Game play-out test method, device and electronic equipment

Abstract

This disclosure provides a game simulation testing method and apparatus. During game simulation testing, game configuration parameters and test standards are loaded, and multiple virtual agents are initialized. A corresponding virtual session is created for each virtual agent, and the virtual session is attached to a physical computing node. This allows the physical computing node to perform game logic calculations based on the current state data of the virtual agents, obtaining the latest state data of the virtual agents. While the physical computing node is running normally, test indicators are obtained based on the latest state data corresponding to the multiple virtual agents. If the test indicators meet the test standards, the game configuration parameter test is considered passed. If not, an exception is triggered, and new game configuration parameters are generated. The game simulation test continues in the game test environment generated based on the new game configuration parameters. This method improves the efficiency and quality of game simulation testing with low resource consumption.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and in particular to a game simulation testing method and apparatus. Background Technology

[0002] With the rapid development of the video game industry, the size and complexity of modern games have increased exponentially. Throughout the entire lifecycle of a game, from project initiation to official launch, the dynamic deduction of game states and the automated balancing of numerical systems are the most crucial aspects determining the success or failure of a game product. In traditional game development processes, balance testing and state deduction mainly rely on manual black-box testing by a large quality assurance team, or Monte Carlo simulations by designers based on static tables. However, when there are thousands of entities with independent behavioral logic in the game world, and these entities have complex nonlinear interactions (such as dynamic market transactions, guild territory battles, and ecological food chain evolution), static numerical deduction and limited manual testing cannot reach the system's boundary conditions and emergent behaviors. Summary of the Invention

[0003] In view of this, the purpose of this disclosure is to provide a game simulation testing method and apparatus to improve the efficiency and quality of game simulation testing with low resource consumption.

[0004] In a first aspect, embodiments of this disclosure provide a game simulation testing method, which includes: responding to a game simulation testing instruction, loading game configuration parameters and testing standards, and initializing multiple virtual agents to join a game testing environment generated by the game configuration parameters; for each of the multiple virtual agents, creating a virtual session corresponding to the virtual agent, and attaching the virtual session to a physical computing node, so that the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent; obtaining test indicators based on the latest state data corresponding to the multiple virtual agents, and determining whether the test indicators meet the testing standards; if they meet the standards, determining that the game configuration parameter test has passed; if they do not meet the standards, performing an abnormal circuit breaker and generating new game configuration parameters, so as to continue the game simulation test in the game testing environment generated based on the new game configuration parameters.

[0005] Secondly, this disclosure also provides a game simulation testing device, which includes: an initialization module, used to respond to a game simulation testing command, load game configuration parameters and test standards, and initialize multiple virtual agents to join the game test environment generated by the game configuration parameters; a session mounting module, used to create a virtual session corresponding to each of the multiple virtual agents, and mount the virtual session to a physical computing node, so that the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent; a result judgment module, used to obtain test indicators based on the latest state data corresponding to the multiple virtual agents, and determine whether the test indicators meet the test standards; and a test result determination module, used to determine that the game configuration parameter test is passed if the test is satisfied; if the test is not satisfied, an abnormal circuit breaker is performed and new game configuration parameters are generated so that the game simulation test can continue to be performed in the game test environment generated based on the new game configuration parameters.

[0006] Thirdly, this disclosure provides an electronic device including a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the above-described game simulation test method.

[0007] Fourthly, this disclosure provides a computer-readable storage medium storing computer-executable instructions that, when invoked and executed by a processor, cause the processor to implement the aforementioned game simulation test method.

[0008] The embodiments disclosed herein bring the following beneficial effects: This disclosure provides a game simulation testing method and apparatus. First, in response to a game simulation testing command, game configuration parameters and testing standards are loaded, and multiple virtual agents are initialized and added to the game testing environment generated by the game configuration parameters. For each virtual agent, a corresponding virtual session is created and attached to a physical computing node. This allows the physical computing node to perform game logic calculations based on the current state data of the virtual agent, obtaining the latest state data of the virtual agent. Based on the latest state data of the multiple virtual agents, test indicators are obtained, and it is determined whether the test indicators meet the testing standards. If they do, the game configuration parameter test is considered passed; otherwise, an exception is triggered, and new game configuration parameters are generated to continue the game simulation test in the game testing environment generated based on the new game configuration parameters. In this method, the virtual agent acts as a lightweight virtual session, and this virtual session is dynamically attached to the physical computing node, achieving complete decoupling between physical computing resources and the agent's logical state, and also realizing multiplexing of the physical computing node, thereby improving the efficiency and quality of game simulation testing.

[0009] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure are realized and obtained through the structures particularly pointed out in the description, claims and drawings.

[0010] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the specific embodiments of this disclosure or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0012] Figure 1 A flowchart of a game simulation testing method provided in this embodiment of the disclosure; Figure 2 This is a diagram illustrating a virtual session decoupling and remounting structure provided in an embodiment of this disclosure. Figure 3 A state diagram of an A2A dynamic topology communication mesh provided in this embodiment of the disclosure; Figure 4 A dynamic behavior tree and two-layer decision architecture diagram provided in this disclosure embodiment; Figure 5 This is a schematic diagram of the structure of a game simulation testing device provided in an embodiment of the present disclosure; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0014] Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely to illustrate selected embodiments of the disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0015] With the maturation of reinforcement learning and generative agent technologies driven by Large Language Models (LLM), the industry has begun to explore the introduction of AI (Artificial Intelligence) agents into game testing. By deploying a large number of automated testing agents on the server or client side, the behavior of real players is simulated in order to discover potential numerical vulnerabilities, economic system collapse points, or level design flaws before the game is launched.

[0016] However, applying massive numbers of AI agents to large-scale game state simulations presents extremely stringent engineering challenges. The underlying architecture of modern game engines was not designed from the outset to support tens of thousands of AI agents with deep thinking and high-frequency interaction capabilities simultaneously.

[0017] In the current game development pipeline, there are three main existing technical solutions for large-scale automated testing and state simulation. However, these solutions all have theoretical flaws and engineering bottlenecks in terms of architecture design, resource scheduling, and communication mechanisms.

[0018] Option 1, based on a distributed cluster simulation of all client instances, is based on the core idea of ​​launching a large number of complete game client instances in a cloud server cluster. Each instance is injected with a scripted AI based on a behavior tree or a simple state machine. Specifically, the system assigns an independent process to each virtual player, which includes a complete graphics rendering pipeline, a physics engine calculation module, and a network synchronization module.

[0019] Assuming the game test scenario requires simulating N agents, and each client instance has a base memory footprint of M... base The CPU overhead for physical and logical operations is C. logic The network synchronization overhead is C. net The total resource consumption R of the cluster is then... total It can be represented as:

[0020] Here, O(NlogN) represents the additional scheduling overhead on the server side for handling these concurrent client requests. Based on this, the drawback of the above method is: extreme resource waste, inability to support ultra-large-scale simulations, and because each agent is bound to a heavy client instance, M... base The constant term is extremely large (typically in the GB range). When N > 10,000, hardware costs explode linearly or even superlinearly, making large-scale ecosystem simulations of hundreds of thousands of clients impossible. In economic system simulations lasting several weeks, if a physical node carrying thousands of clients crashes or experiences network instability, these client processes will be directly destroyed by the operating system.

[0021] Option two involves headless agent inference based on a centralized message bus. To address the issue of overly cumbersome client-side architecture, some leading vendors have stripped away the rendering and physical layers, running lightweight logical agents directly on the server. However, to enable interaction between these agents (e.g., simulating player trading, team formation, and chat), existing technologies typically employ a centralized message bus architecture. Specifically, in this approach, all agent state updates and interaction requests must first be sent to a centralized message queue, sorted and matched by a central scheduler, and then broadcast to other agents.

[0022] However, this approach has the following drawbacks: When faced with high-frequency agent-to-agent communication involving hundreds of thousands of agents, the centralized message bus inevitably experiences severe network congestion and read / write lock contention. Existing agents cannot establish dynamic point-to-point communication topologies based on "spatial proximity" or "social relationship graphs" like real players. All messages must travel through a central node, resulting in extremely low inference efficiency and an inability to simulate the rapid spread of local information in the real world (such as panic selling due to a sudden price surge in a certain area of ​​the game).

[0023] To address the aforementioned issues, this disclosure provides a game simulation testing method and apparatus, which can be applied to game testing scenarios.

[0024] To facilitate understanding of this disclosure, a game simulation testing method provided by an embodiment of this disclosure will first be described in detail, such as... Figure 1 As shown, the method includes the following specific steps: In step S102, in response to the game simulation test command, the game configuration parameters and test standards are loaded, and multiple virtual agents are initialized to join the game test environment generated by the game configuration parameters.

[0025] In practice, when the game development team decides to conduct automated simulation testing on a particular game version, the operator initiates the game simulation testing task, which means sending a game simulation testing command to the system. After receiving the command, the system loads the game configuration parameters. These parameters define the basic rules of the game world, including: the attributes of various game monsters (health points, attack power, item drops and drop probabilities), the basic parameters of the economic system (base item prices, transaction tax rates, gold production rates), and the character growth curves (experience points required for leveling up, coefficients in skill damage formulas), etc. Essentially, the game configuration parameters can be understood as the game's rulebook; the entire game simulation test will run under these rules.

