Game copy automatic test case generation method and terminal

Abstract

According to the game copy automatic test case generation method and the terminal, a copy digital twinning scene is constructed through an environment simulator, an intelligent agent is generated, and the intelligent agent makes a decision by using a dual-network mechanism of a layered DQN controller; training agents in a grading manner by utilizing dynamic course learning, gradually improving difficulty according to copy complexity and migrating low-level training weights, and guiding the agents to explore in combination with a multi-dimensional reward function; generating a test case according to the agent exploration trajectory; according to the method, the digital twinning scene is constructed through the environment simulator, and the intelligent agent autonomously explores and covers a hidden path and a dynamic mechanism based on layered DQN and dynamic course learning, so that the test coverage rate can be effectively improved, the effectiveness of test case generation is ensured, and the dynamic course learning can effectively improve the training efficiency and quickly adapt to copy updating; the defects of low random test efficiency and poor static algorithm adaptability are overcome, and high-coverage, efficient and self-adaptive test case generation is realized.

Description

Technical Field

[0001] The present invention relates to the field of game testing technology, and in particular to a method and terminal for generating automated test cases for game copies. Background Art

[0002] In MMO game testing, dungeon path coverage is a core challenge, especially for complex dungeons (such as the "Void Corridor" in the game "Magic World"), which may contain dynamic mechanisms and multi-BOSS linkage mechanisms.

[0003] Therefore, the traditional method has the following problems: (1) It relies on the experience of testers, has a low coverage rate (about 60%-70%), and cannot cover hidden paths (such as the "floating bridge mechanism trigger path").

[0004] (2) Random testing is inefficient and requires traversing a large number of invalid paths, which takes up to several weeks.

[0005] (3) Static algorithms have poor adaptability. Genetic algorithms and other algorithms are difficult to adapt to dynamic events (such as changes in BOSS skills). Use cases need to be redesigned after the copy is updated. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and terminal for automatically generating test cases for game copies, thereby improving the effectiveness, dynamic adaptability and efficiency of test case generation.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for generating automated test cases for a game copy, comprising the steps of: Build a replica digital twin scenario through the environment simulator and generate an intelligent agent that makes decisions using a dual-network mechanism of a hierarchical DQN controller; Use dynamic curriculum learning to hierarchically train agents, gradually increase difficulty based on instance complexity and migrate low-level training weights, and combine multi-dimensional reward functions to guide agent exploration; Generate test cases based on the agent's exploration trajectory.

[0008] In order to solve the above technical problems, another technical solution adopted by the present invention is: A terminal for generating automated test cases for game copies comprises a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for generating automated test cases for game copies are implemented.

[0009] The beneficial effects of the present invention are: a method and terminal for automatically generating test cases for game copies of the present invention constructs a digital twin scene through an environmental simulator, and the intelligent agent independently explores based on hierarchical DQN and dynamic course learning, covering hidden paths and dynamic mechanisms, which can effectively improve the test coverage and ensure the effectiveness of test case generation. Dynamic course learning can effectively improve training efficiency, quickly adapt to copy updates, solve the defects of low efficiency of random testing and poor adaptability of static algorithms, and achieve high-coverage, efficient and adaptive test case generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flowchart of a method for generating automated test cases for game copies according to an embodiment of the present invention; Figure 2 A structural diagram of a terminal for generating automated test cases for a game copy according to an embodiment of the present invention; Description of labels: 1. A terminal for generating automated test cases for game copies; 2. A processor; 3. A memory. DETAILED DESCRIPTION

[0011] To illustrate the technical content, achieved objectives and effects of the present invention in detail, the following description is given in conjunction with the embodiments and accompanying drawings.

