A reinforcement learning-based adversarial game strategy real-time optimization method and system

Abstract

The application provides a kind of method and system for real-time optimization of reinforcement learning-based confrontation game strategy, obtains historical game situation characteristics;Real-time capture of dynamic motion parameter data of target object to generate physical dynamic vector;The physical dynamic vector is input to the preset real-time data compression circuit for high-dimensional signal processing, to generate low-dimensional electrical response signal;Correlation analysis of the historical decision characteristics of the preset confrontation game strategy to be optimized and the historical game situation characteristics, to generate strategy evolution characteristics;Fusion strategy evolution characteristics and the time sequence continuity characteristics of low-dimensional electrical response signal, to generate dynamic confrontation environment state data;The dynamic confrontation environment state data is input into the preset reinforcement learning model, to generate the optimized confrontation game strategy.The application realizes the cross-domain collaborative closed loop from physical motion data to decision strategy, breaks through the response delay bottleneck of traditional game strategy optimization, and ensures that the confrontation strategy is updated in milliseconds with the target object motion state.

Description

Technical Field

[0001] This invention relates to the field of real-time optimization technology for adversarial game strategies, and in particular to a method and system for real-time optimization of adversarial game strategies based on reinforcement learning. Background Technology

[0002] With the development of video games and intelligent adversarial systems, real-time dynamic decision-making scenarios place higher demands on the responsiveness and adaptability of strategy generation. In environments such as multiplayer online competitive games and unmanned adversarial systems, the behavior of both sides exhibits high uncertainty and nonlinearity, making it difficult for traditional static strategies to meet the needs of real-time game development in complex environments. Therefore, there is an urgent need for a technical means that can integrate multi-source dynamic information, quickly respond to changes in opponent behavior, and continuously optimize decision-making paths during the adversarial process to achieve efficient and accurate real-time strategy generation and adjustment.

[0003] Current mainstream approaches combine multimodal data fusion with deep reinforcement learning. This involves collecting historical data on opponent behavior and current environmental state information to construct a multi-layered state representation model. Convolutional neural networks extract visual perception information, and recurrent neural networks model historical decision sequences, generating context-aware environmental state representations. These representations are then input into a deep reinforcement learning framework to enable online learning and optimization of adversarial strategies. This improves the real-time performance and adaptability of strategy generation, particularly in moderately complex adversarial environments. However, existing approaches have inherent limitations. These include reliance on high-dimensional perceptual data processing capabilities, making it difficult to guarantee the stability of state representations in complex environments with signal noise and data latency. Furthermore, the lack of deep coupling modeling between physical motion characteristics and strategy evolution when modeling opponent behavior trends leads to policy updates lagging behind changes in opponent behavior, affecting the foresight and accuracy of adversarial decisions. Finally, the high computational cost of these models makes it difficult to meet millisecond-level response requirements, limiting their effectiveness in high-speed adversarial scenarios. Summary of the Invention

[0004] This invention provides a method and system for real-time optimization of adversarial game strategies based on reinforcement learning. It addresses several issues in existing technologies, including reliance on high-dimensional perceptual data processing capabilities, difficulty in ensuring the stability of state representation in complex environments with signal noise and data latency, lack of deep coupling modeling between physical motion characteristics and strategy evolution when modeling opponent behavior evolution trends (resulting in strategy updates lagging behind opponent behavior changes and affecting the foresight and accuracy of adversarial decision-making), and high computational overhead that makes it difficult to meet the real-time requirement of millisecond-level response, thus limiting its application effectiveness in high-speed adversarial scenarios.

[0005] In a first aspect, the present invention provides a real-time optimization method for adversarial game strategies based on reinforcement learning, comprising:

[0006] Obtain historical game situation characteristics;

[0007] The dynamic motion parameter data of the target object is captured in real time using a pre-set inertial sensor to generate a physical dynamic vector;

[0008] The physical dynamic vector is input to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal.

[0009] The historical decision characteristics of the preset adversarial game strategy to be optimized are correlated with the historical game situation characteristics to generate strategy evolution characteristics.

[0010] The strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal are fused to generate dynamic adversarial environment state data.

[0011] The dynamic adversarial environment state data is input into a preset reinforcement learning model to generate an optimized adversarial game strategy.

[0012] Optionally, historical game situation characteristics can be obtained, including:

[0013] Receive historical adversarial game record data, wherein the historical adversarial game record data contains a complete sequence of interaction states for multiple historical game rounds;

[0014] Key adversarial nodes are identified from each complete interaction state sequence, and strategy selection data of the game operator at the key adversarial nodes and environmental constraint data of the adversarial decision-maker at the key adversarial nodes are collected.

[0015] The strategy selection data and the environmental constraint data are combined to generate multiple key adversarial situation segments;

[0016] By connecting the key adversarial situation segments in the chronological order of the historical game rounds, historical game situation features are generated.

[0017] Optionally, a preset inertial sensor is used to capture the dynamic motion parameter data of the target object in real time to generate a physical dynamic vector, including:

[0018] The raw motion signal of the target object is acquired using a pre-set inertial sensor;

[0019] The original motion signal is decomposed into three mutually orthogonal axial motion components.

[0020] Each of the aforementioned axial motion components is converted into dynamic motion parameter data;

[0021] Establish a mapping relationship between the motion direction of the target object and the physical space coordinate system of the target object, and bind the dynamic motion parameter data to the direction based on the mapping relationship to generate a motion parameter dataset with direction labels;

[0022] The motion parameter dataset with directional labels is encapsulated to generate a physical dynamic vector.

[0023] Optionally, the physical dynamic vector is input to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal, including:

[0024] The physical dynamic vector is subjected to axial data separation processing to obtain multiple axial motion parameter data;

[0025] The continuous physical quantities in each of the axial motion parameter data are quantized and scale-matched to obtain the level range.

[0026] Each of the continuous physical quantities is mapped to the level range to generate a discrete level range sequence;

[0027] Based on preset compression coding rules, the coaxial discrete level interval sequences are aggregated to generate compressed signal block units;

[0028] Differential encoding is performed on each of the compressed signal block units to generate a relative encoding group of level changes within each of the compressed signal block units;

[0029] The relative coding groups of each axis are reassembled and encapsulated according to the original encapsulation structure of the physical dynamic vector to obtain a low-dimensional electrical response signal.

[0030] Optionally, the historical decision characteristics of the preset adversarial game strategy to be optimized are correlated with the historical game situation characteristics to generate strategy evolution characteristics, including:

[0031] Obtain the sequence of key adversarial situation segments from the historical game situation characteristics;

[0032] Extract decision action data from the historical decision features of the preset adversarial game strategy to be optimized, which are at the same moment as each of the key adversarial situation segments;

[0033] The decision action data is compared with the strategy selection data in the key adversarial situation segment corresponding to the decision action data to extract the strategy deviation data from the comparison results.