[0026] After recording the game configuration parameters, the system also determines the testing standards, which are usually pre-set by the game designers. These standards define what numerical states are considered healthy. For example, the design team might set the following: the server-wide gold inflation rate should be kept below 5% per day; the difference in average DPS (Damage Per Second) between any class should not exceed 15%; and the wealth Gini coefficient for a certain type of rare item should not exceed 0.4. The testing standards also define allowable tolerance ranges. For instance, if a test metric occasionally deviates slightly from the standard range, the system will not immediately trigger an alert, but will only take action when the deviation exceeds the tolerance range.

[0027] After receiving the game simulation test command, the system will initialize multiple virtual agents and add them to the game test environment generated by the game configuration parameters. That is, the system will create multiple virtual agents (for example, 100,000 or 150,000 virtual agents) and set initial state data for each virtual agent. This initial state data includes: initial location (randomly distributed in various areas of the game world), initial assets (each virtual player has a certain amount of initial gold coins and basic equipment), and personality settings (assigned different personalities and behavioral tendencies through a large language model, such as "conservative player", "aggressive merchant", "social expert", etc.).

[0028] It should be noted that the aforementioned virtual intelligent agents are automated test entities used in game simulation testing scenarios to simulate the behavior of real human players using AI. These test entities can also be called virtual players. Before the game's official launch, the game development team needs to know whether the game's economic system will collapse and whether the combat balance will be broken if 100,000 real players are online simultaneously. However, before the game's launch, the development team cannot actually find 100,000 real players to test with. Therefore, this invention uses AI to create hundreds of thousands of virtual players, allowing them to move, trade, fight, and socialize in the game world like real humans, thereby conducting game simulation testing.

[0029] Step S104: For each virtual agent among the multiple virtual agents, create a virtual session corresponding to the virtual agent and attach the virtual session to the physical computing node so that the physical computing node can perform game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent.

[0030] The system creates a virtual session for each virtual agent and assigns these virtual sessions to different physical computing nodes. For example, virtual agents 1 and 2 are assigned to physical computing node P1, and virtual agent 3 is assigned to physical computing node P2, and so on. The state data of the virtual agents is stored in a pre-defined distributed memory to ensure that the state data of the virtual agents is not lost.

[0031] During normal operation, the physical computing nodes continuously execute game logic calculations to determine the latest state data of the virtual agent. Simultaneously, this latest state data is written to the distributed memory.

[0032] Step S106: Based on the latest state data corresponding to multiple virtual intelligent agents, obtain the test indicators and determine whether the test indicators meet the test standards.

[0033] The system aggregates the latest status data of all virtual agents and calculates test indicators reflecting the game's economy and combat balance. These test indicators may include: the total amount of gold coins in the entire server and the inflation rate, the average market price of various items, the average combat power of different professions, and the Gini coefficient of wealth distribution, etc.

[0034] Determine whether the test indicators meet the test standards. If they do, the game simulation test is normal; if they do not, the game simulation test is abnormal.

[0035] Step S108: If the conditions are met, the game configuration parameter test is confirmed to be passed; if not, an abnormal circuit breaker is triggered and new game configuration parameters are generated to continue the game simulation test in the game test environment generated based on the new game configuration parameters.

[0036] If the test metrics meet the test standards, the current game configuration parameters are considered to have passed the test and can be used in the official launch. If the test metrics do not meet the test standards, all simulation activities based on the current game configuration parameters will be stopped, and new game configuration parameters will be generated to allow game simulation tests to continue based on the new parameters, in order to find suitable game configuration parameters when the test is successful.

[0037] This disclosure provides a game simulation testing method in which a virtual agent is used as a lightweight virtual session, and the virtual session is dynamically mounted on a physical computing node. This achieves complete decoupling of physical computing resources and the logical state of the agent, and also realizes multiplexing of physical computing nodes, thereby improving the efficiency and quality of game simulation testing.

[0038] The following examples describe how virtual sessions mount physical computing nodes.

[0039] Specifically, the process of creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node may include: creating a corresponding virtual session for the virtual agent at a preset virtual session layer, and attaching the virtual session to the corresponding physical computing node through a dynamic mapping router.

[0040] In its implementation, this disclosure abstracts an independently running virtual session layer between the physical computing node and the game logic layer. Each virtual agent no longer exists as a heavy client process, but rather resides in the virtual session layer as a lightweight virtual session layer. The computing resources (CPU / memory) of the physical computing node are dynamically mounted to the virtual session only as a rendering / computing host.

[0041] In practical applications, the set of physics calculation nodes in a game simulation system is defined as follows: The set of virtual players (equivalent to virtual intelligent agents) that need to be simulated is: ,generally For any virtual intelligent agent , its in The state data at a given time is defined as a high-dimensional tensor. This tensor contains three subspaces:

[0042] in, Indicates the environmental perception state (current coordinates, entities within the field of view); Represents long-term memory states (historical transaction records, hate values, social graphs); Indicates the state of behavioral intent (the currently executing decision tree node or the prompt word of the preset large language model).

[0043] Within the architecture of the virtual session layer, define a dynamic mapping function. Used in Virtual sessions at all times Bind to physical computing node On the computation thread:

[0044] When the mapping is established, it is called the mounted state, at which point the virtual session... Mounted on physical computing nodes Above, thus physical computing nodes Responsible for calculation State transition equation:

[0045] in, This is the underlying logic calculation function for the game. For environmental input, For virtual sessions State data at time t For virtual sessions State data at time t-1.

[0046] In practical applications, the underlying logic calculation functions of a game calculate the state changes for the next moment based on the state of the previous moment and the environmental input. These functions can also perform combat calculations, economic system calculations, position movement and collision detection, mission progress updates, and skill cooldown and effect calculations.

[0047] In combat settlement, when a virtual player (which can be understood as a virtual intelligent agent) attacks another target, the game's underlying logic function needs to calculate "how much damage this attack will cause." It comprehensively considers various factors such as the attacker's attack power, the target's defense, whether it's a critical hit, and skill bonuses, ultimately calculating a damage value. Then, the target's health decreases by the corresponding amount, and the attacker may gain experience points. For example, if a virtual player's current state is "attack power 100, attacking a monster with defense 30," the game's underlying logic function calculates "causing 70 points of damage," so the monster's health changes from 500 to 430—this is one state transition.

[0048] In the economic system's settlement process, when virtual players engage in buying and selling transactions in the game market, the game's underlying logic settlement function is responsible for calculating the changes in assets for both parties. For example, if player A buys an iron sword for 100 gold coins, the game's underlying logic settlement function will calculate: player A's gold coins decrease by 100, and their inventory gains an iron sword; the merchant's gold coins increase by 100, and their inventory decreases by an iron sword. Additionally, if the game has a transaction tax, the game's underlying logic settlement function will also deduct the corresponding tax.

[0049] In position movement and collision detection, when a virtual player moves in the game world, the game's underlying logic calculation function needs to calculate its new coordinates, check whether it has collided with a wall, or entered a special area (such as a safe zone or a danger zone), and update its status accordingly. In task progress updates, when a virtual player completes a task objective (such as "kill 10 wolves"), the game's underlying logic calculation function checks whether the task completion conditions are met and updates the task's progress status.

[0050] In skill cooldown and effect calculation, skills in the game usually have a cooldown time (for example, after releasing a fireball, you need to wait 5 seconds before you can release it again). The game's underlying logic calculation function will update the cooldown timer of each skill in real time and calculate the continuous effect (for example, when you are poisoned, you lose 10 health points per second).

[0051] The aforementioned environmental inputs are the collection of all external stimuli and information from the game environment and outside the virtual players themselves. Specifically, these environmental inputs can include global game environment parameters, local environment information, other player interaction inputs, and system configuration update parameters. Among them, global game environment parameters refer to the state of the game world itself, which affects all virtual players. These parameters include in-game time information, weather system status (whether the current weather in the game world is sunny, rainy, or blizzard, and weather affects character movement speed, field of vision, and the effects of certain skills), global event trigger signals (e.g., "The server-wide guild war will begin in 30 minutes," "A rare BOSS has appeared in a certain area," "Double experience event is activated," etc.), and server-wide parameters (such as the current number of online players and system-level configuration information such as the server's resource output multiplier). Local environmental information describes the environmental conditions around the virtual player's location. The local environment seen by each virtual player may be different, specifically including entity information within the field of vision (what other characters are around the virtual player, where are these characters located, are they friendly or hostile, and what are they currently doing), terrain and interactive object information (whether there are any mineable ores, trees that can be cut down, caves that can be entered, obstacles that need to be climbed, etc. near the virtual player), and area attribute information (whether the area where the virtual player is currently located is a safe zone or a danger zone, and how abundant are the resources in the area).

[0052] The aforementioned other player interaction inputs come from the direct behavioral influence of other virtual players on the current virtual player. Specifically, these include: attack information (e.g., other virtual players or monsters are attacking the current virtual player), trade requests (e.g., other virtual players initiate trade requests to the current virtual player), social interaction information (e.g., party invitations, guild invitations, chat messages, etc. sent by other virtual players), and price changes in the surrounding market (e.g., the latest price of an item in the auction house).

[0053] In an optional embodiment, the specific process of attaching the virtual session to the corresponding physical computing node may include: obtaining the running status data of multiple preset computing nodes: determining the physical computing node configured for the virtual session from multiple preset computing nodes based on the game area where the virtual session is located and the running status data, and attaching the virtual session to the corresponding physical computing node.

