks with dependencies in the task graph are uniformly optimized using a recursive optimization method to generate a global task optimization result.

[0015] S6. Taking the global task optimization results as input, combined with the lighting parameters, obstacle parameters and delay parameters in the simulated environment, the global task strategy and local scene strategy are optimized through a reinforcement learning model based on the Actor-Critic architecture.

[0016] Optionally, the S2 specifically includes:

[0017] S21, constructing a multimodal feature extraction network, wherein the multimodal extraction network is composed of a bidirectional long short-term memory network, a ResNet34 network and a multi-layer perceptron;

[0018] S22. Use a bidirectional long short-term memory network to extract operation records and generate operation behavior features:

[0019] F op =f LSTM (X op );

[0020] Among them, F op Indicates the characteristics of the operation behavior, X op Indicates operation record, f LSTM represents a bidirectional long short-term memory network;

[0021] S23. For visual information, the ResNet34 network is used to extract spatial features in the image and generate visual features:

[0022] F vis =f CNN (X vis );

[0023] Among them, F vis represents visual features, X vis represents visual information, f CNN Represents the ResNet34 network;

[0024] S24. For the game rules, the rule logic is embedded through a multi-layer perceptron to generate rule features:

[0025] F rule =f MLP (X rule );

[0026] Among them, F rule represents the rule feature, X rule represents the rules of the game, f MLP represents a multi-layer perceptron;

[0027] S25. Analyze the spatial relationship in the visual features, the task constraints in the rule features, and the operation logic in the operation behavior features to generate a task graph, which describes the dependency, parallel relationship, and mutual exclusion relationship between tasks.

[0028] Optionally, the S3 specifically includes:

[0029] S31. Use an improved contrastive learning method to optimize the expression of operational behavior characteristics;

[0030] S32. Use multi-view fusion technology to optimize visual features, construct visual features for each view, perform weighted fusion on the features of each view, and generate optimized visual features:

[0031]

[0032] in, represents the optimized visual features, α m represents the weight of the mth perspective, represents the visual features of the mth perspective, M represents the number of perspectives, W represents the trainable parameters, represents the visual features of the j-th perspective, exp represents the natural exponential function;

[0033] S33, processing the rule features, optimizing the rule features using an embedded coding method, and generating optimized rule features;

[0034] S34. Perform feature fusion on the optimized features, use the self-attention mechanism to calculate the importance weight of the features, and calculate the weight matrix of each feature:

[0035]

[0036] Among them, W i represents the weight matrix of feature i, Q i and K i They represent the query and key mapping of the feature vector, d represents the feature dimension, T represents the transposition operation, and softmax represents the normalization function;

[0037] S36. Generate multimodal features:

[0038]

[0039] in, represents multimodal features, V i Represents the value mapping of the eigenvector, and n represents the number of modes.

[0040] Optionally, the S31 specifically includes:

[0041] S311, based on operational behavior characteristics F op , construct contrastive learning sample pairs, the contrastive learning sample pairs include positive sample pairs And negative sample pairs And select the hard negative samples that are most similar to the positive sample features from the negative sample set to form a negative sample index set H(i);

[0042] S312, using the main encoder and momentum encoder to analyze the operation behavior feature F op Perform feature projection and generate a stable feature representation Z by optimizing the weights of the momentum encoder op :

[0043] θ′=mθ′+(1-m)θ;

[0044] Where θ' represents the momentum encoder weight, θ represents the main encoder weight, and m represents the momentum coefficient;

[0045] S313. Express Z based on stable features op Calculate the symmetric contrast loss for positive and negative sample pairs, including forward similarity optimization and reverse similarity optimization:

[0046]

[0047] Among them, L sym represents the symmetric contrast loss of positive and negative sample pairs, L op represents the one-way contrast loss, represents the positive sample feature representation generated by the momentum encoder, represents the reverse positive sample feature representation generated by the momentum encoder, represents the negative sample feature representation generated by the momentum encoder, sim represents the cosine similarity function, τ represents the temperature coefficient, H(i) represents the negative sample index set, and represents the feature representation generated by the momentum encoder;

[0048] S314, operation behavior characteristics F op Perform multi-scale feature extraction to generate operational behavior feature representations at different scales Calculate the symmetric contrast loss of each scale separately and sum them:

[0049]

[0050] Among them, L multi Multi-scale contrastive learning total loss, represents the characteristic representation of the operational behavior at the s-th scale, and P represents the total number of characteristic scales;

[0051] S315. Generate optimized operation behavior features by minimizing the total loss of multi-scale contrastive learning

[0052] Optionally, the S4 specifically includes:

[0053] S41, constructing a multi-task learning model, inputting multimodal features into a shared feature extraction network to extract shared features, the multi-task learning model includes a shared feature extraction network and a task branch network, the shared feature extraction network is composed of a multi-layer convolutional neural network, and the task branch network is composed of a shallow convolutional neural network;

[0054] S42, constructing an independent task branch network for each task according to the task nodes in the task graph, and extracting task features through the task branch network;

[0055] S43. According to the dependency, competition or mutual exclusion relationship between tasks in the task graph, the task features are adjusted through the relevance of shared features:

[0056]

[0057] R ij =1-Mutex(t i ,t j );

[0058] in, represents the task features adjusted by shared features, R ij represents the inhibitory factor, W p , W s and W k They represent the weighted parameters of the parent task, shared features, and subtasks, respectively, and are used to control the contribution of features to the adjustment. i represents the subtask set, P i represents the parent task set, F shared represents shared features, and Respectively represent the specific features of the parent task and the specific features of the subtask, Mutex(t i ,t j ) represents task t i and t j Mutually exclusive relationship;

[0059] S43, calculating the immediate utility for each task according to the task graph and the adjusted task characteristics;

[0060] S44. Dynamically adjust task loss weights based on immediate utility:

[0061]

[0062] Among them, w i represents the loss weight, ∈ represents the smoothing factor to prevent the denominator from being zero, and u i and u j Respectively represent tasks t i and t j Immediate utility;

[0063] S45. Extract the dependencies between tasks according to the task graph, build a dependency chain for each task, recursively optimize the dependency chain from parent tasks to child tasks, and update the features of the nested model:

[0064]

[0065] Among them, F nested (t i ) represents task t i The nested optimization feature, f nested represents a nested optimization function, and Indicates task characteristics;

[0066] S46. Generate task priorities based on immediate utility and dependencies between tasks.

[0067] Optionally, the S5 specifically includes:

[0068] S51. According to the task graph, extract the set of tasks that have a mutually exclusive relationship with the tasks, introduce mutually exclusive constraints on the nested features of the tasks, and adjust the task features through the mutually exclusive optimization function:

[0069]

[0070] Among them, f mutex represents the mutually exclusive optimization function, F nested (t i ) represents task t i The optimization feature, F nested (t k ) represents the optimization feature of mutually exclusive tasks, sim represents the cosine similarity function, λ represents the weight coefficient of the mutually exclusive constraint, which is used to control the strength of the penalty term, Mutex(t i ) represents a set of mutually exclusive tasks;

[0071] S52. According to the task graph, extract a set of tasks that can be executed in parallel, build a global consistency optimization target based on the nested features of the parallel task set, jointly optimize the features of the parallel task set, and update the task features:

[0072]

[0073] Among them, L parallel represents the global consistency optimization objective function, Parallel(t i ) represents a set of parallel tasks, F nested (t k ) represents the optimization features of mutually exclusive tasks, sim represents the cosine similarity function, F shared The shared features represent the information shared between tasks;

[0074] S53. Calculate the resource requirements of the task according to the task graph and task priority, dynamically adjust the resource allocation weight according to the task priority and resource requirements, and generate a global task optimization result.

[0075] Optionally, the S6 specifically includes:

[0076] S61. Build a reinforcement learning model based on the Actor-Critic architecture, taking the global task optimization results and environmental parameters as input to generate unified state features:

[0077] H ti=ReLU(W t ·F nested (t i )+b t );

[0078]

[0079] H env =ReLU(W e ·E env +b e );

[0080]

[0081] Among them, H ti represents the embedded task features, F nested (t i ) represents task t i The nested optimization feature, H global represents the global task characteristics, k represents the number of tasks, E env represents the set of environmental parameters, H env represents the global environment characteristics, S represents the unified state characteristics, and W t , W e and W f They represent the weights of task embedding, environment embedding and feature fusion, respectively, and b t 、b e and b f They represent the corresponding biases, μ and σ represent normalization parameters;

[0082] S62. Output strategy through Actor network:

[0083] π(a|s)=softmax(W actor ·S+b actor );

[0084] Among them, π(a|s) represents the output strategy, W actor and b actor Represent the weight and bias of the Actor network respectively;

[0085] S63. Generate state value function through critic network:

[0086] V(s)=W critic ·S+b critic ;

[0087] Among them, V(s) represents the state value function, W critic and b critic Respectively represent the weight and bias of the Critic network;

[0088] S64, optimize the global task strategy, calculate the advantage function through the immediate reward and cumulative return, update the Actor network parameters, and update the Critic network parameters by minimizing the Critic network loss function:

[0089]

[0090] A(s,a)=R t -V(s);

[0091]

[0092] Among them, R t represents the cumulative return, γ represents the discount factor, r t represents the immediate reward, A(s,a) represents the global advantage function, α represents the learning rate, θ actor Represents the Actor network parameters, Denotes the parameter θ actor The gradient of L critic represents the Critic network loss function, θ critic represents the Critic network parameters, Denotes the parameter θ critic The gradient of

[0093] S65. Extract the local state features of the environmental parameters and generate the local scene strategy by optimizing the local strategy loss function:

[0094]

[0095] π local (a|s)=softmax(W actor-local ·S local +b actor-local );

[0096] L local =-log2π local (a|s)·A local (s,a);

[0097] Among them, H L represents the illumination characteristics, H O represents the overall obstacle characteristics, H D represents the delay characteristic, μ and σ represent the normalization parameters, S local represents the local state feature, W f and b f Respectively represent the weight and bias of the fusion network, [H L ,H O ,H D ] indicates that H L , H O and HD For splicing, L local represents the local strategy loss function, A local (s,a) represents the local advantage function, π local (a|s) represents the local scene strategy, W actor-local and b actor-local Respectively represent the weight and bias of the Actor network under environmental parameters;

[0098] S66. Output the optimized global task strategy and local scene strategy.