[0034] The strategy deviation data is encapsulated to generate a strategy deviation vector;

[0035] The strategy deviation vector is associated and bound with the environmental constraint data of the key adversarial situation segment corresponding to the strategy deviation vector to generate multiple associated data pairs;

[0036] The associated data pairs are aggregated according to time sequence to generate strategy evolution features.

[0037] Optionally, the strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal are fused to generate dynamic adversarial environment state data, including:

[0038] The policy evolution features are subjected to temporal deconstruction processing to obtain multiple policy evolution units;

[0039] Dynamic change indicators are extracted from the temporal continuity characteristics of the low-dimensional electrical response signal to generate a dynamic change indicator set;

[0040] Bind the dynamic change index set to the strategy evolution unit corresponding to the time window as a state variable to generate bound state data;

[0041] The binding state data is structured and encapsulated to generate multiple fused data units;

[0042] The fused data units are aggregated according to the signal timing sequence to generate dynamic adversarial environment state data.

[0043] Optionally, the dynamic adversarial environment state data is input into a preset reinforcement learning model to generate an optimized adversarial game strategy, including:

[0044] The decision-making unit of the dynamic adversarial environment state data is deconstructed to obtain multiple sub-environment decision-making units;

[0045] Each of the sub-environment decision units is input into the policy inference engine of the preset reinforcement learning model to perform policy inference, and optimization factors are extracted from the policy inference results to obtain multiple policy optimization factors.

[0046] The optimization factors of each strategy and the constraints in each sub-environment decision unit are adaptively bound to generate multiple bound adaptive units.

[0047] The various bound adaptive units are aggregated to generate an optimized adversarial game strategy.

[0048] Secondly, the present invention provides a real-time optimization system for adversarial game strategies based on reinforcement learning, comprising:

[0049] The acquisition module is used to acquire historical game situation characteristics;

[0050] The capture module is used to capture the dynamic motion parameter data of the target object in real time using a pre-set inertial sensor in order to generate a physical dynamic vector;

[0051] The processing module is used to input the physical dynamic vector to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal.

[0052] The analysis module is used to perform correlation analysis between the historical decision characteristics of the preset adversarial game strategy to be optimized and the historical game situation characteristics to generate strategy evolution characteristics.

[0053] The fusion module is used to fuse the strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal to generate dynamic adversarial environment state data.

[0054] The generation module is used to input the dynamic adversarial environment state data into a preset reinforcement learning model to generate an optimized adversarial game strategy.

[0055] Thirdly, the present invention provides a computing device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a real-time optimization method for adversarial game strategies based on reinforcement learning as described in any of the first aspects.

[0056] Fourthly, the present invention provides a computer storage medium storing computer program instructions thereon, wherein the computer program instructions, when executed by a processor, implement a real-time optimization method for adversarial game strategies based on reinforcement learning as described in any one of the first aspects.

[0057] This invention uses pre-installed inertial sensors to capture the dynamic motion parameters of a target object in real time and generate a physical dynamic vector. A pre-installed data compression circuit then reduces the dimensionality of the high-dimensional physical signal to generate a low-dimensional electrical response signal. This data is further integrated with historical game situation characteristics and real-time signal timing features to construct dynamic adversarial environment state data. Finally, a reinforcement learning model generates an optimization strategy, achieving a cross-domain collaborative closed loop from physical motion data to decision-making strategy. This method overcomes the response delay bottleneck of traditional game strategy optimization. In real-time dynamic adversarial scenarios, through hard collaboration between the physical sensing layer and the electrical processing layer, it ensures that the adversarial strategy is updated in milliseconds according to the target object's motion state.

[0058] Furthermore, by performing axial separation, quantization scale matching, and discrete sequence generation operations on the physical dynamic vector, continuous physical quantities are dynamically mapped into discrete level interval sequences. These sequences are then processed through compressed signal block aggregation and differential coding to generate high-compression-ratio coded groups, which are finally reassembled into low-dimensional electrical signals according to the original encapsulation structure. This core scheme overcomes the processing delay problem of high-dimensional motion signals. Through the synergistic mechanism of dynamic level interval partitioning and axial structure-preserving encapsulation, reinforcement learning models can synchronously analyze the physical motion change characteristics, providing a millisecond-level response electrical signal foundation for real-time adversarial decision-making.

[0059] These or other aspects of the invention will become more apparent from the following description of the embodiments. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 A flowchart illustrating a real-time optimization method for adversarial game strategies based on reinforcement learning, provided in an embodiment of the present invention.

[0062] Figure 2 A schematic diagram of the structure of a real-time optimization system for adversarial game strategies based on reinforcement learning, provided in an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of the structure of a computing device provided in an embodiment of the present invention. Detailed Implementation

[0064] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0065] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Figure 1 This invention provides a flowchart of a real-time optimization method for adversarial game strategies based on reinforcement learning, as shown in the following embodiment. Figure 1 As shown, the method includes:

[0068] Existing technologies suffer from three core shortcomings in real-time dynamic adversarial scenarios: First, the processing of high-dimensional perception data relies on complex neural network architectures, which can lead to instability in state representation and a sharp drop in decision reliability due to signal noise and transmission delay interference. Second, the lack of a deep coupling mechanism between physical motion features and strategy evolution in opponent behavior modeling causes strategy updates to lag behind actual adversarial changes, resulting in a loss of game foresight. Finally, the excessive computational load of the model cannot meet the millisecond-level response requirements of high-speed adversarial scenarios, severely limiting its practicality. To address these issues, this invention proposes the following approach: First, it uses pre-installed inertial sensors to directly capture the physical motion features of the target object, generating dynamic vectors to circumvent the bottleneck of high-dimensional perception data processing. Second, it introduces a real-time data compression circuit to perform hardware-level dimensionality reduction on the physical vectors, compressing the electrical signal delay to a negligible level. Third, it simultaneously integrates the strategy evolution features of historical game situations with the temporal continuity features of real-time electrical signals to construct a deeply coupled model of physical motion and strategy decision-making. Finally, it uses a reinforcement learning model to generate optimized strategies driven by dynamic environment state data. This solution replaces traditional software algorithm computation with a hard collaboration between the physical sensing layer and the electrical processing layer, achieving end-to-end millisecond-level response closed loop while ensuring the depth of strategy evolution. This completely solves the industry pain points of decision lag and instability in high-speed adversarial scenarios. Based on this, this invention provides a real-time optimization method for adversarial game strategies based on reinforcement learning, such as... Figure 1 ,include:

[0069] Step 101: Obtain historical game situation characteristics.

[0070] In this step, the historical game situation characteristics refer to the time-series data formed by linking key confrontation situation segments, including the strategy choices and environmental constraints of both sides at the strategic conflict nodes, reflecting the evolution pattern of the historical confrontation situation.