[0054] For each physical computing node, the system monitors its operational status data in real time. This data may include CPU (Central Processing Unit) utilization, memory utilization, the number of virtual players already hosted, and remaining network bandwidth. The system then combines these operational statuses to calculate a load score. The lower the load score, the more idle the physical computing node is, and the more suitable it is to receive new virtual agents.

[0055] In the game world, virtual agents within the same area frequently interact (for example, hunters in the same forest may need to cooperate in hunting). To reduce communication latency, virtual agents geographically close in the game world are preferentially assigned to the same or adjacent physical computing nodes. Combining information from both the load rating of the physical computing nodes and the spatial distance between the virtual agents, a final decision is output: which physical computing node should the virtual agent be assigned to. Specifically, among all available physical computing nodes, the one with the lowest load rating and the best match for the game area where the virtual agent is located is selected to mount the virtual agent.

[0056] In an optional embodiment, after creating a virtual session corresponding to a virtual agent, the state data of the virtual agent corresponding to the virtual session is stored in a preset distributed memory.

[0057] In practical implementation, the virtual session layer automatically stores the state data of the virtual agent corresponding to the virtual session in a preset distributed memory at a preset frequency to avoid data loss and facilitate subsequent use. This preset frequency can be 10 times / second or 50 times / second, etc.

[0058] In practical applications, when a physical computing node crashes, the system can retrieve the state data of the virtual intelligent agent attached to that physical computing node from distributed memory, thereby reducing the loss rate of the implemented state data.

[0059] In one optional embodiment, the physical computing node retrieves the current state data of the virtual agent from distributed memory and performs game logic calculations based on the current state data to obtain the latest state data of the virtual agent. Then, at a preset frequency, the final state data of the virtual agent is synchronously written into the distributed memory. This method helps to preserve the final state data of the virtual agent and prevents the loss of state data.

[0060] After creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, determine whether the physical computing node is malfunctioning. If the physical computing node is malfunctioning, interrupt the game logic calculations performed on the physical computing node and remove the attachment of the virtual session to the physical computing node. Re-determine the target computing node from multiple preset computing nodes and re-attach the virtual session to the target computing node. If the physical computing node is running normally, execute the steps described above to obtain test indicators based on the latest status data corresponding to multiple virtual agents.

[0061] When remounting the virtual session to the target computing node, the target computing node retrieves the latest state data of the virtual agent corresponding to the virtual session from the distributed memory, performs game logic calculations based on the latest state data, and updates the latest state data of the virtual agent.

[0062] In practical implementation, when a failure is detected on a physical computing node, all virtual agents attached to that physical computing node are immediately cleared, i.e., F_map(V_i, t_crash) → (Empty set). Within a very short time (milliseconds) after being cleared, the mapping function F_map will find new physical computing nodes for these virtual intelligent entities, i.e., F_map(V_i, t_crash + Δt) → P_new; where t_crash is the failure time of the physical computing node, and Δt represents the time interval from the physical computing node crashing (t_crash) to the virtual intelligent entity being successfully remounted to the new physical computing node (P_new).

[0063] like Figure 2 The diagram shown illustrates a virtual session decoupling and remounting architecture provided in this disclosure. The entire architecture is divided into three layers from top to bottom: the top layer is a distributed memory grid, the middle layer is a virtual session layer (VSL), and the bottom layer is a physical computing cluster. These three layers cooperate with each other through data connections and control commands, collectively forming the core technical architecture of this disclosure.

[0064] Among them, the distributed memory grid is a high-speed storage system specifically designed to store the state data of virtual agents (equivalent to the aforementioned distributed memory). Figure 2The system comprises N storage units, labeled as follows: memory unit S_1 stores the state data of the first virtual agent, memory unit S_2 stores the state data of the second virtual agent, and memory unit S_N stores the state data of the Nth virtual agent. This state data includes all information such as the virtual agent's location coordinates, owned items, completed tasks, social relationships, and historical behavior records. It is stored in a distributed memory grid rather than locally on the physical computing nodes, ensuring that even if the physical computing nodes fail, the state data remains securely stored in this independent storage system.

[0065] The virtual session layer sits between the storage layer and the computing layer, acting as a "bridge" and "scheduling center." Figure 2 In this architecture, the virtual session layer contains the following components: Virtual Session 1, Virtual Session 2, and Virtual Session N: these are the virtual sessions corresponding to each virtual agent. Each virtual session is a lightweight logical unit. Virtual sessions are connected to the distributed memory grid above via bidirectional arrows, allowing them to read and write the state data of their corresponding virtual agents at any time. Figure 2 The dynamic mapping router is the core control component in the virtual session layer. All virtual sessions are connected to this router. The router's responsibility is to determine which physical computing node each virtual session should be attached to for computation, or to unattach the virtual session from the physical computing node when it fails.

[0066] Figure 2 The physical computing cluster in the context includes actual server hardware, that is, physical computing nodes that actually perform computing tasks. Figure 2 The diagram shows three physical computing nodes: Physical Node P1 (normal operation): This is a physical computing node in good working condition, currently handling and computing the logic of the virtual intelligent agent; Physical Node P2 (crashed): This is a physical computing node that has failed, indicating that it can no longer function properly; Physical Node P3 (idle / standby): This is a standby physical computing node in a ready state, prepared to take over the work of the crashed physical computing node at any time.

[0067] exist Figure 2 In this context, there are three different types of connections between the dynamic mapping router and the physical computing cluster, each representing one of three operational states: 1. Attach: A solid line connection from the router to the physical node P1 indicates that the router has successfully assigned some virtual sessions to P1, and P1 is performing computational tasks for the virtual agents corresponding to these virtual sessions.

[0068] 2. Detach (Heartbeat Loss State): The dashed connection between the router and physical node P2 indicates that the router has detected the loss of heartbeat signal from P2 (i.e., P2 has crashed) and is unmounting the virtual session on P2 and disconnecting the mapping relationship.

[0069] 3. Re-attach (State Recovery): A solid line connection from the router to the physical node P3 indicates that the router is reattaching the virtual sessions that were unattached from P2 to the idle P3. P3 will read the latest state data from the distributed memory grid and take over from P2 to continue providing computing services for these virtual agents.

[0070] In its implementation, after system startup, the virtual session layer creates a virtual session for each virtual agent and assigns these virtual sessions to different physical computing nodes via a dynamic mapping router. Assume virtual session 1 and virtual session 2 are assigned to physical node P1, and virtual session N is assigned to physical node P2. During normal operation, the physical nodes continuously execute game logic calculations to determine the latest state data of the virtual agents. Simultaneously, the virtual session layer synchronously writes the latest state data to the corresponding storage units in the distributed memory grid at a preset frequency. The physical nodes send a heartbeat signal to the virtual session layer every preset frequency to indicate "I am still working normally." At some point during the simulation, physical node P2 crashes due to a memory leak or hardware failure. All computations running on P2 immediately cease, and P2 can no longer send heartbeat signals. The dynamic mapping router of the virtual session layer detects that it has not received heartbeat signals from P2 for several consecutive times and determines that P2 has crashed. The router immediately performs a Detach operation: unbinding all virtual sessions originally mapped to P2 (such as virtual session N) from P2 and marking them as "awaiting remounting." At this point, these virtual sessions are in a "paused" state—their logical operations are temporarily suspended, but their state data is securely stored in the distributed memory grid. The router then initiates a remount process: it checks which physical computing nodes in the physical computing cluster are idle and finds that P3 is in an idle standby state. The router remaps virtual session N to P3. Upon receiving the instruction, P3 immediately reads the latest state data corresponding to virtual session N from the distributed memory grid. After the data loading is complete, P3 begins to execute computational tasks for virtual session N, and the derivation process seamlessly continues from the last saved state point.

[0071] The entire process from failure occurrence to recovery takes only milliseconds to seconds, which is negligible for simulations that can last for days or even weeks. This is the complete working mechanism of the VSL virtual session decoupling and remounting architecture. Its core value lies in completely separating state storage (distributed memory grid), session management (virtual session layer), and computation execution (physical computing cluster), ensuring that the failure of any physical computing node will not lead to the loss of simulation data, thus truly achieving high reliability for large-scale, long-term simulations.

[0072] In an optional embodiment, after creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, for every two virtual agents among the multiple virtual agents, the following operations are performed: the spatial distance and social relationship between the first and second virtual agents in each pair of virtual agents are obtained in real time; the relationship weight is determined based on the spatial distance and social relationship; it is determined whether the relationship weight exceeds a preset weight threshold. If it does, a communication connection is established between the first and second virtual agents to enable the first and second virtual agents to communicate point-to-point.

[0073] In practical implementation, the aforementioned relationship weights are used to indicate the closeness of communication between the first and second virtual intelligent agents. Specifically, the closer the control distance between the first and second virtual intelligent agents in the game scene, the higher the weight of the relationship weight in the spatial dimension; the closer the social relationship between the first and second virtual intelligent agents (e.g., being in the same guild, having had multiple transactions, being friends, etc.), the higher the weight of the relationship weight in the social dimension.

[0074] The system will only establish a communication connection between two virtual agents when the relationship weight exceeds the preset weight threshold. Virtual agents with relationship weights below the preset weight threshold will not establish a direct connection to avoid unnecessary communication overhead.