[0099] The beneficial effects of the present invention are:

[0100] First of all, the present invention conducts in-depth mining and optimization of data of different modes by constructing a multimodal feature extraction network, and combining an improved comparative learning method and multi-view fusion technology. This method can not only effectively improve the expressive power of operation behavior features, visual features and rule features, but also better capture the details and complexity in the game scene. By fusing multimodal information, the system can understand the game environment more comprehensively and make more reasonable decisions. This intelligent fusion of multimodal data greatly improves the accuracy of game operations, can cope with game tasks of different types and complexities, and enables the system to adaptively adjust the operation strategy in various complex scenarios, thereby optimizing the player's gaming experience.

[0101] Secondly, the present invention innovatively introduces a multi-task learning framework based on task graphs, and optimizes the processing of task features through a dynamic weight allocation mechanism. The task graph can accurately describe the task relationships in the game, including dependencies, mutual exclusions, and parallel relationships between tasks. Through this mechanism, the system can not only reasonably allocate resources, but also optimize according to the priority, urgency, and resource requirements of the tasks. This means that the system can reasonably schedule between different tasks to ensure that critical tasks can be processed first without being interfered with by the execution of low-priority tasks. Dynamically adjusting task priorities and resource allocation makes operations in multi-task scenarios more efficient, reduces task conflicts and resource waste, and thus improves the overall work efficiency of the system.

[0102] In addition, the present invention adopts a reinforcement learning model based on the Actor-Critic architecture, which can not only optimize the global task strategy, but also optimize the strategy for environmental factors such as lighting, obstacles, and delays in local scenes. Through continuous learning and adjustment, the system can flexibly respond to changes in the current environment to ensure efficient and accurate operation in complex and changing game scenes. Unlike traditional static rules or predetermined strategies, the reinforcement learning model can adjust the strategy in real time according to the player's behavior and environmental changes to achieve true adaptive operation.

[0103] Finally, the present invention can greatly improve the intelligence and adaptability of game operations in practical applications, not only optimizing the execution efficiency of game tasks, but also enhancing the interactive experience of players. For example, in a scenario where multiple tasks are executed in parallel, the system can reasonably schedule operation tasks to avoid a decrease in operation efficiency due to conflicts between tasks; in a dynamically changing environment, the system can adjust the operation strategy in real time according to environmental changes to ensure that tasks can be executed smoothly. More importantly, the system can gradually optimize the task execution strategy based on the player's behavior and feedback from the game scene, and provide personalized operation plans, thereby enhancing the player's sense of immersion and the comfort of operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0105] Figure 1 This is a flow chart of a method for adaptively operating complex scenes of mobile games based on multi-task learning proposed by the present invention;

[0106] Figure 2 This is a structural diagram of a multimodal feature extraction network for a complex scene adaptive operation method for mobile games based on multi-task learning proposed in the present invention. DETAILED DESCRIPTION

[0107] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0108] refer to Figure 1 and Figure 2 , a complex scene adaptive operation method for mobile games based on multi-task learning, comprising the following steps:

[0109] S1. Collect data from complex scenes of mobile games, extract operation records, visual information and game rules, and preprocess the data to generate original multimodal data sets;

[0110] S2. Based on the original multimodal dataset, a multimodal feature extraction network is constructed to extract visual features, operation behavior features and rule features, and a task graph is generated through the spatial relationship in the visual features, the task constraints in the rule features and the operation logic in the operation behavior features;

[0111] S3. Enhance the visual features, operation behavior features and rule features, use the improved contrast learning method to optimize the expression of operation behavior features, and combine the multi-view fusion technology and embedded coding method to optimize the visual features and rule features respectively, and perform feature fusion on the optimized features to generate multimodal features;

[0112] S4, taking multimodal features and task graphs as input, building a multi-task learning model, optimizing task features, adjusting task loss weights in combination with a dynamic weight allocation mechanism, and generating task priorities;

[0113] S5. Based on the task priority and the optimized task characteristics, the tasks with dependencies in the task graph are uniformly optimized using a recursive optimization method to generate a global task optimization result.

[0114] S6. Taking the global task optimization results as input, combined with the lighting parameters, obstacle parameters and delay parameters in the simulated environment, the global task strategy and local scene strategy are optimized through a reinforcement learning model based on the Actor-Critic architecture.

[0115] In this implementation, S2 specifically includes:

[0116] S21, constructing a multimodal feature extraction network, wherein the multimodal extraction network is composed of a bidirectional long short-term memory network, a ResNet34 network and a multi-layer perceptron;

[0117] S22. Use a bidirectional long short-term memory network to extract operation records and generate operation behavior features:

[0118] F op =f LSTM (X op );

[0119] Among them, F op Indicates the characteristics of the operation behavior, X op Indicates operation record, f LSTM represents a bidirectional long short-term memory network;

[0120] S23. For visual information, the ResNet34 network is used to extract spatial features in the image and generate visual features:

[0121] F vis =f CNN (X vis );

[0122] Among them, F vis represents visual features, X vis represents visual information, f CNN Represents the ResNet34 network;

[0123] S24. For the game rules, the rule logic is embedded through a multi-layer perceptron to generate rule features:

[0124] F rule =f MLP (X rule );

[0125] Among them, F rule represents the rule feature, X rule represents the rules of the game, f MLP represents a multi-layer perceptron;

[0126] S25. Analyze the spatial relationship in the visual features, the task constraints in the rule features, and the operation logic in the operation behavior features to generate a task graph, which describes the dependency, parallel relationship, and mutual exclusion relationship between tasks.