[0071] In this embodiment of the invention, historical adversarial game record data is first received, which contains multiple complete interaction state sequences; second, key adversarial nodes are identified from each sequence, and the strategy selection data of the game operator at that node and the environmental constraint data of the adversarial decision-maker are collected; then, the strategy selection data and environmental constraint data are combined into key adversarial situation segments; finally, all segments are concatenated in historical time sequence to generate historical game situation features that reflect the evolution of the adversarial situation.

[0072] Step 102: Use a pre-set inertial sensor to capture the dynamic motion parameter data of the target object in real time to generate a physical dynamic vector.

[0073] In this step, the target object refers to the operating subject that generates physical motion input, and its dynamic motion is captured by inertial sensors and transformed into decision-making basis; dynamic motion parameter data refers to physical quantity data that reflects the motion characteristics of the target object in three-dimensional space, collected by inertial sensors, including standard dimension parameters such as acceleration and angular velocity; physical dynamic vector refers to a structured vector formed by encapsulating the motion parameter dataset with direction labels, preserving the directional characteristics of the target object's spatial motion.

[0074] In this embodiment of the invention, the original motion signal of the target object in three-dimensional physical space is first continuously acquired by a pre-set inertial sensor; then the original signal is decomposed into three mutually orthogonal axial motion components; next, each component is converted into motion parameter data with standard physical dimensions; then a mapping relationship between the motion direction of the target object and the physical coordinate system is established, and the motion parameter data is bound with direction labels based on the mapping relationship; finally, the dataset with direction labels is encapsulated to generate a physical dynamic vector that retains the spatial motion characteristics.

[0075] Step 103: Input the physical dynamic vector into a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal.

[0076] In this step, high-dimensional signal processing refers to the electrical processing of reducing the dimensionality of physical motion vectors through quantization matching, discrete sequence generation, and differential coding; low-dimensional electrical response signal refers to the electrical signal generated after compression and recombination that retains the physical orientation characteristics, used for real-time transmission of motion change information.

[0077] In this embodiment of the invention, firstly, an axial separation operation is performed on the physical dynamic vector to obtain motion parameter data for each axis; secondly, the continuous physical quantities of each axis are dynamically divided into level ranges through quantization scale matching; then, the continuous physical quantities are mapped to the corresponding level ranges to generate discrete sequences; subsequently, the coaxial discrete sequences are aggregated into compressed signal block units based on preset compression coding rules; then, differential coding is performed on each signal block to generate a level change relative coding group; finally, each axial coding group is reorganized according to the original encapsulation structure to generate a low-dimensional electrical response signal that retains the physical direction characteristics.

[0078] Step 104: Perform correlation analysis between the historical decision characteristics of the preset adversarial game strategy to be optimized and the historical game situation characteristics to generate strategy evolution characteristics.

[0079] In this step, the preset adversarial game strategy to be optimized refers to the set of historical decision rules that serve as the optimization benchmark; historical decision characteristics refer to the record of decision actions of the strategy to be optimized in historical adversarial situations; correlation analysis refers to the process of comparing strategy selection data with decision action data and extracting deviation vectors; strategy evolution characteristics refer to the features formed by binding and aggregating strategy deviation vectors with environmental constraint data, revealing the direction of strategy improvement.

[0080] In this embodiment of the invention, firstly, the sequence of key adversarial situation segments in the historical game situation features is obtained; secondly, the decision action data at the same moment as each segment is extracted from the historical decision features of the strategy to be optimized; then, the consistency comparison operation is performed between the decision action data and the strategy selection data of the corresponding segment; subsequently, the strategy deviation data is extracted from the comparison results and encapsulated into a vector; finally, the strategy deviation vector is bound to the environmental constraint data of the corresponding segment, and the data is aggregated in time sequence to generate strategy evolution features that reflect the evolution law of strategy.

[0081] Step 105: Fuse the strategy evolution features and the temporal continuity features of the low-dimensional electrical response signal to generate dynamic adversarial environment state data.

[0082] In this step, temporal continuity characteristics refer to the fluctuation frequency and amplitude indicators in low-dimensional electrical signals that reflect the continuity of motion changes; fusion operation refers to the cross-domain coupling process that binds the strategy evolution unit with the dynamic change indicators; dynamic adversarial environment state data refers to real-time decision input data that integrates the temporal characteristics of physical signals and the laws of strategy evolution.

[0083] In this embodiment of the invention, firstly, the strategy evolution characteristics are deconstructed temporally to separate multiple strategy evolution units; secondly, a set of dynamic change indicators is extracted from the temporal continuity characteristics of the low-dimensional electrical response signal; then, the set of dynamic change indicators is bound to the state variables of the strategy evolution units corresponding to the time window; subsequently, the binding results are encapsulated in a structured manner to generate fused data units; finally, all fused units are aggregated in temporal order to generate dynamic adversarial environment state data that integrates physical signals and strategy evolution.

[0084] Step 106: Input the dynamic adversarial environment state data into the preset reinforcement learning model to generate an optimized adversarial game strategy.

[0085] In this step, the preset reinforcement learning model refers to the machine learning architecture of the built-in policy inference engine; the optimized adversarial game strategy refers to the decision rules generated by the aggregated adaptive units that adapt to the real-time environment and historical patterns.

[0086] In this embodiment of the invention, firstly, a decision unit deconstruction operation is performed on the dynamic adversarial environment state data to separate multiple sub-environment decision units; secondly, each unit is input into the policy inference engine of the reinforcement learning model for policy inference; then, policy optimization factors are extracted from the inference results; subsequently, the optimization factors are adaptively bound to the environmental constraints of the corresponding units to generate adaptive units; finally, all adaptive units are aggregated and encapsulated into an optimized adversarial game strategy that takes into account both the real-time environment and historical evolution patterns.

[0087] For example, in a real-time video game scenario, the system first collects raw 3D motion signals using the inertial sensor built into the player's controller. These signals are then spatially decomposed and oriented to generate a physical dynamic vector with axial labels. Next, this vector is input into a hardware compression circuit, where it undergoes quantization, discretization, signal block aggregation, and differential encoding to reassemble into a low-dimensional electrical signal. Simultaneously, strategy selection and environmental constraint data from key nodes in historical gameplay are extracted. The player's historical decisions are compared with their strategy choices to generate a strategy deviation vector, which is then bound to environmental constraints and aggregated into strategy evolution features. These features are then deconstructed into units, bound to dynamic change indicators of the electrical signal, and encapsulated into fusion units to generate environmental state data. Finally, this data is input into a reinforcement learning model, where it undergoes decision unit deconstruction, strategy deduction, and constraint binding to aggregate and output a real-time optimized game strategy.