[0075] In practical applications, all virtual intelligent agents in the game world are abstracted into a dynamic graph. ,in A collection of virtual intelligent agent nodes. for The set of communication edges between virtual agents at any given time. This set of communication edges indicates the communication connections between virtual agents.

[0076] First, define two intelligent agents. and The spatial distance in the game world is The distance of their social relationship in the game's social graph (such as whether they are in the same guild or have a history of transactions) is... Based on this, relation weights for:

[0077] in, These are the weighting coefficients. This is the spatial perception attenuation parameter. It applies if and only if the relation weight exceeds a preset weight threshold. At that time, and Establish a direct, memory-mapped, or lightweight RPC (Remote Procedure Call) point-to-point communication edge between them. In this way, virtual agents can bypass the global server and directly broadcast information (e.g., "Rare mineral vein discovered at coordinates (150, 200)" or "Buy iron ore for 50 gold coins") to other "nearby" or "closely related" agents. This greatly reduces the complexity of global communication, enabling the system to support concurrent inference by hundreds of thousands of agents.

[0078] In one specific embodiment, when two virtual agents move to the same game area or join the same guild, causing the relationship weight between them to exceed a preset weight threshold, the system will create a new communication edge between them. When the two virtual agents leave each other's range or leave their shared guild, causing the relationship weight to drop below the preset relationship threshold, the system will disconnect the communication edge between them. Specifically, even if the communication edge always exists, the relationship weight will be dynamically updated as the distance between the two virtual agents changes and their social relationship evolves.

[0079] To simulate high-frequency transactions and information transmission among real players in the game's economic system simulation, this application designs an A2A dynamic topology communication mesh. For example... Figure 3 The diagram shown is an A2A dynamic topology communication grid state diagram provided in an embodiment of this disclosure. The A2A dynamic topology communication grid state diagram shows two different regions in a game world, and how virtual agents in these regions communicate directly in a peer-to-peer (P2P) manner without going through a traditional centralized message queue. Figure 3 The document also specifically showcases a decommissioned global message queue (MQ) center to illustrate the fundamental difference between the communication method disclosed herein and traditional methods.

[0080] Figure 3Game area A (Main City Trading Area) simulates a bustling commercial trading area in the game, much like a pedestrian shopping street or a farmers' market in the real world. This area contains three virtual agents: Agent 1 (Merchant): This is a virtual player playing the role of a merchant. In the game's progression, its behavior pattern is to collect and hoard various resources, then sell them on the market for profit. Its large language model gives it the personality of a "merchant": adept at analyzing market conditions and pursuing profit maximization. Agent 2 (Buyer): This is a virtual player playing the role of an ordinary buyer. It needs to purchase various resources from the market to complete its game objectives (such as buying weapons to fight monsters, buying ingredients to cook). Agent 3 (Buyer): Another virtual player playing the role of a buyer, similar to Agent 2, with its own purchasing needs.

[0081] In game area A, the following communication connections exist: Between Agent1 and Agent2: High-frequency P2P transaction communication (represented by a thick solid double-headed arrow). This is a frequently used direct communication channel where merchants and buyers need to exchange transaction information frequently: merchants post product prices, buyers send purchase requests, and both parties negotiate prices. This communication is real-time and high-frequency, hence its importance and frequency of use are indicated by thick lines. Between Agent1 and Agent3: High-frequency P2P transaction communication (also represented by a thick solid double-headed arrow). Similar to the above, the merchant also maintains high-frequency transaction communication with another buyer. Between Agent2 and Agent3: Spatial proximity broadcasting (represented by a thin dashed unidirectional arrow). Since these two buyers are in the same trading area, a weaker communication connection exists between them: spatial proximity broadcasting. This means that when Agent2 learns that a product has been discounted, this information may automatically "spread" to the nearby Agent3, simulating the information diffusion phenomenon in reality where "if a stall in a market lowers its price, people nearby will also notice."

[0082] Figure 3Game area B (Outdoor Combat Zone) simulates an outdoor combat zone far from the city, where guild wars or monster hunting activities are taking place. This area also features three virtual agents: Agent4 (Guild Leader): A virtual player acting as the guild leader, directing guild members in coordinated combat. Agent5 (Guild Member): An ordinary member under the Guild Leader, receiving instructions and executing specific combat missions. Agent6 (Random Person): A lone wolf not belonging to any guild, simply passing through the area. Within game area B, the following communication connections exist: Between Agent4 and Agent5: Social Topology Communication (represented by a thick solid double-headed arrow). This communication edge is not based on spatial distance but on social relationships, as they belong to the same guild. Even if they are temporarily separated on the battlefield, the guild channel remains open, allowing the guild leader to issue combat commands to members at any time. Between Agent5 and Agent6: Spatial Proximity Warning (represented by a thin dashed single-headed arrow). Although Agent6 is not a guild member, Agent5 might send it some spatial proximity broadcast messages, such as "Dangerous monster ahead, be careful," because it happens to be nearby. Cross-regional communication: between Agent1 and Agent4.

[0083] Figure 3 The example also demonstrates a communication connection spanning two regions: a "cross-region private chat (social topology)" connection (represented by a thin dashed line) exists between Agent1, a merchant in the main city, and Agent4, a guild leader in the wilderness. This illustrates that even if two virtual agents are in completely different geographical areas, as long as they have a social relationship (e.g., Agent1 is a supplier for Agent4's guild), the system will maintain a communication edge for them. Through this edge, Agent4 can directly contact Agent1 on the battlefield to say, "Send another batch of potions."

[0084] exist Figure 3 On one side, there is a "deprecated global message queue (MQ) center" marked with a gray dashed line. The connections between Agent1 and Agent4 and this center are crossed out (indicating the connection is broken). This design clearly expresses the core innovation of this disclosure: In a traditional architecture, if Agent1 wants to communicate with Agent4, the message must first be sent to the central message queue, and then forwarded to Agent4 by the center. However, in the A2A architecture of this disclosure, Agent1 can communicate directly with Agent4 point-to-point, completely bypassing the central node, significantly reducing communication latency and system bottleneck risks. When the simulation begins, the system calculates the relationship weight based on the spatial distance and social relationship between each pair of virtual players. For virtual agent pairs whose relationship weight exceeds a threshold, the system establishes a direct P2P communication edge.

[0085] In the main city's trading area, merchant Agent1 posts an ad: "Selling iron swords, 80 gold each." This ad is sent directly to nearby buyers Agent2 and Agent3 via a P2P communication edge. Agent2 decides to buy and replies via the same communication edge: "I want to buy an iron sword." The transaction is completed immediately, without any intermediaries. In the wilderness, guild leader Agent4 issues an order to Agent5 via the guild channel: "Focus fire on the BOSS's left flank." Simultaneously, Agent4 contacts merchant Agent1 via cross-server private chat: "We need supplies, please send 50 healing potions to the wilderness camp." These communications are all direct, point-to-point transactions. As the simulation progresses, virtual players move continuously. When a buyer leaves the main city's trading area for the wilderness, the spatial distance between them and the merchant increases, the relationship weight drops below a threshold, and the system automatically disconnects their trading communication edge. Simultaneously, after arriving in the wilderness, this buyer approaches nearby new players, and the system establishes a new communication edge for them. The entire communication network is thus constantly reorganized as virtual players move and their social relationships change.

[0086] By enabling point-to-point direct communication based on spatial proximity and social relationships, the traditional centralized message bus has been completely replaced, allowing tens of thousands of virtual intelligent agents to engage in efficient, self-organized social interactions just like real humans.

[0087] In one optional embodiment, the specific process by which the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent may include: acquiring the current state data and social information of the virtual agent through the physical computing node; wherein, the social information is external information obtained through point-to-point communication; converting the current state data and social information into natural language text, and inputting the natural language text into a preset large language model, so that the preset large language model performs reasoning analysis based on the natural language text to obtain an output result; generating a dynamic behavior tree based on the output result and the current state data; wherein, the dynamic behavior tree includes multiple actions to be executed arranged in a preset order; controlling the virtual agent to execute multiple actions to be executed sequentially in a preset order, and updating the current state data of the virtual agent after executing multiple actions to be executed to obtain the latest state data of the virtual agent.

[0088] In practical implementation, the aforementioned social information refers to various social and market information obtained by virtual agents from other virtual agents through the A2A dynamic topology communication network. This social information can include market information (e.g., dynamic commodity prices, supply and demand trends, abnormal transaction information, etc.), guild and organization information (e.g., guild commands, member dynamics, organizational resource status, etc.), regional security and threat information (e.g., information on the presence of hostile players, high-value targets, regional event information), information on changes in social relationships (e.g., changes in alliance and adversarial relationships, changes in personal relationships, reputation and evaluation, etc.), and social feedback on global events (e.g., the reactions and discussions of other virtual players when a global event occurs in the game).

[0089] In a specific implementation, the above output results include an intent vector and a target reward; where the intent vector is used to indicate the strategic decision direction, and the target reward is used to indicate the evaluation results of multiple behaviors under the strategic decision direction; based on this, the specific process of generating a dynamic behavior tree according to the output results and current state data can include: inputting the intent vector, target reward and current state data into a preset conditional policy network so that the conditional policy network generates a dynamic behavior tree.