[0127] In this implementation, S3 specifically includes:

[0128] S31. Use an improved contrastive learning method to optimize the expression of operational behavior characteristics;

[0129] S32. Use multi-view fusion technology to optimize visual features, construct visual features for each view, perform weighted fusion on the features of each view, and generate optimized visual features:

[0130]

[0131] in, represents the optimized visual features, α m represents the weight of the mth perspective, represents the visual features of the mth perspective, M represents the number of perspectives, W represents the trainable parameters, represents the visual features of the j-th perspective, exp represents the natural exponential function;

[0132] S33, processing the rule features, optimizing the rule features using an embedded coding method, and generating optimized rule features;

[0133] S34. Perform feature fusion on the optimized features, use the self-attention mechanism to calculate the importance weight of the features, and calculate the weight matrix of each feature:

[0134]

[0135] Among them, W i represents the weight matrix of feature i, Q i and K i They represent the query and key mapping of the feature vector, d represents the feature dimension, T represents the transposition operation, and softmax represents the normalization function;

[0136] S36. Generate multimodal features:

[0137]

[0138] in, represents multimodal features, V i Represents the value mapping of the eigenvector, and n represents the number of modes.

[0139] In this implementation manner, the S31 specifically includes:

[0140] S311, based on operational behavior characteristics F op , construct contrastive learning sample pairs, the contrastive learning sample pairs include positive sample pairs And negative sample pairs And select the hard negative samples that are most similar to the positive sample features from the negative sample set to form a negative sample index set H(i);

[0141] S312, using the main encoder and momentum encoder to analyze the operation behavior feature F op Perform feature projection and generate a stable feature representation Z by optimizing the weights of the momentum encoder op :

[0142] θ′=mθ′+(1-m)θ;

[0143] Where θ' represents the momentum encoder weight, θ represents the main encoder weight, and m represents the momentum coefficient;

[0144] S313. Express Z based on stable features op Calculate the symmetric contrast loss for positive and negative sample pairs, including forward similarity optimization and reverse similarity optimization:

[0145]

[0146] Among them, L sym represents the symmetric contrast loss of positive and negative sample pairs, L op represents the one-way contrast loss, represents the positive sample feature representation generated by the momentum encoder, represents the reverse positive sample feature representation generated by the momentum encoder, represents the negative sample feature representation generated by the momentum encoder, sim represents the cosine similarity function, τ represents the temperature coefficient, H(i) represents the negative sample index set, and represents the feature representation generated by the momentum encoder;

[0147] S314, operation behavior characteristics F opPerform multi-scale feature extraction to generate operational behavior feature representations at different scales Calculate the symmetric contrast loss of each scale separately and sum them:

[0148]

[0149] Among them, L multi Multi-scale contrastive learning total loss, represents the characteristic representation of the operational behavior at the s-th scale, and P represents the total number of characteristic scales;

[0150] S315. Generate optimized operation behavior features by minimizing the total loss of multi-scale contrastive learning

[0151] In this implementation, S4 specifically includes:

[0152] S41, constructing a multi-task learning model, inputting multimodal features into a shared feature extraction network to extract shared features, the multi-task learning model includes a shared feature extraction network and a task branch network, the shared feature extraction network is composed of a multi-layer convolutional neural network, and the task branch network is composed of a shallow convolutional neural network;

[0153] S42, constructing an independent task branch network for each task according to the task nodes in the task graph, and extracting task features through the task branch network;

[0154] S43. According to the dependency, competition or mutual exclusion relationship between tasks in the task graph, the task features are adjusted through the relevance of shared features:

[0155]

[0156] R ij =1-Mutex(t i ,t j );

[0157] in, represents the task features adjusted by shared features, R ij represents the inhibitory factor, W p , W s and W k They represent the weighted parameters of the parent task, shared features, and subtasks, respectively, and are used to control the contribution of features to the adjustment. i represents the subtask set, P i represents the parent task set, F shared represents shared features, and Respectively represent the specific features of the parent task and the specific features of the subtask, Mutex(t i ,tj ) represents task t i and t j Mutually exclusive relationship;

[0158] S43, calculating the immediate utility for each task according to the task graph and the adjusted task characteristics;

[0159] S44. Dynamically adjust task loss weights based on immediate utility:

[0160]

[0161] Among them, w i represents the loss weight, ∈ represents the smoothing factor to prevent the denominator from being zero, and u i and u j Respectively represent tasks t i and t j Immediate utility;

[0162] S45. Extract the dependencies between tasks according to the task graph, build a dependency chain for each task, recursively optimize the dependency chain from parent tasks to child tasks, and update the features of the nested model:

[0163]

[0164] Among them, F nested (t i ) represents task t i The nested optimization feature, f nested represents a nested optimization function, and F ti Indicates task characteristics;

[0165] S46. Generate task priorities based on immediate utility and dependencies between tasks.