[0088] This invention uses a pre-installed inertial sensor to capture physical motion in real time and generate directional vectors, which are then converted into low-latency electrical signals by a hardware-level compression circuit. Simultaneously, it integrates historical policy evolution patterns with real-time signal dynamic features to construct environmental state data. This drives a reinforcement learning model to generate optimized decisions that take into account both motion changes and policy evolution, breaking through the bottlenecks of traditional methods in signal processing delay, policy response lag, and insufficient environmental adaptability, and achieving millisecond-level accurate policy updates in high-speed adversarial scenarios.

[0089] To address the issues of missing key nodes and fragmented game elements in historical game data analysis, this step identifies policy conflict nodes and combines policy selection and environmental constraint data to construct a time-series sequence of key adversarial situation segments. This invention provides a specific embodiment, step 101, which involves acquiring historical game situation characteristics, specifically including the following steps:

[0090] Step 111: Receive historical adversarial game record data, wherein the historical adversarial game record data contains a complete sequence of interactive states for multiple historical game rounds.

[0091] In this step, historical adversarial game record data refers to a data set that stores multiple complete game processes, including all interaction states and environmental parameters of both sides from the start to the end; historical game rounds refer to an independent adversarial game cycle, including the entire process record of the start, adversarial phase and end; complete interaction state sequence refers to the set of states arranged in chronological order in a single round, reflecting the continuous trajectory of strategic interactions between the two sides and environmental changes.

[0092] In this embodiment of the invention, firstly, historical adversarial game record data containing a complete sequence of interactive states from multiple historical game rounds is received; secondly, all interactive states of the adversaries from the beginning to the end of each round are parsed; then, key data fields reflecting the game process are extracted from the state sequence; finally, the parsed complete sequence is used as the input source for subsequent key node identification.

[0093] Step 112: Identify key adversarial nodes from each complete interaction state sequence, and collect strategy selection data of the game operator at the key adversarial nodes and environmental constraint data of the adversarial decision-maker at the key adversarial nodes.

[0094] In this step, the key adversarial node refers to the critical moment when strategic conflict occurs during the game, which is manifested as the state point where the decision intentions of both parties are directly opposed; the game operator refers to the operator that actively executes the strategy, and its strategy selection data reflects the active decision intention; the strategy selection data refers to the specific action instructions and parameter configurations taken by the operator at the key node; the adversarial decision-maker refers to the decision-making entity that passively responds to the strategy, and its environmental constraint data constitutes the decision-making constraints; the environmental constraint data refers to the rule constraint parameters imposed by the decision-maker, including adversarial conditions such as resource limits and action range.

[0095] In this embodiment of the invention, firstly, the key adversarial nodes with significant policy conflicts are dynamically located by scanning the complete interactive state sequence; secondly, the policy selection data executed by the game operator at the node are collected; at the same time, the environmental constraint data imposed by the adversarial decision-maker at the same node are collected; then, the spatiotemporal consistency of the two types of data is verified; finally, the time-stamped policy selection dataset and environmental constraint dataset are output.

[0096] Step 113: Combine the strategy selection data and the environmental constraint data to generate multiple key adversarial situation fragments.

[0097] In this step, the combination operation refers to the structured processing of binding strategy selection data and environmental constraint data according to spatiotemporal relationships; the key adversarial situation segment refers to the combination of game elements at a single key node, including the operator's strategy, the decision-maker's constraints, and the node's spatiotemporal identifier.

[0098] In this embodiment of the invention, firstly, the strategy selection data and environmental constraint data of the same key adversarial node are aligned by timestamp; secondly, a mapping relationship between strategy selection and environmental constraints is established; then, key adversarial situation fragments containing two-way game elements are generated through data structuring and binding operations; finally, node identifiers and time stamps are added to each fragment.

[0099] Step 114: Connect the key adversarial situation segments in sequence according to the time order of the historical game rounds to generate historical game situation features.

[0100] In this step, time sequence refers to the rule of arranging historical events in chronological order; serial operation refers to the processing mechanism of connecting discrete situation segments in time sequence to form a continuous data chain.

[0101] In this embodiment of the invention, all key adversarial situation segments are first grouped according to the historical game round number; then, within each group, the segments are arranged in ascending order according to their time markers; subsequently, the segments are concatenated into a continuous data stream through a time-series connection operation; finally, all round data streams are integrated to generate historical game situation features that preserve the overall picture of the evolution of the adversarial situation.

[0102] This invention addresses the problems of coarse-grained historical data analysis, omission of key nodes, and fragmentation of game elements in traditional methods by dynamically identifying strategic conflict nodes and separating and collecting strategy and constraint data of both sides of the game. It generates adversarial situation fragments through structured binding and then constructs historical situation features by sequentially connecting them. This provides a high-fidelity historical adversarial situation foundation for strategy evolution.

[0103] To address the issue of lost orientation information during motion signal acquisition, this step generates a labeled set of motion parameters through spatial axial decomposition and orientation binding, which is then encapsulated into a physical dynamic vector that retains spatial characteristics. This invention provides a specific embodiment: Step 102 utilizes a pre-set inertial sensor to capture the dynamic motion parameter data of the target object in real time to generate the physical dynamic vector, specifically including the following steps:

[0104] Step 201: Use a preset inertial sensor to collect the original motion signal of the target object.

[0105] In this step, the raw motion signal refers to the voltage waveform data output by the inertial sensor, which reflects the instantaneous motion state of the target object in three-dimensional space.

[0106] In this embodiment of the invention, the original voltage signal in the three-dimensional physical space is first collected in real time by an inertial sensor pre-installed on the target object; then the voltage signal is subjected to noise filtering; then continuous signal segments are extracted at a fixed sampling frequency; finally, the original motion signal stream reflecting the instantaneous motion state of the target object is output.

[0107] Step 202: Perform spatial axial decomposition on the original motion signal to obtain three mutually orthogonal axial motion components.

[0108] In this step, spatial axial decomposition processing refers to the digital signal processing operation that separates the composite motion signal into independent components according to the orthogonal axes of the physical coordinate system; the axial motion component refers to the subset of motion parameters along a specific coordinate axis direction, including the axial projection values ​​of acceleration or angular velocity.

[0109] In this embodiment of the invention, the orthogonal axes of the physical space coordinate system are first defined; then the original motion signal is projected along the coordinate system axes; subsequently, the axial component separation calculation is performed by a digital signal processor; finally, three axial motion component data that are independent of each other and orthogonal in direction are generated.

[0110] Step 203: Convert each of the axial motion components into dynamic motion parameter data.

[0111] In this step, the conversion operation refers to the calculation process of converting the voltage signal into motion parameters with standard physical dimensions based on the sensor calibration coefficient.

[0112] In this embodiment of the invention, the voltage amplitude of each axial motion component is first read; then, scaling transformation is performed according to the physical quantity conversion coefficient preset by the sensor; subsequently, the voltage value is converted into motion parameters with standard physical dimensions; finally, a dataset of acceleration and angular velocity with axial identification is output.