[0090] In practical applications, the virtual agent transmits current state data at a low frequency (e.g., once per minute, or triggered when the environment undergoes drastic changes). and social information obtained from the A2A communication mesh The text is converted into natural language and input into a pre-defined large language model (LLM) deployed in the cloud. The pre-defined large language model (LLM) outputs a high-dimensional intent vector. and corresponding target rewards :

[0091] For example, after analyzing the current market situation, LLM generates an intent vector of "give up fighting monsters and go to the main city to stockpile iron ore in preparation for the upcoming guild war".

[0092] The system receives the intent vector Subsequently, the system does not directly allow the pre-set large model to output specific operations (because the latency is too high and it is easy to produce illusions). Instead, the system utilizes a pre-trained reinforcement learning policy network. ,by As a condition, a dynamic behavior tree is generated or reassembled in real time. The generated dynamic behavior tree is then sent to the execution engine of the physical computing node to execute specific micro-actions at the highest frequency. (Such as finding a way, clicking on an NPC, or entering a transaction amount).

[0093] The optimization objective of the above conditional policy network is to maximize the expected cumulative reward, specifically calculated as follows:

[0094] in, It is a mathematical operator that finds the value that makes the following expression take the maximum value. This represents the expected value calculated under the conditional policy network. This summation symbol represents adding up the rewards from the current moment (k=0) to infinity in the future (k=∞). γ is a value between 0 and 1 (such as 0.99 or 0.95), called the discount factor. For the target reward, it accepts two inputs: state data at time t+k. and the action performed at time t+k Output a reward value.

[0095] The above formula is used to find the best set of decision parameters (θ). This allows the virtual player to maximize the average weighted cumulative reward (with higher weight for closer rewards and lower weight for more distant rewards) from the current moment to all future moments when the virtual player acts according to the strategy guided by this set of parameters, under the guidance of the macro-intention formulated by the large language model.

[0096] like Figure 4 The diagram shown is a dynamic behavior tree and two-layer decision architecture provided in this embodiment. The entire architecture is divided into three main areas from top to bottom: the top is the Virtual Session Layer (VSL) (responsible for providing data input), the middle is the Slow Thinking Center System2 (responsible for strategic decision-making), and the bottom is the Fast Thinking Center System1 (responsible for tactical execution). Data flows sequentially from top to bottom, forming a complete "perception-thinking-action" closed loop.

[0097] The Virtual Session Layer (VSL) is the starting point of the entire decision-making process. It provides two core types of data required for decision-making: Virtual Agent State Data: This is the complete state information of the current virtual player, stored in a database-shaped icon. It contains all the virtual agent's personal information—current coordinates, item list, gold coin quantity, quest progress, skill level, health, historical behavior records, etc. A2A Topology Communication Data: This is external information (i.e., social information) received by the virtual agent from other virtual agents through the A2A dynamic topology communication network. It includes market price dynamics, guild commands, the behavior of nearby players, regional security status, etc. The Slow Thinking Hub (System 2) (Cloud / Asynchronous) is the strategic brain in the decision-making architecture. It runs on a cloud server and operates at a low frequency (e.g., triggered once per minute, or triggered when significant environmental changes occur). It contains three sub-components: Prompt Generator: This is an information transformation module. Its function is to convert structured data (state data and social information from A2A communication output) from VSL into natural language text (prompt) that the large language model can understand. For example, it will convert structured data such as "Gold: 5000, Iron Ore: 100, Location: Main City, Nearby Commodity Price Increase Signal: true" into a natural language description like "You are a merchant, currently in the main city, you have 5000 gold coins and 100 iron ore. You have observed that the price of enhancement stones in the nearby market is rising, and your guild leader has announced that there will be a guild war next week. Please analyze the current situation and formulate your next strategy."

[0098] Large Language Model (LLM): This is the core of System2—an AI language model similar to ChatGPT. It receives natural language descriptions from the Prompt generator, performs deep thinking, and then outputs strategic-level decision-making recommendations. The advantage of the Large Language Model lies in its human-like reasoning, analysis, and judgment capabilities, enabling it to understand complex market dynamics and social relationships, and make high-level strategic judgments.

[0099] Intent Vector I_t and Target Reward R_target: These are the outputs of the large language model. I_t is a high-dimensional intent vector that summarizes the strategic decision direction made by the LLM (e.g., "hoarding enhancement stones for guild wars"). R_target is the target reward associated with this strategic direction, used to guide the reinforcement learning network below in evaluating the merits of each specific action (e.g., "buying enhancement stones earns positive points, selling enhancement stones earns negative points"). The Fast Thinking Hub System 1 (physical node / real-time) is the "execution brain" in the decision-making architecture. It runs on physical computing nodes and operates at an extremely high frequency (60 times per second, synchronized with the game's frame rate). It contains three sub-components: Conditional Policy Network: This is a pre-trained reinforcement learning neural network. It receives two inputs: (1) macro-intention I_t and target reward R_target from System 2; (2) real-time state data S_i from VSL. Then, the policy network generates specific action plans in real time based on these inputs. It works very fast—it can make decisions in milliseconds, meeting the real-time requirements of games. Dynamic Behavior Tree DBT: This is the output of the policy network—a behavior tree describing specific action steps. Behavior trees are a classic action organization method in the field of game AI. They decompose complex behaviors into a series of orderly steps. For example, the macro-intention of "hoarding enhancement stones" will be decomposed into a behavior tree: "First open the backpack to check the amount of gold coins → go to the auction house → search for the list of enhancement stones → if the price is lower than the target price, buy → if there are not enough gold coins, sell the iron ore first → repeat the above steps". This behavior tree is "dynamic"—it will be adjusted according to the real-time changes in the environment. Micro-action sequence a_t: This is the final output of the behavior tree after execution—the specific operation actions actually performed by the virtual player in each frame. Examples include "move to coordinates (150, 200)," "click the NPC dialogue button," "enter the purchase quantity 50 in the input box," and "click to confirm purchase." These are the lowest-level operation commands that can be directly sent to the game engine for execution.

[0100] The Virtual Session Layer (VSL) continuously collects two types of data: (1) the latest state of the virtual agent itself (a state snapshot from the distributed memory grid); and (2) social information from other players received through the A2A communication grid. The Prompt generator reads these two types of data from the VSL and assembles them into a structured natural language text. The assembled natural language text is sent to a large language model in the cloud for processing. The large language model performs in-depth analysis and reasoning on the current situation and outputs macro-intention and target reward. This process may take hundreds of milliseconds to several seconds, but because its triggering frequency is very low (once per minute or event-triggered), it will not affect the real-time performance of the system. The intention vector and target reward are passed to the conditional policy network on the physical node. The conditional policy network combines the real-time context state to generate a dynamic behavior tree at a very high speed. The dynamic behavior tree is executed frame by frame, and each frame (every 1 / 60 second) outputs a specific micro-action, which is sent to the game engine for execution. After the micro-action is executed, the game environment will change accordingly (e.g., buying enhancement stones or reducing gold coins). These changes will update the context state of the virtual player, forming a closed loop, and the new state will affect the decision-making in the next round. exist Figure 4 In this context, this feedback is represented by a dashed arrow that returns from the "micro-action sequence" to the "context state (equivalent to the state data mentioned above)".

[0101] The advantage of this two-layer design lies in the fact that while large language models possess high intelligence, their computational speed is slow (each reasoning session takes several seconds), failing to meet the real-time requirement of 60 frames per second in games. Meanwhile, reinforcement learning networks, although extremely fast (millisecond-level), lack high-level strategic thinking capabilities. This disclosure cleverly combines the two, leveraging their strengths and compensating for their weaknesses, enabling the virtual intelligent agent to possess both human-like deep thinking abilities and machine-like rapid reaction capabilities.

[0102] In one optional implementation, the above-mentioned test criteria include: the deviation of the test indicators from the health indicators is less than a preset tolerance value, and the measurement index determined based on the test indicators is less than a preset value; wherein, the measurement index is used to indicate the risk of test failure. The measurement index here may be the Lyapunov index.

[0103] Based on the above description, the specific process for determining whether a test indicator meets the test standard includes: determining a measurement index based on the test indicator, and determining the deviation value between the test indicator and the health indicator; determining whether the deviation value is less than a preset tolerance value and whether the measurement index is less than a preset value; if so, determining that the test indicator meets the test standard.

[0104] If the conditions are met, determine whether the test duration of the game testing environment has reached the preset duration threshold. If yes, confirm that the game configuration parameter test has passed; otherwise, continue executing the step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent, until the test duration reaches the preset duration threshold. The preset duration threshold can be determined according to development needs.

[0105] If the conditions are not met, a freeze command is broadcast to multiple virtual agents to stop the game simulation test; the crash chain that caused the test to fail is identified, and parameter optimization calculations are performed based on the crash chain to obtain new game configuration parameters; the new game configuration parameters are expected to fix the crash chain; the game simulation test continues in the game test environment generated based on the new game configuration parameters.

[0106] Specifically, the process of continuing game simulation testing in the game test environment generated based on the new game configuration parameters may include: reverting the state data of multiple virtual agents to the first state data corresponding to the safe time before the collapse link occurred; loading the new game configuration parameters and unfreezing the command, so that the multiple virtual agents can re-perform game simulation testing based on the first state data.