[0166] In this implementation manner, S5 specifically includes:

[0167] S51. According to the task graph, extract the set of tasks that have a mutually exclusive relationship with the tasks, introduce mutually exclusive constraints on the nested features of the tasks, and adjust the task features through the mutually exclusive optimization function:

[0168]

[0169] Among them, f mutex represents the mutually exclusive optimization function, F nested (t i ) represents task t i The optimization feature, F nested (t k ) represents the optimization feature of mutually exclusive tasks, sim represents the cosine similarity function, λ represents the weight coefficient of the mutually exclusive constraint, which is used to control the strength of the penalty term, Mutex(ti ) represents a set of mutually exclusive tasks;

[0170] S52. According to the task graph, extract a set of tasks that can be executed in parallel, build a global consistency optimization target based on the nested features of the parallel task set, jointly optimize the features of the parallel task set, and update the task features:

[0171]

[0172] Among them, L parallel represents the global consistency optimization objective function, Parallel(t i ) represents a set of parallel tasks, F nested (t k ) represents the optimization features of mutually exclusive tasks, sim represents the cosine similarity function, F shared The shared features represent the information shared between tasks;

[0173] S53. Calculate the resource requirements of the task according to the task graph and task priority, dynamically adjust the resource allocation weight according to the task priority and resource requirements, and generate a global task optimization result.

[0174] In this implementation manner, S6 specifically includes:

[0175] S61. Build a reinforcement learning model based on the Actor-Critic architecture, taking the global task optimization results and environmental parameters as input to generate unified state features:

[0176] H ti =ReLU(W t ·F nested (t i )+b t );

[0177]

[0178] H env =ReLU(W e ·E env +b e );

[0179]

[0180] Among them, H ti represents the embedded task features, F nested (t i ) represents task t i The nested optimization feature, H global represents the global task characteristics, k represents the number of tasks, E env represents the set of environmental parameters, Henv represents the global environment characteristics, S represents the unified state characteristics, and W t , W e and W f They represent the weights of task embedding, environment embedding and feature fusion, respectively, and b t 、b e and b f They represent the corresponding biases, μ and σ represent normalization parameters;

[0181] S62. Output strategy through Actor network:

[0182] π(a|s)=softmax(W actor ·S+b actor );

[0183] Among them, π(a|s) represents the output strategy, W actor and b actor Represent the weight and bias of the Actor network respectively;

[0184] S63. Generate state value function through critic network:

[0185] V(s)=W critic ·S+b critic ;

[0186] Among them, V(s) represents the state value function, W critic and b critic Respectively represent the weight and bias of the Critic network;

[0187] S64, optimize the global task strategy, calculate the advantage function through the immediate reward and cumulative return, update the Actor network parameters, and update the Critic network parameters by minimizing the Critic network loss function:

[0188]

[0189] A(s,a)=R t -V(s);

[0190]

[0191] Among them, R t represents the cumulative return, γ represents the discount factor, r t represents the immediate reward, A(s,a) represents the global advantage function, α represents the learning rate, θ actor Represents the Actor network parameters, Denotes the parameter θ actor The gradient of L critic represents the Critic network loss function, θ criticrepresents the Critic network parameters, Denotes the parameter θ critic The gradient of

[0192] S65. Extract the local state features of the environmental parameters and generate the local scene strategy by optimizing the local strategy loss function:

[0193]

[0194] π local (a|s)=softmax(W actor-local ·S local +b actor-local );

[0195] L local =-log2π local (a|s)·A local (s,a);

[0196] Among them, H L represents the illumination characteristics, H O represents the overall obstacle characteristics, H D represents the delay characteristic, μ and σ represent the normalization parameters, S local represents the local state feature, W f and b f Respectively represent the weight and bias of the fusion network, [H L ,H O ,H D ] indicates that H L , H O and H D For splicing, L local represents the local strategy loss function, A local (s,a) represents the local advantage function, π local (a|s) represents the local scene strategy, W actor-local and b actor-local Respectively represent the weight and bias of the Actor network under environmental parameters;

[0197] S66. Output the optimized global task strategy and local scene strategy.

[0198] Embodiment 1:

[0199] In modern mobile games, complex game scenes and diverse operational requirements place high demands on players' operational skills, reaction speed, decision-making ability, and adaptability to scene changes. Especially in some action games or strategy games, players need to make real-time decisions in a constantly changing environment and quickly perform corresponding operations. This not only requires players to have excellent operational skills, but also requires the game system to be able to intelligently adjust its operational prompts, game rules, and even scene elements to provide a more personalized and intelligent gaming experience.