[0012] Please refer to Figure 1 , a method for generating automated test cases for a game copy, comprising the steps of: Build a replica digital twin scenario through the environment simulator and generate an intelligent agent that makes decisions using a dual-network mechanism of a hierarchical DQN controller; Use dynamic curriculum learning to hierarchically train agents, gradually increase difficulty based on instance complexity and migrate low-level training weights, and combine multi-dimensional reward functions to guide agent exploration; Generate test cases based on the agent's exploration trajectory.

[0013] From the above description, it can be seen that the beneficial effects of the present invention are: a method for automatically generating test cases for game copies of the present invention constructs a digital twin scene through an environmental simulator, and the intelligent agent independently explores based on hierarchical DQN and dynamic course learning, covering hidden paths and dynamic mechanisms, which can effectively improve the test coverage and ensure the effectiveness of test case generation. Dynamic course learning can effectively improve training efficiency, quickly adapt to copy updates, solve the defects of low efficiency of random testing and poor adaptability of static algorithms, and achieve high-coverage, efficient and adaptive test case generation.

[0014] Furthermore, dynamic curriculum learning is used to train agents in a hierarchical manner, gradually increasing the difficulty according to the complexity of the instance and migrating low-level training weights, including: The complexity of the dungeon is quantified by the number of dungeon branch paths, the number of dynamic events, and the number of bosses. The difficulty level of the dungeon is divided according to the complexity of the dungeon. The agent is trained from the lowest difficulty level, and when the agent meets the preset upgrade conditions in a preset number of consecutive trainings, a more complex mechanism is introduced based on the optimal network weights obtained from the current level training to increase the difficulty level of the training and adjust the learning rate.

[0015] As can be seen from the above description, by quantifying complexity through the number of paths, events, and bosses, and conducting hierarchical training, the agent gradually transitions from simple scenarios (such as linear paths) to complex ones (such as multi-boss interactions), avoiding the ineffective trial and error of directly facing difficult scenarios. The establishment of upgrade conditions ensures a solid foundation of capabilities, reduces training difficulty, and improves exploration efficiency. Compared to traditional methods, this method reduces training time and covers more complex interaction mechanisms.

[0016] Furthermore, the quantitative calculation of the replica complexity is: ; in, Indicates the number of replica branch paths, Indicates the number of dynamic events, Indicates the number of BOSSes.

[0017] As can be seen from the above description, by scientifically quantifying instance complexity based on these calculations and clarifying the core impact of dynamic events on difficulty, the system can objectively assess instance challenge. During tiered training, the difficulty progression is precisely controlled based on these quantified results, avoiding biases in subjective difficulty assignments, ensuring that training difficulty matches agent capabilities, and improving training stability and efficiency.

[0018] Furthermore, the hierarchical DQN controller adopts a 5-layer fully connected network architecture, including an input layer, two hidden layers, an output layer, and a hidden layer activation function; Among them, the input layer is 256 dimensions, the hidden layer is 512 dimensions, the output layer is 64-dimensional action value, and the hidden layer activation function is LeakyReLU.

[0019] As can be seen from the preceding description, the five-layer fully connected network supports processing high-dimensional state inputs (such as 20×20 coverage heat maps and skill cooldown matrices), improving the agent's ability to extract features from complex scenarios (such as dynamic mechanism coordinates and boss skill ranges). This layered network structure enhances decision-making accuracy, making action selection (such as eight-way movement and skill release timing) more aligned with the dungeon mechanics and reducing ineffective operations.

[0020] Furthermore, the definition of the multi-dimensional reward function is: ; in, Indicates the newly added coverage area. represents the total area of ​​the replica, Indicates the number of times a dynamic event is triggered. Indicates the number of times the area is visited repeatedly. Indicates the blood loss ratio.

[0021] As can be seen from the above description, the multi-dimensional reward function guides the agent to prioritize exploring new areas (coverage reward) and triggering dynamic events (event reward), while also penalizing repeated paths (repetition penalty) and high health loss (survival penalty). This balances exploration breadth with survivability, avoiding blind traversal through random testing, and enabling the agent to efficiently explore hidden paths (such as the floating bridge trigger condition) and risky scenarios.