[0113] Step 204: Establish a mapping relationship between the motion direction of the target object and the physical space coordinate system of the target object, bind the direction of each dynamic motion parameter data based on the mapping relationship, and generate a motion parameter dataset with direction labels.

[0114] In this step, the mapping relationship refers to the spatial angle correspondence model between the actual motion direction of the target object and the physical coordinate system; the direction binding operation refers to the spatial attribute enhancement processing of multiplying the motion parameter data with the direction weight coefficient; the direction label refers to the metadata tag that identifies the spatial direction to which the motion parameter data belongs; the motion parameter dataset refers to the structured data set that integrates direction labels, physical quantity values ​​and time series information.

[0115] In this embodiment of the invention, firstly, a model is established to show the angular relationship between the target object's motion direction vector and the physical space coordinate axis; secondly, the direction mapping weight coefficient is calculated based on the angular relationship; then, the weight coefficient is multiplied by the motion parameter data of the corresponding axis; finally, a direction label is added to the product result to generate a motion parameter dataset that retains spatial attributes.

[0116] Step 205: Encapsulate the motion parameter dataset with directional labels to generate a physical dynamic vector.

[0117] In this step, the encapsulation operation refers to the standardized process of encoding the dataset into machine-readable binary vectors according to a preset protocol.

[0118] In this embodiment of the invention, the dataset with direction labels is first sorted by axial number; then, a structured data unit containing timestamps, axial identifiers, and motion parameter values ​​is constructed; then, the data unit is encoded into a binary stream through a vector encapsulation protocol; finally, a physical dynamic vector that can be directly parsed by the electrical processing circuit is generated.

[0119] This invention preserves the directional characteristics of motion vectors through spatial axial decomposition, generates a spatially labeled dataset through physical quantity conversion and directional weight binding, and finally encapsulates it into a structured vector. This solves the problem of lost directional information in traditional motion signal acquisition and provides high-fidelity spatial motion feature input for subsequent electrical processing.

[0120] To address the issues of delay and orientation distortion in high-dimensional physical signal processing, this step generates a discrete sequence through quantization scale matching, which is then compressed using differential coding and reassembled into a low-dimensional orientation-preserving electrical signal according to its original structure. This invention provides a specific embodiment where step 103 involves inputting the physical dynamic vector to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal. This specifically includes the following steps:

[0121] Step 301: Perform axial data separation processing on the physical dynamic vector to obtain multiple axial motion parameter data.

[0122] In this step, axial data separation processing refers to the operation of splitting the dataset into independent subsets according to the axial identifier of the physical dynamic vector, ensuring that motion parameters in each direction are processed independently; axial motion parameter data refers to the separated data set containing single-axis motion parameters and direction labels, used to maintain spatial orientation characteristics.

[0123] In this embodiment of the invention, firstly, an axial separation operation is performed on the physical dynamic vector to disassemble the data according to the original encapsulation structure; secondly, the motion parameter dataset with direction labels is extracted from the vector; then, the dataset is separated according to the axial label; finally, multiple axial motion parameter datasets containing independent direction data are generated.

[0124] Step 302: Perform quantization and scale matching processing on the continuous physical quantities in each of the axial motion parameter data to obtain the level range.

[0125] In this step, continuous physical quantities refer to the original motion parameter values ​​that are not discretized, such as acceleration or angular velocity; quantization scale matching processing refers to the adaptive operation of dividing the level interval according to the dynamic range of the physical quantity, which is the basis for converting physical quantities into electrical quantities; the level interval range refers to the preset discretized level segments for continuous physical quantities, with each interval corresponding to a specific voltage amplitude range.

[0126] In this embodiment of the invention, firstly, the continuous physical quantity values ​​in the motion parameter data of each axis are read; secondly, the level interval range is dynamically divided according to the preset range; then, the mapping relationship between the physical quantity and the level interval is calculated; finally, a unique level interval division rule is generated for each axis.

[0127] Step 303: Map each of the continuous physical quantities to the level range to generate a discrete level range sequence.

[0128] In this step, the discrete level interval sequence refers to the ordered data chain with interval identifiers formed after mapping continuous physical quantities, preserving the original timing relationship.

[0129] In this embodiment of the invention, continuous physical quantities are first classified according to the level interval rules of the corresponding axis; then, the discrete level interval to which the physical quantity belongs is determined; subsequently, an interval identifier is added to each data point; finally, a discrete level interval sequence that preserves the timing relationship is generated.

[0130] Step 304: Based on the preset compression coding rules, the coaxial discrete level interval sequences are aggregated to generate compressed signal block units.

[0131] In this step, the preset compression coding rule refers to the hardware processing specifications that define the signal block length and aggregation method; the aggregation operation refers to the process of integrating discrete sequences into signal blocks according to time windows; the compressed signal block unit refers to the data block with axial marking generated by aggregation, which serves as the input unit for differential coding.

[0132] In this embodiment of the invention, a preset compression encoding rule is first loaded; then, the coaxial discrete sequence is divided into equal-length time windows according to the rule; subsequently, the data within the window is aggregated to generate compressed signal block units; finally, an axial identifier and a time stamp are added to each unit.

[0133] Step 305: Perform differential encoding on each of the compressed signal block units to generate a relative encoding group of level changes within each of the compressed signal block units.

[0134] In this step, differential coding refers to the processing mechanism that calculates and encodes the changes in adjacent levels within a signal block; relative coding group refers to the set of codes composed of level changes, reflecting the trend of signal fluctuation.

[0135] In this embodiment of the invention, the discrete levels within the signal block unit are first arranged in chronological order; then the difference between adjacent level changes is calculated; subsequently, the difference is encoded into a relative change; and finally, a relative coding group composed of the relative change is generated.

[0136] Step 306: Reassemble and encapsulate the relative coding groups of each axis according to the original encapsulation structure of the physical dynamic vector to obtain a low-dimensional electrical response signal.

[0137] In this step, the original encapsulation structure refers to the arrangement order of axial data and the orientation label system in the physical dynamic vector; the re-encapsulation refers to the operation of integrating the coding group according to the original structure to generate a low-dimensional signal while maintaining spatial orientation consistency.

[0138] In this embodiment of the invention, firstly, each relative coding group is grouped according to the axial identification; secondly, the reassembly order is determined according to the original encapsulation structure of the physical dynamic vector; then, the coaxial coding groups are spliced ​​together according to the time sequence; finally, it is encapsulated into a low-dimensional electrical response signal that retains the spatial orientation characteristics.