[0200] In this context, a complex scene adaptive operation method for mobile games based on multi-task learning came into being. By combining multi-task learning and reinforcement learning, this embodiment can adaptively adjust the game's operating logic and strategy according to the player's performance in the game and the dynamic changes of the game environment, thereby providing players with a smoother and more efficient gaming experience. Take a typical real-time strategy game as an example, in which players need to control multiple units and perform tasks in complex scenes. The scenes in the game include multiple dynamically changing elements, such as obstacles, enemies, resource points, etc. The player's operation not only needs to adjust tactics according to these changes, but also needs to reasonably allocate resources among multiple tasks and perform different operations.

[0201] In this complex scenario, the game system needs to have task allocation and resource scheduling, scene adaptive operation, and multimodal feature extraction and processing. Specifically, the system first extracts features of the player's operating behavior, visual information in the game scene, and game rules through a multimodal feature extraction network. Next, the system constructs a task graph and optimizes tasks through a multi-task learning model, and finally adjusts the priority of tasks based on the optimized task results. Through the reinforcement learning model of the Actor-Critic architecture, combined with scene parameters such as lighting, obstacles, and delays, the system continuously optimizes global task strategies and local scene strategies, and ultimately provides players with the best operation suggestions.

[0202] In order to verify the effectiveness of the present invention, multiple experiments were conducted to evaluate the impact of the multi-task learning model and reinforcement learning strategy optimization method on game performance and player experience in complex scenarios.

[0203] Table 1 Experimental comparison data of traditional method and method of the present invention

[0204] index Traditional methods Method of the present invention Task completion time (seconds) 40.5 32.8 Operational Fluency Rating (out of 10) 7.5 9.3 Intelligent operation score (out of 10) 6.3 9.0 Overall satisfaction rating (out of 10) 7.1 9.2 Average task completion accuracy 72.4 92.1 Average resource utilization 68.9 84.3 Stability (no lag) 3 0

[0205] The comparison of task completion time shows that the method of the present invention significantly shortens the time required for players to complete tasks. Specifically, players using traditional methods need an average of 40.5 seconds to complete the task, while players using the method of the present invention only need 32.8 seconds, saving nearly 19% of time. This difference shows that through the optimization of multi-task learning and reinforcement learning, the game system can more intelligently adjust the execution order and resource allocation of tasks, thereby improving the efficiency of task completion.

[0206] In terms of operation fluency, the players who adopted the method of the present invention scored a high score of 9.3 points, compared with 7.5 points for the traditional method. This difference reflects the optimization of the fluency and smooth experience of the players in the game operation process by the method of the present invention. By adjusting the operation logic and game environment in real time, the present invention can provide players with a smoother control experience, making the player's operation in complex scenes more natural and comfortable.

[0207] The intelligent operation score further proves the advantages of the method of the present invention in intelligent operation. The intelligent operation score of the traditional method is 6.3 points, while the score of the method of the present invention is 9.0 points, which is a significant difference. This difference reflects that the present invention combines multimodal feature extraction with multi-task learning to achieve in-depth analysis and understanding of player behavior, game scenes and rules, so that the game system can make more intelligent adjustments and optimizations according to environmental changes and player needs.

[0208] In terms of overall satisfaction rating, the method of the present invention also significantly improved the game satisfaction of players, with a rating of 9.2, while the traditional method was 7.1. This shows that the present invention not only improves the operating efficiency, but also has a positive impact on the overall perception of the player experience. Players are more willing to accept the system's intelligent adjustment and optimization operations during the game. This improvement is not only reflected in the game process, but also in the player's evaluation of the overall game experience.

[0209] In terms of average task completion accuracy, players using the method of the present invention can complete tasks more accurately, with a task completion accuracy of 92.1%, which is about 20% higher than the 72.4% of the traditional method. This improvement shows that through the optimization of task features and intelligent adjustment of task priorities by the multi-task learning model, players can be more focused and reduce invalid operations when performing tasks, thereby improving the accuracy and efficiency of task execution.

[0210] The improvement in average resource utilization further demonstrates the advantages of the method of the present invention in resource allocation and optimization in games. The resource utilization rate of the traditional method is 68.9%, while the resource utilization rate of players using the method of the present invention is increased to 84.3%. This improvement shows that through the strategy of combining multi-task learning with reinforcement learning, the game can more effectively schedule and allocate resources, helping players to efficiently complete more tasks under limited resource conditions, thereby improving the overall game execution efficiency.

[0211] Finally, the comparison of stability also further shows the advantages of the method of the present invention. Under the traditional method, there were 3 freezes during the game, while the method of the present invention achieved 0 freezes. This means that the present invention can avoid the common performance bottleneck problems in traditional methods while maintaining the smoothness of game operation, ensuring the continuity and stability of the game experience.

[0212] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for adaptive operation of complex scenes in mobile games based on multi-task learning, characterized in that: The steps include: S1. Collect data from complex scenes of mobile games, extract operation records, visual information and game rules, and preprocess the data to generate original multimodal data sets; S2. Based on the original multimodal dataset, a multimodal feature extraction network is constructed to extract visual features, operation behavior features and rule features, and a task graph is generated through the spatial relationship in the visual features, the task constraints in the rule features and the operation logic in the operation behavior features; S3. Enhance the visual features, operation behavior features and rule features, use the improved contrast learning method to optimize the expression of operation behavior features, and combine the multi-view fusion technology and embedded coding method to optimize the visual features and rule features respectively, and perform feature fusion on the optimized features to generate multimodal features; S4, taking multimodal features and task graphs as input, building a multi-task learning model, optimizing task features, adjusting task loss weights in combination with a dynamic weight allocation mechanism, and generating task priorities; S5. Based on the task priority and the optimized task characteristics, the tasks with dependencies in the task graph are uniformly optimized using a recursive optimization method to generate a global task optimization result. S6. Taking the global task optimization results as input, combined with the lighting parameters, obstacle parameters and delay parameters in the simulated environment, the global task strategy and local scene strategy are optimized through a reinforcement learning model based on the Actor-Critic architecture.