[0022] Furthermore, the dual network mechanism of the hierarchical DQN controller includes: The MainNet network is responsible for the agent's action decision-making in real time, and the TargetNet network outputs the target Q value of the DQN network. The target Q value is used to measure the quality of the MainNet network's current decision, thereby providing feedback to update the MainNet network. The MainNet network is trained for a first preset number of times, and the network weights are updated by a gradient descent algorithm; The TargetNet network synchronizes network weights from the MainNet once every time it performs a second preset number of trainings, where the second preset number of trainings is greater than the first preset number of trainings.

[0023] Furthermore, the first preset number is 4, and the second preset number is 1000.

[0024] As can be seen from the preceding description, the dual-network mechanism (MainNet updates every 4 steps, TargetNet synchronizes every 1000 steps) addresses the instability issues inherent in traditional DQN training by separating online decision-making from target calculation. MainNet optimizes its strategies in real time to cope with dynamic environments (such as boss skill changes), while TargetNet provides stable target Q values, reducing training oscillations and accelerating convergence. The frequency of weight synchronization is varied, balancing exploration and exploitation, improving learning efficiency.

[0025] Furthermore, in the hierarchical DQN controller, the definition of the state space includes the coordinates, health, skill cooldown, and coverage heat map of the agent; The action space is defined to include movement, skills, and interaction commands.

[0026] As can be seen from the preceding description, the state space includes coordinates, health, skill cooldowns, and a coverage heatmap, ensuring the agent's complete environmental awareness. The action space encompasses directional movement, skills, and interactions, supporting complex operations. The combination of these two allows the agent to simulate abnormal player behavior, addressing edge cases not covered by traditional methods and improving vulnerability discovery.

[0027] Furthermore, the agent uses the dual-network mechanism of the hierarchical DQN controller to make decisions, including: The replica digital twin scene collects the status information of the intelligent agent in real time, and transmits the status information to the intelligent agent using a preset communication protocol, and the intelligent agent makes decisions based on the status information.

[0028] As can be seen from the preceding description, real-time transmission of state information via a transport protocol ensures that agent decisions are based on the latest data and adapt to dynamic replica changes. This real-time interaction enables agents to adjust their strategies promptly, preventing the policy failures of traditional static algorithms caused by data lags and improving the system's ability to cover real-time mechanisms.

[0029] Please refer to Figure 2 A game copy automated test case generation terminal includes a processor, a memory, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the steps in the above-mentioned game copy automated test case generation method are implemented.

[0030] From the above description, it can be seen that the beneficial effects of the present invention are: a game copy automatic test case generation terminal of the present invention constructs a digital twin scene through an environmental simulator, and the intelligent agent independently explores based on hierarchical DQN and dynamic course learning, covering hidden paths and dynamic mechanisms, which can effectively improve the test coverage and ensure the effectiveness of test case generation. Dynamic course learning can effectively improve training efficiency, quickly adapt to copy updates, solve the defects of low efficiency of random testing and poor adaptability of static algorithms, and achieve high-coverage, efficient and adaptive test case generation.

[0031] A method and terminal for generating automated test cases for game copies of the present invention are suitable for generating automated test cases for MMO game copies.

[0032] Please refer to Figure 1 , embodiment 1 of the present invention is: A method for generating automated test cases for a game copy, comprising the steps of: A replica digital twin scenario is constructed through an environment simulator, and an intelligent agent is generated, which makes decisions using a dual-network mechanism of a hierarchical DQN controller.

[0033] The replica digital twin scene collects the status information of the intelligent agent in real time, and transmits the status information to the intelligent agent using a preset communication protocol, and the intelligent agent makes decisions based on the status information.