[0139] This invention preserves the spatial characteristics of motion vectors through axial separation, achieves efficient conversion of physical quantities to electrical quantities through adaptive level division and discrete sequence generation, compresses data dimensions through signal block aggregation and differential coding, and finally reassembles the original structure to ensure lossless transmission of directional information, thus solving the problems of high-dimensional signal processing delay and directional distortion, and providing low-latency, directional electrical signal input for real-time decision-making.

[0140] To address the disconnect between strategy evaluation and environmental constraints, this step extracts strategy deviation data through time-series alignment, encapsulates it into a vector, and then binds it to environmental constraints to aggregate it into evolutionary features. This invention provides a specific embodiment: Step 104 involves performing correlation analysis between the historical decision features of the preset adversarial game strategy to be optimized and the historical game situation features to generate strategy evolution features, specifically including the following steps:

[0141] Step 401: Obtain the sequence of key adversarial situation segments from the historical game situation features.

[0142] In this step, the key adversarial situation segment sequence refers to the set of adversarial node data arranged in chronological order, with each segment containing the dual elements of strategy selection and environmental constraints.

[0143] In this embodiment of the invention, firstly, a historical game situation feature library is loaded; secondly, a set of key adversarial situation segments in the feature library is located; then, all segments are sorted by timestamp; finally, a sequence of key adversarial situation segments with time sequence identifiers is generated.

[0144] Step 402: Extract decision action data from the historical decision features of the preset adversarial game strategy to be optimized, which are at the same moment as each of the key adversarial situation segments.

[0145] In this step, decision action data refers to the operation instructions and parameter records actually executed by the strategy to be optimized in historical confrontations.

[0146] In this embodiment of the invention, the historical decision features of the adversarial game strategy to be optimized are first analyzed; then the occurrence time of each key adversarial situation segment is matched; then the historical decision action records at the same time are extracted; finally, a decision action dataset aligned with the time sequence of the segments is output.

[0147] Step 403: Perform a data consistency comparison between the decision action data and the strategy selection data in the key adversarial situation segment corresponding to the decision action data, and extract the strategy deviation data from the comparison results.

[0148] In this step, the data consistency comparison operation refers to the field-level difference detection process between the strategy selection data and the decision action data; the comparison result refers to the set of quantitative difference values ​​output by the consistency comparison; and the strategy deviation data refers to the decision execution deviation metric extracted from the comparison result.

[0149] In this embodiment of the invention, the decision action data is first compared with the strategy selection data of the corresponding key confrontation situation segment by field; then the difference value between the two data is calculated; then the strategy execution deviation is extracted from the difference value; finally, strategy deviation data reflecting the degree of decision deviation is generated.

[0150] Step 404: Encapsulate the policy deviation data to generate a policy deviation vector.

[0151] In this step, the encapsulation operation refers to the process of organizing the deviation data into a structured vector with time-series labels; the policy deviation vector refers to a multi-dimensional data structure that integrates deviation intensity and time-series attributes.

[0152] In this embodiment of the invention, the time-series markers and deviation amounts of the strategy deviation data are first integrated; then the data elements are arranged according to a preset vector structure; subsequently, they are encapsulated into a structured data body containing deviation intensity and time-series attributes; and finally, a strategy deviation vector is generated.

[0153] Step 405: Associate and bind the strategy deviation vector with the environmental constraint data of the key adversarial situation segment corresponding to the strategy deviation vector to generate multiple associated data pairs.

[0154] In this step, the association binding operation refers to the processing mechanism that establishes a coupling relationship between the strategy deviation vector and the environmental constraint data; the association data pair refers to the binding body of the strategy deviation and the environmental constraints, revealing the pattern of decision-making being constrained by the environment.

[0155] In this embodiment of the invention, firstly, environmental constraint data of key adversarial situation segments are read; secondly, a mapping relationship between the strategy deviation vector and the constraint data is established; then, a bidirectional correlation body is generated through data structure binding operations; finally, the correlation data pairs with environmental constraint labels are output.

[0156] Step 406: Aggregate the associated data pairs according to the time sequence to generate strategy evolution features.

[0157] In this step, the aggregation operation refers to the process of joining related data pairs in time sequence to form continuous features.

[0158] In this embodiment of the invention, all associated data pairs are first arranged in chronological order; then, a continuous data chain is constructed through a time-series join operation; subsequently, the data chain is aggregated to generate policy evolution features that reflect the coupling evolution law between policy and environment.

[0159] This invention extracts the deviation data between decision actions and policy selection through temporal alignment, encapsulates it into a vector, dynamically binds it with environmental constraints, and then aggregates it to generate policy evolution features. This solves the problem of the separation between policy evaluation and environmental constraints in traditional methods, and provides reinforcement learning with evolutionary features that integrate decision deviation and environmental constraints.

[0160] To address the spatiotemporal separation between historical strategies and real-time physical signals, this step binds strategy evolution units and dynamic change indicators through time windows, encapsulating and aggregating them into spatiotemporally unified dynamic environment data. This invention provides a specific embodiment: Step 105, fusing the temporal continuity characteristics of the strategy evolution features and the low-dimensional electrical response signal to generate dynamic adversarial environment state data, specifically including the following steps:

[0161] Step 501: Perform temporal deconstruction processing on the policy evolution features to obtain multiple policy evolution units.

[0162] In this step, temporal deconstruction refers to the operation of splitting continuous features into independent units according to time attributes, while preserving the original temporal relationships; the policy evolution unit refers to the smallest temporal analysis body generated by the deconstruction of policy evolution features, which includes the binding relationship between policy deviation and environmental constraints within a single time window.

[0163] In this embodiment of the invention, the temporal structure of the policy evolution features is first parsed; then, it is split into independent units according to the original time attributes; subsequently, a time identifier is added to each unit; finally, a set of policy evolution units that retains the smallest logical unit of policy evolution is generated.

[0164] Step 502: Extract dynamic change indicators from the temporal continuity characteristics of the low-dimensional electrical response signal to generate a dynamic change indicator set.

[0165] In this step, the dynamic change index refers to the quantized parameters extracted from the time-series characteristics of the electrical signal, including real-time motion characteristics such as the frequency and amplitude of signal fluctuations.

[0166] In this embodiment of the invention, the temporal continuity characteristics of the low-dimensional electrical response signal are first loaded; then, the changing trend characteristics in the signal fluctuation are identified; subsequently, the rate of change and amplitude parameters are extracted; finally, the parameters are integrated to generate a dynamic change index set that reflects real-time motion changes.

[0167] Step 503: Bind the dynamic change index set to the policy evolution unit corresponding to the time window as state variables to generate bound state data.

[0168] In this step, the time window refers to the time period benchmark unit for aligning the strategy evolution unit with the dynamic indicator set; the state variable binding operation refers to the process of establishing the mathematical coupling relationship between strategy evolution parameters and real-time motion indicators; and the bound state data refers to the fused data volume formed by binding strategy elements and motion indicators, revealing the correlation between strategy and real-time motion.