2. According to claim 1, a method for adaptive operation of complex scenes of mobile games based on multi-task learning is characterized in that: The S2 specifically includes: S21, constructing a multimodal feature extraction network, wherein the multimodal extraction network is composed of a bidirectional long short-term memory network, a ResNet34 network and a multi-layer perceptron; S22. Use a bidirectional long short-term memory network to extract operation records and generate operation behavior features: F op =f LSTM (X op ); Among them, F op Indicates the characteristics of the operation behavior, X op Indicates operation record, f LSTM represents a bidirectional long short-term memory network; S23. For visual information, the ResNet34 network is used to extract spatial features in the image and generate visual features: F vis =f CNN (X vis ); Among them, F vis represents visual features, X vis represents visual information, f CNN Represents the ResNet34 network; S24. For the game rules, the rule logic is embedded through a multi-layer perceptron to generate rule features: F rule =f MLP (X rule ); Among them, F rule represents the rule feature, X rule represents the rules of the game, f MLP represents a multi-layer perceptron; S25. Analyze the spatial relationship in the visual features, the task constraints in the rule features, and the operation logic in the operation behavior features to generate a task graph, which describes the dependency, parallel relationship, and mutual exclusion relationship between tasks.

3. The method for adaptively operating complex scenes of mobile games based on multi-task learning according to claim 1, characterized in that: The S3 specifically includes: S31. Use an improved contrastive learning method to optimize the expression of operational behavior characteristics; S32. Use multi-view fusion technology to optimize visual features, construct visual features for each view, perform weighted fusion on the features of each view, and generate optimized visual features: in, represents the optimized visual features, α m represents the weight of the mth perspective, represents the visual features of the mth perspective, M represents the number of perspectives, W represents the trainable parameters, represents the visual features of the j-th perspective, exp represents the natural exponential function; S33, processing the rule features, optimizing the rule features using an embedded coding method, and generating optimized rule features; S34. Perform feature fusion on the optimized features, use the self-attention mechanism to calculate the importance weight of the features, and calculate the weight matrix of each feature: Among them, W i represents the weight matrix of feature i, Q i and K i They represent the query and key mapping of the feature vector, d represents the feature dimension, T represents the transposition operation, and softmax represents the normalization function; S36. Generate multimodal features: in, represents multimodal features, V i Represents the value mapping of the eigenvector, and n represents the number of modes.

4. The method for adaptively operating complex scenes of mobile games based on multi-task learning according to claim 3, characterized in that: The S31 specifically includes: S311, based on operational behavior characteristics F op , construct contrastive learning sample pairs, the contrastive learning sample pairs include positive sample pairs And negative sample pairs And select the hard negative samples that are most similar to the positive sample features from the negative sample set to form a negative sample index set H(i); S312, using the main encoder and momentum encoder to analyze the operation behavior feature F op Perform feature projection and generate a stable feature representation Z by optimizing the weights of the momentum encoder op : θ=mθ+(1-m)θ; Where θ' represents the momentum encoder weight, θ represents the main encoder weight, and m represents the momentum coefficient; S313. Express Z based on stable features op Calculate the symmetric contrast loss for positive and negative sample pairs, including forward similarity optimization and reverse similarity optimization: Among them, L sym represents the symmetric contrast loss of positive and negative sample pairs, L op represents the one-way contrast loss, represents the positive sample feature representation generated by the momentum encoder, represents the reverse positive sample feature representation generated by the momentum encoder, represents the negative sample feature representation generated by the momentum encoder, sim represents the cosine similarity function, τ represents the temperature coefficient, H(i) represents the negative sample index set, and represents the feature representation generated by the momentum encoder; S314, operation behavior characteristics F op Perform multi-scale feature extraction to generate operational behavior feature representations at different scales Calculate the symmetric contrast loss of each scale separately and sum them: Among them, L multi Multi-scale contrastive learning total loss, represents the characteristic representation of the operational behavior at the s-th scale, and P represents the total number of characteristic scales; S315. Generate optimized operation behavior features by minimizing the total loss of multi-scale contrastive learning 5. The method for adaptive operation of complex scenes of mobile games based on multi-task learning according to claim 1, characterized in that: The S4 specifically includes: S41, constructing a multi-task learning model, inputting multimodal features into a shared feature extraction network to extract shared features, the multi-task learning model includes a shared feature extraction network and a task branch network, the shared feature extraction network is composed of a multi-layer convolutional neural network, and the task branch network is composed of a shallow convolutional neural network; S42, constructing an independent task branch network for each task according to the task nodes in the task graph, and extracting task features through the task branch network; S43. According to the dependency, competition or mutual exclusion relationship between tasks in the task graph, the task features are adjusted through the relevance of shared features: R ij =1-Mutex(t i ,t j ); in, represents the task features adjusted by shared features, R ij represents the inhibitory factor, W p , W s and W k They represent the weighted parameters of the parent task, shared features, and subtasks, respectively, and are used to control the contribution of features to the adjustment. i represents the subtask set, P i represents the parent task set, F shared represents shared features, and Respectively represent the specific features of the parent task and the specific features of the subtask, Mutex(t i ,t j ) represents task t i and t j Mutually exclusive relationship; S43, calculating the immediate utility for each task according to the task graph and the adjusted task characteristics; S44. Dynamically adjust task loss weights based on immediate utility: Among them, w i represents the loss weight, ∈ represents the smoothing factor to prevent the denominator from being zero, and u i and u j Respectively represent tasks t i and t j Immediate utility; S45. Extract the dependencies between tasks according to the task graph, build a dependency chain for each task, recursively optimize the dependency chain from parent tasks to child tasks, and update the features of the nested model: Among them, F nested (t i ) represents task t i The nested optimization feature, f nested represents a nested optimization function, and Indicates task characteristics; S46. Generate task priorities based on immediate utility and dependencies between tasks.