[0107] In practical implementation, when the test metrics fail to meet the test standard system, a "global freeze" command is broadcast to all virtual agents, immediately suspending all simulation activities. The crash path is then extracted, and Bayesian backpropagation is performed to calculate a new set of game configuration parameters C_new, expected to fix the problem. The system utilizes historical state data stored in a distributed memory grid to roll back the state data of all virtual agents to a safe point in time before the crash. Then, the system loads the new game configuration parameters C_new (modified drop rates, transaction tax rates, etc.), lifts the global freeze, and allows all virtual agents to resume activity under the new rules, continuing simulations under the new parameters to verify whether the repaired parameters can maintain the long-term stability of the system.

[0108] This disclosure introduces phase space theory and Lyapunov exponents from physics into the game's numerical monitoring system, constructing an automated closed loop of "abnormal circuit breaking - root cause analysis - reverse optimization - state rollback - continued deduction". This enables the game to proactively capture and automatically repair fatal design flaws such as rampant inflation, power imbalances, and ecosystem disruptions during the critical stages from project initiation to official launch.

[0109] In summary, this disclosure not only solves the problems of computational bottleneck and state volatility in large-scale game automated testing, but also provides a highly adaptive and robust industrial-grade solution for the numerical balancing of modern complex game systems, demonstrating significant theoretical innovation and enormous commercial application value.

[0110] Corresponding to the above method embodiments, this disclosure also provides a game deduction testing device, such as... Figure 5 As shown, the device includes: Initialization module 50 is used to respond to game simulation test commands, load game configuration parameters and test standards, and initialize multiple virtual agents to join the game test environment generated by the game configuration parameters.

[0111] The session mounting module 51 is used to create a virtual session for each virtual agent among multiple virtual agents, and mount the virtual session to the physical computing node so that the physical computing node can perform game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent.

[0112] The result judgment module 52 is used to obtain test indicators based on the latest state data of multiple virtual intelligent agents and to determine whether the test indicators meet the test standards.

[0113] The test result determination module 53 is used to determine whether the game configuration parameter test is passed if the conditions are met; otherwise, it performs an exception circuit breaker and generates new game configuration parameters to continue the game simulation test in the game test environment generated based on the new game configuration parameters.

[0114] The aforementioned game simulation testing device uses a virtual agent as a lightweight virtual session, which is dynamically mounted on physical computing nodes. This achieves complete decoupling of physical computing resources and the logical state of the agent, as well as multiplexing of physical computing nodes, thereby improving the efficiency and quality of game simulation testing.

[0115] Furthermore, the aforementioned session mounting module 51 is used to: create a corresponding virtual session for the virtual agent in a preset virtual session layer, and mount the virtual session to the corresponding physical computing node through a dynamic mapping router.

[0116] Furthermore, the aforementioned session mounting module 51 is also used to: obtain the running status data corresponding to multiple preset computing nodes; determine the physical computing node configured for the virtual session from multiple preset computing nodes based on the game area where the virtual session is located and the running status data, and mount the virtual session to the corresponding physical computing node.

[0117] Furthermore, the above-mentioned device also includes a data storage module, used to: after creating a virtual session corresponding to a virtual agent, store the state data of the virtual agent corresponding to the virtual session in a preset distributed memory.

[0118] Furthermore, the aforementioned session mounting module 51 is also used for: the physical computing node to obtain the current state data of the virtual agent from the distributed memory, and to perform game logic calculations based on the current state data to obtain the latest state data of the virtual agent.

[0119] Furthermore, the aforementioned device also includes a data synchronization module, used to: after the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent, synchronously write the final state data of the virtual agent into the distributed memory at a preset frequency.

[0120] Furthermore, the aforementioned device also includes a remounting module, used for: after creating a virtual session corresponding to a virtual agent and mounting the virtual session to a physical computing node, determining whether the physical computing node is malfunctioning; if the physical computing node is malfunctioning, interrupting the game logic calculation performed on the physical computing node and demounting the virtual session from the physical computing node; re-determining a target computing node from multiple preset computing nodes and remounting the virtual session to the target computing node; if the physical computing node is functioning normally, executing the step of obtaining test indicators based on the latest status data corresponding to multiple virtual agents.

[0121] Furthermore, the aforementioned device also includes a logic calculation module, used to: after remounting the virtual session to the target computing node, obtain the latest state data of the virtual agent corresponding to the virtual session from the distributed memory through the target computing node, perform game logic calculations based on the latest state data, and update the latest state data of the virtual agent.

[0122] Furthermore, the aforementioned device also includes a communication establishment module, used to: after creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, perform the following operations for every two virtual agents among the multiple virtual agents: obtain in real time the spatial distance and social relationship between the first virtual agent and the second virtual agent in every two virtual agents; determine the relationship weight based on the spatial distance and social relationship; determine whether the relationship weight exceeds a preset weight threshold; if it does, establish a communication connection between the first virtual agent and the second virtual agent so that the first virtual agent and the second virtual agent can communicate point-to-point.

[0123] Furthermore, the aforementioned session mounting module 51 is also used to: acquire the current state data and social information of the virtual intelligent agent through a physical computing node; wherein the social information is external information acquired through point-to-point communication; convert the current state data and social information into natural language text, and input the natural language text into a preset large language model, so that the preset large language model performs reasoning analysis based on the natural language text to obtain output results; generate a dynamic behavior tree based on the output results and the current state data; wherein the dynamic behavior tree includes multiple actions to be executed arranged in a preset order; control the virtual intelligent agent to execute multiple actions to be executed in a preset order, and after executing multiple actions to be executed, update the current state data of the virtual intelligent agent to obtain the latest state data of the virtual intelligent agent.

[0124] Furthermore, the above output includes an intent vector and a target reward; wherein the intent vector is used to indicate the strategic decision direction, and the target reward is used to indicate the evaluation results of multiple behaviors under the strategic decision direction; based on this, the above session mounting module 51 is also used to: input the intent vector, target reward and current state data into a preset conditional policy network, so that the conditional policy network generates a dynamic behavior tree.

[0125] Furthermore, the aforementioned test criteria include: the deviation of the test indicator from the health indicator is less than a preset tolerance value, and the measurement index determined based on the test indicator is less than a preset value; wherein, the measurement index is used to indicate the risk of test failure; based on this, the aforementioned result judgment module 52 is used to: determine the measurement index based on the test indicator, and determine the deviation value between the test indicator and the health indicator; determine whether the deviation value is less than the preset tolerance value, and whether the measurement index is less than the preset value; if so, determine that the test indicator meets the test criteria.

[0126] Furthermore, the test result determination module 53 is used to: if satisfied, determine whether the test duration of the game test environment has reached the preset duration threshold; if yes, determine that the game configuration parameter test has passed; otherwise, continue to execute the step of the physical calculation node performing game logic calculation based on the current state data of the virtual intelligent agent to obtain the latest state data of the virtual intelligent agent, until the test duration reaches the preset duration threshold.

[0127] Furthermore, the aforementioned test result determination module 53 is also used to: broadcast a freeze command to multiple virtual agents so that all virtual agents stop the game simulation test; identify the crash link that caused the test to fail, and perform parameter optimization calculations based on the crash link to obtain new game configuration parameters; wherein the new game configuration parameters are expected to be able to repair the crash link; and continue the game simulation test in the game test environment generated based on the new game configuration parameters.

[0128] Furthermore, the aforementioned test result determination module 53 is also used to: revert the state data of multiple virtual agents to the first state data corresponding to the safe time before the collapse link occurred; load new game configuration parameters and unfreeze the command, so that multiple virtual agents can re-perform game simulation tests based on the first state data.

[0129] The game simulation testing device provided in this disclosure has the same implementation principle and technical effects as the aforementioned method embodiments. For the sake of brevity, any parts not mentioned in the device embodiments can be referred to the corresponding content in the aforementioned method embodiments.

[0130] This embodiment also provides an electronic device, such as... Figure 6 As shown, the electronic device includes a processor and a memory. The memory stores computer-executable instructions that can be executed by the processor. The processor executes the computer-executable instructions to implement the aforementioned game simulation testing method. This electronic device can be a server or a terminal device.

[0131] Specifically, the aforementioned game simulation testing method includes: responding to a game simulation testing command, loading game configuration parameters and testing standards, and initializing multiple virtual agents to join the game testing environment generated by the game configuration parameters; for each of the multiple virtual agents, creating a virtual session corresponding to the virtual agent, and attaching the virtual session to a physical computing node, so that the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent; obtaining test indicators based on the latest state data corresponding to the multiple virtual agents, and determining whether the test indicators meet the testing standards; if they meet the standards, determining that the game configuration parameter test has passed; if they do not meet the standards, performing an exception circuit breaker and generating new game configuration parameters, so that the game simulation test can continue in the game testing environment generated based on the new game configuration parameters.

[0132] The aforementioned game simulation testing method uses a virtual agent as a lightweight virtual session, which is dynamically mounted on physical computing nodes. This achieves complete decoupling of physical computing resources and the logical state of the agent, as well as multiplexing of physical computing nodes, thereby improving the efficiency and quality of game simulation testing.

[0133] In an optional embodiment, the steps of creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node include: creating a corresponding virtual session for the virtual agent in a preset virtual session layer, and attaching the virtual session to the corresponding physical computing node through a dynamic mapping router.

[0134] In an optional embodiment, the step of attaching the virtual session to the corresponding physical computing node includes: obtaining the running status data corresponding to multiple preset computing nodes: determining the physical computing node configured for the virtual session from multiple preset computing nodes based on the game area where the virtual session is located and the running status data, and attaching the virtual session to the corresponding physical computing node.