[0034] In this example, a digital twin of a dungeon is constructed using the Unity environment simulator, collecting player status information in real time. This simulator accurately reproduces MMO dungeon scenes (such as the "Void Corridor" in the game "Magic Online") and supports dynamic parameter injection (e.g., trap locations, boss skills). Player status (e.g., coordinates, health, skill cooldowns, and coverage heatmaps) is transmitted to the DQN controller in real time at 60Hz via the gRPC protocol.

[0035] Environmental Modeling: Use Unity to build a replica digital twin and define the floating bridge triggering rules (player equipment weight ≥ 200 units).

[0036] Get real-time status via API: Python def get_state(): return { "position": (x, y), # Player coordinates "HP": player.health, # Current health "skill_CD": [1.2, 3.5, 0.0], # Fireball / Heal / Shield Cooldown "coverage_map": load_grid() # Overlay heat map } In this embodiment, the hierarchical DQN controller adopts a 5-layer fully connected network architecture, including an input layer, two hidden layers, an output layer, and a hidden layer activation function; Among them, the input layer is 256-dimensional, the hidden layer is 512-dimensional, the output layer is 64-dimensional action value, and the hidden layer activation function is LeakyReLU (α=0.01).

[0037] The dual network mechanism of the hierarchical DQN controller includes: The MainNet network is responsible for the agent's action decision-making in real time, and the TargetNet network outputs the target Q value of the DQN network. The target Q value is used to measure the quality of the MainNet network's current decision, thereby providing feedback to update the MainNet network. The MainNet network is trained for a first preset number of times, and the network weights are updated by a gradient descent algorithm; The TargetNet network synchronizes network weights from the MainNet once every time it performs a second preset number of trainings, where the second preset number of trainings is greater than the first preset number of trainings.

[0038] In this embodiment, the first preset number is 4, and the second preset number is 1000.

[0039] MainNet (online network): Gradient descent updates every 4 steps, making real-time action decisions.

[0040] TargetNet: synchronizes weights from MainNet every 1000 steps to stabilize the training process.

[0041] In the hierarchical DQN controller, the state space is defined including the agent's coordinates, health, skill cooldown, and coverage heatmap. The action space is defined to include movement, skills, and interaction commands.

[0042] In this embodiment, the state space contains the player's real-time state and environment information: ; Among them, (x, y) represents coordinates; HP represents health value (blood volume); Indicates skill cooldown time; Represents a 20×20 coverage heatmap matrix.

[0043] The state space consists of the following four parts: Player coordinates: (x,y)∈[0,100]×[0,100] (normalized coordinate system); HP and skill cooldown: HP∈[0,100%], skill cooldown time CD_skill∈[0,15 seconds]; Coverage heat map: 256×256 matrix G, where element G(i,j) represents the coverage of coordinate (i,j).

[0044] The action space contains 3 categories and 12 actions: action_space = { "move":["N","S","E","W","NE","NW","SE","SW"], #8 direction movement "skill":["fireball","heal","shield"],#3 kinds of skills release "interact":["chest","lever","npc"]#3 kinds of interactive actions } The agent uses a dual-network mechanism of a hierarchical DQN controller to make decisions, including: Use dynamic curriculum learning to hierarchically train agents, gradually increase difficulty based on instance complexity and migrate low-level training weights, and combine multi-dimensional reward functions to guide agent exploration; In this embodiment, the dynamic curriculum is used to learn hierarchical training of the agent, gradually increasing the difficulty according to the complexity of the copy and migrating the low-level training weights, including: The complexity of the dungeon is quantified by the number of dungeon branch paths, the number of dynamic events, and the number of bosses. The difficulty level of the dungeon is divided according to the complexity of the dungeon. The quantitative calculation of the replica complexity is: ; in, Indicates the number of replica branch paths, Indicates the number of dynamic events, Indicates the number of BOSSes.

[0045] The agent is trained from the lowest difficulty level, and when the agent meets the preset upgrade conditions in a preset number of consecutive trainings, a more complex mechanism is introduced based on the optimal network weights obtained from the current level training to increase the difficulty level of the training and adjust the learning rate.