6. The method for adaptive operation of complex scenes of mobile games based on multi-task learning according to claim 1, characterized in that: The S5 specifically includes: S51. According to the task graph, extract the set of tasks that have a mutually exclusive relationship with the tasks, introduce mutually exclusive constraints on the nested features of the tasks, and adjust the task features through the mutually exclusive optimization function: Among them, f mutex represents the mutually exclusive optimization function, F nested (t i ) represents task t i The optimization feature, F nested (t k ) represents the optimization feature of mutually exclusive tasks, sim represents the cosine similarity function, λ represents the weight coefficient of the mutually exclusive constraint, which is used to control the strength of the penalty term, Mutex(t i ) represents a set of mutually exclusive tasks; S52. According to the task graph, extract a set of tasks that can be executed in parallel, build a global consistency optimization target based on the nested features of the parallel task set, jointly optimize the features of the parallel task set, and update the task features: Among them, L parallel represents the global consistency optimization objective function, Parallel(t i ) represents a set of parallel tasks, F nested (t k ) represents the optimization features of mutually exclusive tasks, sim represents the cosine similarity function, F shared The shared features represent the information shared between tasks; S53. Calculate the resource requirements of the task according to the task graph and task priority, dynamically adjust the resource allocation weight according to the task priority and resource requirements, and generate a global task optimization result.

7. The method for adaptively operating complex scenes of mobile games based on multi-task learning according to claim 1, characterized in that: The S6 specifically includes: S61. Build a reinforcement learning model based on the Actor-Critic architecture, taking the global task optimization results and environmental parameters as input to generate unified state features: A env =ReLU(W e ·HAVE BEEN env +b e ); in, represents the embedded task features, F nested (t i ) represents task t i The nested optimization feature, H global represents the global task characteristics, k represents the number of tasks, E env represents the set of environmental parameters, H env represents the global environment characteristics, S represents the unified state characteristics, and W t , W e and W f They represent the weights of task embedding, environment embedding and feature fusion, respectively, and b t , b e and b f They represent the corresponding biases, μ and σ represent normalization parameters; S62. Output strategy through Actor network: π(a|s)=softmax(W actor ·S+b actor ); Among them, π(a|s) represents the output strategy, W actor and b actor Represent the weight and bias of the Actor network respectively; S63. Generate state value function through critic network: V(s)=W critic ·S+b critic ; Among them, V(s) represents the state value function, W critic and b critic Respectively represent the weight and bias of the Critic network; S64, optimize the global task strategy, calculate the advantage function through the immediate reward and cumulative return, update the Actor network parameters, and update the Critic network parameters by minimizing the Critic network loss function: A(s,a)=R t -V(s); Among them, R t represents the cumulative return, γ represents the discount factor, r t represents the immediate reward, A(s,a) represents the global advantage function, α represents the learning rate, θ actor Represents the Actor network parameters, Denotes the parameter θ actor The gradient of L critic represents the Critic network loss function, θ critic represents the Critic network parameters, Denotes the parameter θ critic The gradient of S65. Extract the local state features of the environmental parameters and generate the local scene strategy by optimizing the local strategy loss function: p local (a|s)=softmax(W actor-local ·S local +b actor-local ); L local =-log2π local (a|s)·A local (s,a); Among them, H L represents the illumination characteristics, H O represents the overall obstacle characteristics, H D represents the delay characteristic, μ and σ represent the normalization parameters, S local represents the local state feature, W f and b f Respectively represent the weight and bias of the fusion network, [H L ,H O ,H D ] indicates that H L , H O and H D For splicing, L local represents the local strategy loss function, A local (s,a) represents the local advantage function, π local (a|s) represents the local scene strategy, W actor-local and b actor-local Respectively represent the weight and bias of the Actor network under environmental parameters; S66. Output the optimized global task strategy and local scene strategy.

Images