[0169] In this embodiment of the invention, firstly, the time windows of the alignment strategy evolution unit and the dynamic change index set are aligned; secondly, the mapping relationship between strategy elements and index parameters within the unit is established; then, data binding operation is performed to generate a coupled body; finally, bidirectional binding state data of the fusion strategy and real-time motion state is output.

[0170] Step 504: Perform structured encapsulation processing on the binding state data to generate multiple fused data units.

[0171] In this step, structured encapsulation refers to the operation of reorganizing data elements according to a preset template to generate standardized data packets; fused data unit refers to the smallest data entity that simultaneously carries the strategy evolution law and real-time motion characteristics after encapsulation.

[0172] In this embodiment of the invention, the structural elements of the bound state data are first read; then the elements are reorganized according to a preset data template; subsequently, they are encapsulated into independent data packets with timestamps; and finally, a fused data unit containing both strategy evolution and motion changes is generated.

[0173] Step 505: Aggregate the fused data units according to the signal timing sequence to generate dynamic adversarial environment state data.

[0174] In this step, the signal timing sequence refers to the time sequence rule arranged according to the order in which electrical signals occur; the aggregation operation refers to the processing mechanism that connects discrete data units in time sequence to construct a continuous data chain.

[0175] In this embodiment of the invention, all fused data units are first sorted according to the timing of electrical signals; then adjacent units are connected through a data link protocol; subsequently, a continuous timing data stream is constructed; and finally, dynamic adversarial environment state data that fully reflects the dynamic changes of the adversarial environment is generated.

[0176] This invention addresses the problem of disconnect between historical policies and real-time physical signals by deconstructing policy evolution features into units and precisely binding them with dynamic indicators of electrical signals according to time windows. After being structured and encapsulated into fusion units, these units are aggregated into environmental state data, thus solving the problem of disconnect between historical policies and real-time physical signals and providing spatiotemporally unified dynamic decision input for reinforcement learning.

[0177] To address the issues of insufficient policy feasibility and optimization lag in complex constrained environments, this step deconstructs the decision-making unit to parallel deduce optimization factors, and then aggregates them through constraint binding to generate adaptive policies. This invention provides a specific embodiment where step 106 involves inputting the dynamic adversarial environment state data into a preset reinforcement learning model to generate an optimized adversarial game strategy, specifically including the following steps:

[0178] Step 601: Deconstruct the decision unit of the dynamic adversarial environment state data to obtain multiple sub-environment decision units.

[0179] In this step, the decision unit refers to the smallest data body with complete decision logic in the dynamic adversarial environment state data, which includes environmental state and strategy elements; the deconstruction operation refers to the process of splitting the decision unit into independent sub-units according to the internal logical relationship; the sub-environment decision unit refers to the data entity containing single-scenario decision elements after deconstruction, which carries the binding relationship between environmental state and strategy parameters.

[0180] In this embodiment of the invention, the internal structure of the dynamic adversarial environment state data is first analyzed; then, the complete decision logic units contained in the data are identified; subsequently, the data body is decomposed according to the smallest decision unit; and finally, a set of sub-environment decision units that retain independent decision elements is generated.

[0181] Step 602: Input each of the sub-environment decision units into the policy inference engine of the preset reinforcement learning model to perform policy inference, extract optimization factors from the policy inference results, and obtain multiple policy optimization factors.

[0182] In this step, the policy inference engine refers to the computational core of the reinforcement learning model that simulates policy paths based on state inputs; the policy inference operation refers to the mathematical simulation process of executing the environment state to policy output in the engine; the policy inference result refers to the set of policy paths output by the inference operation and its effect evaluation data; the optimization factor refers to the policy improvement elements extracted from the inference result, including the amount of action parameter adjustment and the timing of execution; the policy optimization factor refers to the set of optimization elements with weighted coefficients, which quantifies the direction and intensity of policy improvement.

[0183] In this embodiment of the invention, the sub-environment decision unit is first input into the policy inference engine of the reinforcement learning model; then, policy simulation calculation based on environmental state and historical patterns is performed; subsequently, policy improvement elements are identified from the inference results; and finally, multiple sets of policy optimization factors reflecting the policy optimization direction are extracted and output.

[0184] Step 603: Adaptively bind the constraints in each of the strategy optimization factors and each of the sub-environment decision units to generate multiple bound adaptive units.

[0185] In this step, constraints refer to environmental rule parameters that limit the feasibility of strategies in sub-environment decision units; adaptive binding operation refers to the calculation process of establishing the mathematical fit relationship between optimization factors and constraints; and the bound adaptive unit refers to the executable strategy unit generated after the optimization factors have been fitted with constraints.

[0186] In this embodiment of the invention, the environmental constraints in the sub-environment decision unit are first read; then, an adaptation relationship model between the strategy optimization factors and the constraints is established; subsequently, the mathematical binding operation between the factors and the constraints is performed; finally, a bound adaptive unit that takes into account both optimization potential and environmental feasibility is generated.

[0187] Step 604: Aggregate the bound adaptive units to generate an optimized adversarial game strategy.

[0188] In this step, the aggregation operation refers to the processing mechanism that connects strategy units according to temporal logic to form a complete decision chain.

[0189] In this embodiment of the invention, all bound adaptive units are first sorted according to the decision occurrence sequence; then adjacent units are connected through a policy link protocol; subsequently, a complete policy logic chain is constructed; and finally, an executable optimized adversarial game strategy is encapsulated and generated.

[0190] This invention achieves parallel policy optimization by deconstructing the environmental decision-making unit, extracting improvement factors through deduction and dynamically binding them with environmental constraints, and finally aggregating to generate an optimization scheme that takes into account both real-time environmental adaptability and policy improvement potential, thus solving the problems of insufficient policy feasibility and optimization lag in traditional reinforcement learning in complex constrained environments.

[0191] Figure 2 This invention provides a schematic diagram of the structure of a real-time optimization system for adversarial game strategies based on reinforcement learning, as shown in the embodiment of the invention. Figure 2 As shown, the system includes:

[0192] Module 21 is used to acquire historical game situation characteristics;

[0193] The capture module 22 is used to capture the dynamic motion parameter data of the target object in real time using a preset inertial sensor in order to generate a physical dynamic vector;

[0194] Processing module 23 is used to input the physical dynamic vector to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal;

[0195] Analysis module 24 is used to perform correlation analysis between the historical decision characteristics of the preset adversarial game strategy to be optimized and the historical game situation characteristics to generate strategy evolution characteristics.

[0196] The fusion module 25 is used to fuse the strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal to generate dynamic adversarial environment state data.

[0197] The generation module 26 is used to input the dynamic adversarial environment state data into a preset reinforcement learning model to generate an optimized adversarial game strategy.