[0046] In this embodiment, the course learning strategy is: Initial stage: ε = 0.9 (exploration rate), decay coefficient 0.9995 per step.

[0047] Upgrade conditions: 3 consecutive training sessions with coverage ≥ 90% and HP loss ≤ 20%.

[0048] Migration rule: Load low-level weights during high-level training, and the learning rate decays to 50%.

[0049] Parameter configuration: Python batch_size = 64 # Experience replay batch size gamma = 0.99 # Future reward discount factor epsilon_start = 0.9 # Initial exploration rate epsilon_decay = 0.9995 # Exploration rate decay coefficient Course learning process: Lv1→Lv3: Train each level until coverage is ≥90%.

[0050] Lv4→Lv5: Allow HP loss ≤30%, add BOSS hatred linkage mechanism.

[0051] Graded training rules: Lv1 (C≤1.0): Linear path, no dynamic events.

[0052] Lv5 (C>4.5): Multiple BOSS linkage + complex mechanism combination, HP loss allowed ≤30%.

[0053] The dynamic curriculum mechanism reduces the number of training steps to reach 90% coverage by 44% (280,000 steps vs. 500,000 steps for traditional DQN).

[0054] Stability: By migrating parameters between layers, we avoid the gradient explosion problem (e.g., Q-value oscillation caused by boss skill combinations) that occurs when traditional DQN directly trains high-complexity replicas.

[0055] The definition of the multi-dimensional reward function: ; in, Indicates the newly added coverage area. represents the total area of ​​the replica, Indicates the number of times a dynamic event is triggered. Indicates the number of times the area is visited repeatedly. Indicates the blood loss ratio.

[0056] In addition, this embodiment also adopts a hierarchical design for the reward function. The higher the level, the more dynamic the path coverage weight (0.5→0.7) and HP loss penalty (0.3→0.1) in the reward function change, guiding DQN to balance exploration and survival at different difficulty levels.

[0057] Generate test cases based on the agent's exploration trajectory.

[0058] In this embodiment, the agent's exploration trajectory is converted into a Python / Unreal dual-platform script, which supports one-click playback on the TaaS platform, including action sequences (movement, skill release, interaction) and assertion verification (coverage, status check).

[0059] In this embodiment, the script example: Python def test_case_05(): move_to(18, 22)# Trigger the floating bridge mechanism wait_for(2.0)# Wait for the floating bridge to be generated cast_skill("fireball", (19, 23))# Attack BOSS move_to(20, 25)# Avoid range skills assert_coverage(89.3%)# Verify coverage Integration testing: Scripts are imported into the TaaS platform to trigger the automated testing pipeline.

[0060] Experimental results:

[0061] Typical application cases Case 1: Discovered the vulnerability of "coordinates not being reset after the floating bridge disappears" (traditional methods do not cover this path).

[0062] Case 2: 17 use cases for avoiding "double boss linkage skills" were generated, increasing coverage by 22%.

[0063] Technology differentiation comparison:

[0064] Improve path coverage: Through autonomous exploration by intelligent agents, dynamic mechanisms, hidden paths, and abnormal player behavior are covered.

[0065] Shortened training cycle: The dynamic course learning mechanism reduces the number of training steps by 44%.

[0066] Real-time adaptive update: After a new element is added to the replica, a high-coverage use case is generated within 3 hours.

[0067] Please refer to Figure 2 , the second embodiment of the present invention is: A terminal 1 for generating automated test cases for game copies comprises a processor 2, a memory 3, and a computer program stored in the memory 3 and executable on the processor 2. When the processor 2 executes the computer program, the steps in the above-mentioned method for generating automated test cases for game copies are implemented.