[0135] In an optional embodiment, after the step of creating a virtual session corresponding to a virtual agent, the method further includes storing the state data of the virtual agent corresponding to the virtual session in a preset distributed memory.

[0136] In an optional embodiment, the step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent includes: the physical computing node obtaining the current state data of the virtual agent from the distributed memory, and performing game logic calculations based on the current state data to obtain the latest state data of the virtual agent.

[0137] In an optional embodiment, after the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent, the method further includes: synchronously writing the final state data of the virtual agent into the distributed memory at a preset frequency.

[0138] In an optional embodiment, after creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, the method further includes: determining whether the physical computing node is malfunctioning; if the physical computing node is malfunctioning, interrupting the game logic calculation performed on the physical computing node and removing the attachment state between the virtual session and the physical computing node; re-determining a target computing node from multiple preset computing nodes and re-attaching the virtual session to the target computing node; if the physical computing node is functioning normally, executing the step of obtaining test indicators based on the latest state data corresponding to multiple virtual agents.

[0139] In an optional embodiment, after the step of remounting the virtual session to the target computing node, the method further includes: obtaining the latest state data of the virtual agent corresponding to the virtual session from the distributed memory through the target computing node, performing game logic calculations based on the latest state data, and updating the latest state data of the virtual agent.

[0140] In an optional embodiment, after creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, the method further includes: for every two virtual agents among the plurality of virtual agents, performing the following operations: acquiring in real time the spatial distance and social relationship between the first virtual agent and the second virtual agent in each pair of virtual agents; determining the relationship weight based on the spatial distance and social relationship; determining whether the relationship weight exceeds a preset weight threshold; if it exceeds, establishing a communication connection between the first virtual agent and the second virtual agent to enable the first virtual agent and the second virtual agent to perform point-to-point communication.

[0141] In an optional embodiment, the step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent includes: acquiring the current state data and social information of the virtual agent through the physical computing node; wherein, the social information is external information obtained through point-to-point communication; converting the current state data and social information into natural language text, and inputting the natural language text into a preset large language model, so that the preset large language model performs reasoning analysis based on the natural language text to obtain output results; generating a dynamic behavior tree based on the output results and the current state data; wherein, the dynamic behavior tree includes multiple actions to be executed arranged in a preset order; controlling the virtual agent to execute multiple actions to be executed sequentially in a preset order, and updating the current state data of the virtual agent after executing multiple actions to be executed to obtain the latest state data of the virtual agent.

[0142] In an optional embodiment, the above output results include an intent vector and a target reward; wherein, the intent vector is used to indicate the strategic decision direction, and the target reward is used to indicate the evaluation results of multiple behaviors under the strategic decision direction; based on this, the above step of generating a dynamic behavior tree according to the output results and current state data includes: inputting the intent vector, target reward and current state data into a preset conditional policy network so that the conditional policy network generates a dynamic behavior tree.

[0143] In an optional embodiment, the above-mentioned test standard includes: the deviation of the test indicator from the health indicator is less than a preset fault tolerance value, and the measurement index determined based on the test indicator is less than a preset value; wherein, the measurement index is used to indicate the risk of test failure; based on this, the above-mentioned step of determining whether the test indicator meets the test standard includes: determining the measurement index based on the test indicator, and determining the deviation value between the test indicator and the health indicator; determining whether the deviation value is less than the preset fault tolerance value, and whether the measurement index is less than the preset value; if so, determining that the test indicator meets the test standard.

[0144] In an optional embodiment, the step of determining that the game configuration parameter test has passed if the conditions are met includes: if the conditions are met, determining whether the test duration of the game test environment has reached a preset duration threshold; if yes, determining that the game configuration parameter test has passed; otherwise, continuing to execute the step of the physical computing node performing game logic calculations based on the current state data of the virtual intelligent agent to obtain the latest state data of the virtual intelligent agent, until the test duration reaches the preset duration threshold.

[0145] In an optional embodiment, the steps of performing abnormal circuit breaking and generating new game configuration parameters to continue game simulation testing in a game test environment generated based on the new game configuration parameters include: broadcasting a freeze command to multiple virtual agents to stop the game simulation testing; identifying the crash chain that caused the test to fail, and performing parameter optimization calculations based on the crash chain to obtain new game configuration parameters; wherein the new game configuration parameters are expected to be able to repair the crash chain; and continuing game simulation testing in a game test environment generated based on the new game configuration parameters.

[0146] In an optional embodiment, the step of continuing the game simulation test in the game test environment generated according to the new game configuration parameters includes: reverting the state data of multiple virtual agents to the first state data corresponding to the safe time before the collapse link occurred; loading the new game configuration parameters and unfreezing the command, so that the multiple virtual agents can re-perform the game simulation test based on the first state data.

[0147] Furthermore, Figure 6 The electronic device shown also includes a bus 102 and a communication interface 103, with the processor 101, the communication interface 103 and the memory 100 connected via the bus 102.

[0148] The memory 100 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 102 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0149] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. The processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 100, and processor 101 reads information from memory 100 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0150] This disclosure also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the above-described game simulation testing method. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0151] Specifically, the aforementioned game simulation testing method includes: responding to a game simulation testing command, loading game configuration parameters and testing standards, and initializing multiple virtual agents to join the game testing environment generated by the game configuration parameters; for each of the multiple virtual agents, creating a virtual session corresponding to the virtual agent, and attaching the virtual session to a physical computing node, so that the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent; obtaining test indicators based on the latest state data corresponding to the multiple virtual agents, and determining whether the test indicators meet the testing standards; if they meet the standards, determining that the game configuration parameter test has passed; if they do not meet the standards, performing an exception circuit breaker and generating new game configuration parameters, so that the game simulation test can continue in the game testing environment generated based on the new game configuration parameters.

[0152] The aforementioned game simulation testing method uses a virtual agent as a lightweight virtual session, which is dynamically mounted on physical computing nodes. This achieves complete decoupling of physical computing resources and the logical state of the agent, as well as multiplexing of physical computing nodes, thereby improving the efficiency and quality of game simulation testing.

[0153] In an optional embodiment, the steps of creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node include: creating a corresponding virtual session for the virtual agent in a preset virtual session layer, and attaching the virtual session to the corresponding physical computing node through a dynamic mapping router.

[0154] In an optional embodiment, the step of attaching the virtual session to the corresponding physical computing node includes: obtaining the running status data corresponding to multiple preset computing nodes: determining the physical computing node configured for the virtual session from multiple preset computing nodes based on the game area where the virtual session is located and the running status data, and attaching the virtual session to the corresponding physical computing node.

[0155] In an optional embodiment, after the step of creating a virtual session corresponding to a virtual agent, the method further includes storing the state data of the virtual agent corresponding to the virtual session in a preset distributed memory.

[0156] In an optional embodiment, the step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent includes: the physical computing node obtaining the current state data of the virtual agent from the distributed memory, and performing game logic calculations based on the current state data to obtain the latest state data of the virtual agent.

[0157] In an optional embodiment, after the physical computing node performs game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent, the method further includes: synchronously writing the final state data of the virtual agent into the distributed memory at a preset frequency.

[0158] In an optional embodiment, after creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, the method further includes: determining whether the physical computing node is malfunctioning; if the physical computing node is malfunctioning, interrupting the game logic calculation performed on the physical computing node and removing the attachment state between the virtual session and the physical computing node; re-determining a target computing node from multiple preset computing nodes and re-attaching the virtual session to the target computing node; if the physical computing node is functioning normally, executing the step of obtaining test indicators based on the latest state data corresponding to multiple virtual agents.

[0159] In an optional embodiment, after the step of remounting the virtual session to the target computing node, the method further includes: obtaining the latest state data of the virtual agent corresponding to the virtual session from the distributed memory through the target computing node, performing game logic calculations based on the latest state data, and updating the latest state data of the virtual agent.

[0160] In an optional embodiment, after creating a virtual session corresponding to a virtual agent and attaching the virtual session to a physical computing node, the method further includes: for every two virtual agents among the plurality of virtual agents, performing the following operations: acquiring in real time the spatial distance and social relationship between the first virtual agent and the second virtual agent in each pair of virtual agents; determining the relationship weight based on the spatial distance and social relationship; determining whether the relationship weight exceeds a preset weight threshold; if it exceeds, establishing a communication connection between the first virtual agent and the second virtual agent to enable the first virtual agent and the second virtual agent to perform point-to-point communication.

[0161] In an optional embodiment, the step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent includes: acquiring the current state data and social information of the virtual agent through the physical computing node; wherein, the social information is external information obtained through point-to-point communication; converting the current state data and social information into natural language text, and inputting the natural language text into a preset large language model, so that the preset large language model performs reasoning analysis based on the natural language text to obtain output results; generating a dynamic behavior tree based on the output results and the current state data; wherein, the dynamic behavior tree includes multiple actions to be executed arranged in a preset order; controlling the virtual agent to execute multiple actions to be executed sequentially in a preset order, and updating the current state data of the virtual agent after executing multiple actions to be executed to obtain the latest state data of the virtual agent.