[0198] Figure 2 The aforementioned real-time optimization system for adversarial game strategies based on reinforcement learning can execute... Figure 1 The implementation principle and technical effects of the real-time optimization method for adversarial game strategies based on reinforcement learning described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the real-time optimization system for adversarial game strategies based on reinforcement learning in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0199] In one possible design, Figure 2 The real-time optimization system for adversarial game strategies based on reinforcement learning, as shown in the embodiment, can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0200] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0201] The processing component 32 is used to: acquire historical game situation characteristics; use a pre-set inertial sensor to capture dynamic motion parameter data of the target object in real time to generate a physical dynamic vector; input the physical dynamic vector to a pre-set real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal; perform correlation analysis between the historical decision characteristics of the preset adversarial game strategy to be optimized and the historical game situation characteristics to generate strategy evolution characteristics; fuse the strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal to generate dynamic adversarial environment state data; and input the dynamic adversarial environment state data into a preset reinforcement learning model to generate an optimized adversarial game strategy.

[0202] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0203] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0204] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0205] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0206] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0207] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0208] This invention also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1 The embodiment shown is a real-time optimization method for adversarial game strategies based on reinforcement learning.

[0209] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0210] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0211] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time optimization method for adversarial game strategies based on reinforcement learning, characterized in that, include: Obtain historical game situation characteristics; The dynamic motion parameter data of the target object is captured in real time using a pre-set inertial sensor to generate a physical dynamic vector; The physical dynamic vector is input to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal. The historical decision characteristics of the preset adversarial game strategy to be optimized are correlated with the historical game situation characteristics to generate strategy evolution characteristics. The strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal are fused to generate dynamic adversarial environment state data. The dynamic adversarial environment state data is input into a preset reinforcement learning model to generate an optimized adversarial game strategy.

2. The method according to claim 1, characterized in that, Obtain historical game situation characteristics, including: Receive historical adversarial game record data, wherein the historical adversarial game record data contains a complete sequence of interaction states for multiple historical game rounds; Key adversarial nodes are identified from each complete interaction state sequence, and strategy selection data of the game operator at the key adversarial nodes and environmental constraint data of the adversarial decision-maker at the key adversarial nodes are collected. The strategy selection data and the environmental constraint data are combined to generate multiple key adversarial situation segments; By connecting the key adversarial situation segments in the chronological order of the historical game rounds, historical game situation features are generated.

3. The method according to claim 1, characterized in that, The system utilizes pre-set inertial sensors to capture dynamic motion parameter data of the target object in real time, thereby generating a physical dynamic vector, including: The raw motion signal of the target object is acquired using a pre-set inertial sensor; The original motion signal is decomposed into three mutually orthogonal axial motion components. Each of the aforementioned axial motion components is converted into dynamic motion parameter data; Establish a mapping relationship between the motion direction of the target object and the physical space coordinate system of the target object, and bind the dynamic motion parameter data to the direction based on the mapping relationship to generate a motion parameter dataset with direction labels; The motion parameter dataset with directional labels is encapsulated to generate a physical dynamic vector.

4. The method according to claim 1, characterized in that, The physical dynamic vector is input to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal, including: The physical dynamic vector is subjected to axial data separation processing to obtain multiple axial motion parameter data; The continuous physical quantities in each of the axial motion parameter data are quantized and scale-matched to obtain the level range. Each of the continuous physical quantities is mapped to the level range to generate a discrete level range sequence; Based on preset compression coding rules, the coaxial discrete level interval sequences are aggregated to generate compressed signal block units; Differential encoding is performed on each of the compressed signal block units to generate a relative encoding group of level changes within each of the compressed signal block units; The relative coding groups of each axis are reassembled and encapsulated according to the original encapsulation structure of the physical dynamic vector to obtain a low-dimensional electrical response signal.

5. The method according to claim 1, characterized in that, The historical decision characteristics of the preset adversarial game strategy to be optimized are correlated with the historical game situation characteristics to generate strategy evolution characteristics, including: Obtain the sequence of key adversarial situation segments from the historical game situation characteristics; Extract decision action data from the historical decision features of the preset adversarial game strategy to be optimized, which are at the same moment as each of the key adversarial situation segments; The decision action data is compared with the strategy selection data in the key adversarial situation segment corresponding to the decision action data to extract the strategy deviation data from the comparison results. The strategy deviation data is encapsulated to generate a strategy deviation vector; The strategy deviation vector is associated and bound with the environmental constraint data of the key adversarial situation segment corresponding to the strategy deviation vector to generate multiple associated data pairs; The associated data pairs are aggregated according to time sequence to generate strategy evolution features.

6. The method according to claim 1, characterized in that, The strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal are fused to generate dynamic adversarial environment state data, including: The policy evolution features are subjected to temporal deconstruction processing to obtain multiple policy evolution units; Dynamic change indicators are extracted from the temporal continuity characteristics of the low-dimensional electrical response signal to generate a dynamic change indicator set; Bind the dynamic change index set to the strategy evolution unit corresponding to the time window as a state variable to generate bound state data; The binding state data is structured and encapsulated to generate multiple fused data units; The fused data units are aggregated according to the signal timing sequence to generate dynamic adversarial environment state data.

7. The method according to claim 1, characterized in that, The dynamic adversarial environment state data is input into a preset reinforcement learning model to generate an optimized adversarial game strategy, including: The decision-making unit of the dynamic adversarial environment state data is deconstructed to obtain multiple sub-environment decision-making units; Each of the sub-environment decision units is input into the policy inference engine of the preset reinforcement learning model to perform policy inference, and optimization factors are extracted from the policy inference results to obtain multiple policy optimization factors. The optimization factors of each strategy and the constraints in each sub-environment decision unit are adaptively bound to generate multiple bound adaptive units. The various bound adaptive units are aggregated to generate an optimized adversarial game strategy.

8. A real-time optimization system for adversarial game strategies based on reinforcement learning, characterized in that, include: The acquisition module is used to acquire historical game situation characteristics; The capture module is used to capture the dynamic motion parameter data of the target object in real time using a pre-set inertial sensor in order to generate a physical dynamic vector; The processing module is used to input the physical dynamic vector to a preset real-time data compression circuit for high-dimensional signal processing to generate a low-dimensional electrical response signal. The analysis module is used to perform correlation analysis between the historical decision characteristics of the preset adversarial game strategy to be optimized and the historical game situation characteristics to generate strategy evolution characteristics. The fusion module is used to fuse the strategy evolution characteristics and the temporal continuity characteristics of the low-dimensional electrical response signal to generate dynamic adversarial environment state data. The generation module is used to input the dynamic adversarial environment state data into a preset reinforcement learning model to generate an optimized adversarial game strategy.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a real-time optimization method for adversarial game strategies based on reinforcement learning as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements a real-time optimization method for adversarial game strategies based on reinforcement learning as described in any one of claims 1 to 7.

Images