[0068] In summary, the present invention provides a method and terminal for automatically generating test cases for game copies. A digital twin scene is constructed through an environmental simulator, and the intelligent agent autonomously explores based on hierarchical DQN and dynamic course learning, covering hidden paths and dynamic mechanisms. This can effectively improve test coverage and ensure the effectiveness of test case generation. Dynamic course learning can effectively improve training efficiency, quickly adapt to copy updates, solve the defects of low efficiency of random testing and poor adaptability of static algorithms, and achieve high-coverage, efficient and adaptive test case generation.

[0069] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's description and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for generating automated test cases for a game copy, characterized in that: Including steps: Build a replica digital twin scenario through the environment simulator and generate an intelligent agent that makes decisions using a dual-network mechanism of a hierarchical DQN controller; Use dynamic curriculum learning to hierarchically train agents, gradually increase difficulty based on instance complexity and migrate low-level training weights, and combine multi-dimensional reward functions to guide agent exploration; Generate test cases based on the agent's exploration trajectory.

2. A method for generating automated test cases for a game copy according to claim 1, characterized in that: Using dynamic curriculum learning to train agents in a hierarchical manner, gradually increasing the difficulty based on the complexity of the instance and migrating low-level training weights include: The complexity of the dungeon is quantified by the number of dungeon branch paths, the number of dynamic events, and the number of bosses. The difficulty level of the dungeon is divided according to the complexity of the dungeon. The agent is trained from the lowest difficulty level, and when the agent meets the preset upgrade conditions in a preset number of consecutive trainings, a more complex mechanism is introduced based on the optimal network weights obtained from the current level training to increase the difficulty level of the training and adjust the learning rate.

3. A method for generating automated test cases for a game copy according to claim 1 or 2, characterized in that: The quantitative calculation of the replica complexity is: ； in, Indicates the number of replica branch paths, Indicates the number of dynamic events, Indicates the number of BOSSes.

4. A method for generating automated test cases for a game copy according to claim 1, characterized in that: The hierarchical DQN controller adopts a 5-layer fully connected network architecture, including an input layer, two hidden layers, an output layer, and a hidden layer activation function; Among them, the input layer is 256 dimensions, the hidden layer is 512 dimensions, the output layer is 64-dimensional action value, and the hidden layer activation function is LeakyReLU.

5. A method for generating automated test cases for a game copy according to claim 1, characterized in that: The definition of the multi-dimensional reward function: ； in, Indicates the newly added coverage area. represents the total area of ​​the replica, Indicates the number of times a dynamic event is triggered. Indicates the number of times the area is visited repeatedly. Indicates the blood loss ratio.

6. A method for generating automated test cases for a game copy according to claim 1, characterized in that: The dual network mechanism of the hierarchical DQN controller includes: The MainNet network is responsible for the agent's action decision-making in real time, and the TargetNet network outputs the target Q value of the DQN network. The target Q value is used to measure the quality of the MainNet network's current decision, thereby providing feedback to update the MainNet network. The MainNet network is trained for a first preset number of times, and the network weights are updated by a gradient descent algorithm; The TargetNet network synchronizes network weights from the MainNet once every time it performs a second preset number of trainings, where the second preset number of trainings is greater than the first preset number of trainings.

7. A method for generating automated test cases for a game copy according to claim 6, characterized in that: The first preset number is 4, and the second preset number is 1000.

8. A method for generating automated test cases for a game copy according to claim 1, characterized in that: In the hierarchical DQN controller, the state space is defined including the agent's coordinates, health, skill cooldown, and coverage heatmap. The action space is defined to include movement, skills, and interaction commands.

9. A method for generating automated test cases for a game copy according to claim 1, characterized in that: The agent uses a dual-network mechanism of a hierarchical DQN controller to make decisions, including: The replica digital twin scene collects the status information of the intelligent agent in real time, and transmits the status information to the intelligent agent using a preset communication protocol, and the intelligent agent makes decisions based on the status information.

10. A terminal for generating automated test cases for game copies, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps in the method for generating automated test cases for game copies described in any one of claims 1 to 9 are implemented.

Images