[0162] In an optional embodiment, the above output results include an intent vector and a target reward; wherein, the intent vector is used to indicate the strategic decision direction, and the target reward is used to indicate the evaluation results of multiple behaviors under the strategic decision direction; based on this, the above step of generating a dynamic behavior tree according to the output results and current state data includes: inputting the intent vector, target reward and current state data into a preset conditional policy network so that the conditional policy network generates a dynamic behavior tree.

[0163] In an optional embodiment, the above-mentioned test standard includes: the deviation of the test indicator from the health indicator is less than a preset fault tolerance value, and the measurement index determined based on the test indicator is less than a preset value; wherein, the measurement index is used to indicate the risk of test failure; based on this, the above-mentioned step of determining whether the test indicator meets the test standard includes: determining the measurement index based on the test indicator, and determining the deviation value between the test indicator and the health indicator; determining whether the deviation value is less than the preset fault tolerance value, and whether the measurement index is less than the preset value; if so, determining that the test indicator meets the test standard.

[0164] In an optional embodiment, the step of determining that the game configuration parameter test has passed if the conditions are met includes: if the conditions are met, determining whether the test duration of the game test environment has reached a preset duration threshold; if yes, determining that the game configuration parameter test has passed; otherwise, continuing to execute the step of the physical computing node performing game logic calculations based on the current state data of the virtual intelligent agent to obtain the latest state data of the virtual intelligent agent, until the test duration reaches the preset duration threshold.

[0165] In an optional embodiment, the steps of performing abnormal circuit breaking and generating new game configuration parameters to continue game simulation testing in a game test environment generated based on the new game configuration parameters include: broadcasting a freeze command to multiple virtual agents to stop the game simulation testing; identifying the crash chain that caused the test to fail, and performing parameter optimization calculations based on the crash chain to obtain new game configuration parameters; wherein the new game configuration parameters are expected to be able to repair the crash chain; and continuing game simulation testing in a game test environment generated based on the new game configuration parameters.

[0166] In an optional embodiment, the step of continuing the game simulation test in the game test environment generated according to the new game configuration parameters includes: reverting the state data of multiple virtual agents to the first state data corresponding to the safe time before the collapse link occurred; loading the new game configuration parameters and unfreezing the command, so that the multiple virtual agents can re-perform the game simulation test based on the first state data.

[0167] The aforementioned function, when implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal device, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0168] In the description of this disclosure, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this disclosure and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this disclosure. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0169] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, 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 this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A game simulation testing method, characterized in that, The method includes: In response to the game simulation test command, the game configuration parameters and test standards are loaded, and multiple virtual intelligent agents are initialized to join the game test environment generated by the game configuration parameters; For each of the plurality of virtual intelligent agents, a virtual session corresponding to the virtual intelligent agent is created, and the virtual session is attached to the physical computing node, so that the physical computing node performs game logic calculations based on the current state data of the virtual intelligent agent to obtain the latest state data of the virtual intelligent agent; Based on the latest state data corresponding to the multiple virtual intelligent agents, test indicators are obtained, and it is determined whether the test indicators meet the test standards. If the conditions are met, the game configuration parameter test is considered passed; if not, an abnormal circuit breaker is triggered and new game configuration parameters are generated to continue the game simulation test in the game test environment generated based on the new game configuration parameters.

2. The method according to claim 1, characterized in that, The step of creating a virtual session corresponding to the virtual agent and attaching the virtual session to the physical computing node includes: In the preset virtual session layer, a corresponding virtual session is created for the virtual agent, and the virtual session is mounted to the corresponding physical computing node through a dynamic mapping router.

3. The method according to claim 2, characterized in that, The step of attaching the virtual session to the corresponding physical computing node includes: Obtain the running status data corresponding to multiple preset computing nodes: Based on the game area where the virtual session is located and the running status data, the physical computing node configured for the virtual session is determined from the plurality of preset computing nodes, and the virtual session is mounted to the corresponding physical computing node.

4. The method according to any one of claims 1-3, characterized in that, After the step of creating the virtual session corresponding to the virtual agent, the method further includes: The state data of the virtual agent corresponding to the virtual session is stored in a preset distributed memory.

5. The method according to claim 4, characterized in that, The step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent includes: The physical computing node obtains the current state data of the virtual agent from the distributed memory, and performs game logic calculations based on the current state data to obtain the latest state data of the virtual agent.

6. The method according to claim 5, characterized in that, After the step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent, the method further includes: The final state data of the virtual intelligent agent is synchronously written into the distributed memory at a preset frequency.

7. The method according to claim 4, characterized in that, After the steps of creating a virtual session corresponding to the virtual agent and attaching the virtual session to the physical computing node, the method further includes: Determine whether the physical computing node is malfunctioning; If the physical computing node malfunctions, the game logic calculations performed on the physical computing node are interrupted, and the virtual session is unmounted from the physical computing node; a new target computing node is determined from multiple preset computing nodes, and the virtual session is remounted to the target computing node.

8. The method according to claim 7, characterized in that, After the step of remounting the virtual session to the target computing node, the method further includes: The target computing node retrieves the latest state data of the virtual agent corresponding to the virtual session from the distributed memory, performs game logic calculations based on the latest state data, and updates the latest state data of the virtual agent.

9. The method according to claim 1, characterized in that, After the steps of creating a virtual session corresponding to the virtual agent and attaching the virtual session to the physical computing node, the method further includes: For every two virtual agents among the plurality of virtual agents, perform the following operations: Real-time acquisition of the spatial distance and social relationship between the first and second virtual agents in every pair of virtual agents; The relationship weight is determined based on the spatial distance and the social relationship; Determine whether the relationship weight exceeds a preset weight threshold. If it does, establish a communication connection between the first virtual agent and the second virtual agent to enable point-to-point communication between them.

10. The method according to claim 9, characterized in that, The step of the physical computing node performing game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent includes: The current state data and social information of the virtual intelligent agent are obtained through the physical computing node; wherein, the social information is external information obtained through point-to-point communication. The current state data and social information are converted into natural language text, and the natural language text is input into a preset large language model so that the preset large language model can perform reasoning analysis based on the natural language text to obtain output results; A dynamic behavior tree is generated based on the output results and the current state data; wherein the dynamic behavior tree includes multiple actions to be executed arranged in a preset order; The virtual agent is controlled to execute the multiple pending actions in a preset order, and after executing the multiple pending actions, the current state data of the virtual agent is updated to obtain the latest state data of the virtual agent.

11. The method according to claim 10, characterized in that, The output includes an intent vector and a target reward; wherein the intent vector is used to indicate the direction of strategic decision-making, and the target reward is used to indicate the evaluation results of multiple behaviors under the direction of strategic decision-making. The step of generating a dynamic behavior tree based on the output result and the current state data includes: The intent vector, the target reward, and the current state data are input into a preset conditional policy network so that the conditional policy network generates the dynamic behavior tree.

12. The method according to claim 1, characterized in that, The testing criteria include: the deviation of the test indicators from the health indicators is less than a preset tolerance value, and the measurement index determined based on the test indicators is less than a preset value; wherein, the measurement index is used to indicate the risk of test failure. The step of determining whether the test index meets the test standard includes: A measurement index is determined based on the test indicators, and the deviation value between the test indicators and the health indicators is determined. Determine whether the deviation value is less than a preset tolerance value and whether the measurement index is less than the preset value; if so, determine that the test index meets the test standard.

13. The method according to claim 12, characterized in that, If the conditions are met, the step of determining that the game configuration parameter test has passed includes: If the conditions are met, determine whether the test duration of the game test environment has reached the preset duration threshold; If yes, the game configuration parameter test is confirmed to be passed; otherwise, the physical computing node continues to perform game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent, until the test duration reaches the preset duration threshold.

14. The method according to claim 12, characterized in that, The steps of performing abnormal circuit breaking and generating new game configuration parameters, so as to continue game simulation testing in a game test environment generated based on the new game configuration parameters, include: A freeze command is broadcast to the plurality of virtual agents so that all the virtual agents stop the game simulation test; Identify the crash chain that caused the test to fail, and perform parameter optimization calculations based on the crash chain to obtain new game configuration parameters; wherein, the new game configuration parameters are expected to be able to fix the crash chain; The game simulation test will continue in the game test environment generated based on the new game configuration parameters.

15. The method according to claim 14, characterized in that, The steps for continuing game simulation testing in the game test environment generated based on the new game configuration parameters include: The state data of the multiple virtual agents are reverted to the first state data corresponding to the safe time before the collapse link occurred; Load new game configuration parameters and unfreeze the command so that the multiple virtual agents can re-perform game simulation tests based on the first state data.

16. A game simulation testing device, characterized in that, The device includes: An initialization module is used to respond to game simulation test commands, load game configuration parameters and test standards, and initialize multiple virtual agents to join the game test environment generated by the game configuration parameters; The session mounting module is used to create a virtual session corresponding to each of the multiple virtual agents and mount the virtual session to the physical computing node, so that the physical computing node can perform game logic calculations based on the current state data of the virtual agent to obtain the latest state data of the virtual agent. The result judgment module is used to obtain test indicators based on the latest state data corresponding to the multiple virtual intelligent agents, and to determine whether the test indicators meet the test standards. The test result determination module is used to determine that the game configuration parameter test has passed if the conditions are met; otherwise, it performs an exception circuit breaker and generates new game configuration parameters to continue the game simulation test in the game test environment generated based on the new game configuration parameters.

Images