Search

CN-121988041-A - Game AI-oriented node data driving decision-making method and system

CN121988041ACN 121988041 ACN121988041 ACN 121988041ACN-121988041-A

Abstract

The invention discloses a node data driving decision method and a node data driving decision system for a game AI, and relates to the technical field of electronic games. The method comprises the steps of constructing a game AI neural node library, connecting and assembling a graphical decision logic structure by graphical nodes, executing the decision structure in game operation to generate AI instructions, recording differentiable tracks, calculating performance rewards according to game feedback, updating internal network parameters of nodes in the node library by utilizing the performance rewards and the differentiable decision tracks, and optimizing the decision logic structure. The technical problems that the traditional game AI decision logic relies on manual writing rules, is difficult to adapt to a complex dynamic environment, cannot be trained end to end and is difficult to automatically optimize according to game feedback are solved, and the technical effects that a flexible and adjustable decision logic structure is built through graphical node assembly, the end to end gradient optimization is realized by combining a differentiable decision track and performance rewarding feedback, and the adaptability, the intelligence and the optimization efficiency of game AI decision are improved are achieved.

Inventors

  • LIU CHENGSEN

Assignees

  • 深圳虎玥互娱科技有限公司

Dates

Publication Date
20260508
Application Date
20251210

Claims (10)

  1. 1. The node data driving decision method facing to the game AI is characterized by comprising the following steps: Constructing a neural node library of the game AI, wherein the neural node library comprises a neural sensing node and a neural action node; based on the neural node library, assembling a graphical decision logic structure in a graphical node connection mode; Traversing and executing the decision logic structure based on the current game environment state data in the game running process, generating and executing an AI control instruction, and recording execution process data to obtain a differentiable decision track; calculating a performance rewarding value according to game result feedback generated after the AI control command is executed; and updating internal network parameters of the neural nodes in the neural node library by using the performance reward value and the differentiable decision track, and optimizing the decision logic structure.
  2. 2. The game AI-oriented nodal data-driven decision-making method of claim 1, wherein during game play, traversing and executing the decision logic structure based on current game environment state data, generating and executing AI control instructions, comprises: Starting a recursive traversal process from a root node of the decision logic structure in each game frame, calling a forward reasoning calculation diagram in the nerve perception node when traversing to the nerve perception node, and inputting current game environment state data to obtain a confidence score; Comparing the confidence score with a preset dynamic threshold value, and judging whether to activate the current node and the subsequently connected child nodes according to a comparison result; and when traversing to the neural action node, executing forward propagation of the parameter prediction sub-network in the neural action node, generating an AI control instruction and submitting the parameters to a game execution layer.
  3. 3. The game AI-oriented nodal data-driven decision-making method of claim 2, wherein concurrently recording the execution process data to obtain a differentiable decision-making trace, comprising: In the execution process of the decision logic structure, a calculation map shadow is synchronously constructed, and the calculation map shadow is parallel to game execution logic and is used for completely recording the forward propagation input data, internal weight, activation output and the connection relation between nodes of each neural node; Linking the calculated shadow graphs of each decision step according to time sequence to generate a decision track data chain; And carrying out alignment storage on the decision track data chain and the game original state frame corresponding to the time stamp to generate the differentiable decision track.
  4. 4. The game AI-oriented nodal data-driven decision-making method of claim 1, wherein calculating a performance prize value based on game outcome feedback generated after execution of said AI control instructions, comprises: defining a sparse final prize function based on game final goals; Introducing a reward distribution network, taking a differentiable decision track as input, and reversely distributing sparse final rewards to each neural node in the decision track based on the sparse final rewards function to obtain distributed rewards; A dense sub-prize function associated with the intermediate behavior is defined for evaluating the performance of the game AI in terms of movement efficiency, resource management and tactical execution, and the performance prize value is obtained by weighting the base sub-prizes obtained based on the dense sub-prize function with the distribution prizes from the prize distribution network.
  5. 5. The game AI-oriented nodal data-driven decision-making method of claim 1, wherein updating internal network parameters of neural nodes in the neural node library with the performance rewards value and the differentiable decision-making trace, optimizes the decision logic structure, comprising: The collected performance rewards and the corresponding differentiable decision tracks in the game running process are stored in an experience playback buffer area at regular intervals; Sampling batch track data from the experience playback buffer area on a special training server, adopting a distributed training architecture, reversely spreading a plurality of training ends in parallel through a differentiable decision track, calculating gradients, and carrying out gradient aggregation and asynchronous updating by a parameter server; and optimizing the decision logic structure by using the updated neural node network parameters.
  6. 6. The game AI-oriented nodal data-driven decision-making method of claim 1, wherein optimizing the decision logic structure further comprises: continuously monitoring the use frequency and average rewarding contribution of each node in the decision logic structure; when the duration of the average contribution of any node is lower than the preset threshold value and is longer than the preset duration, marking the corresponding node as the node to be optimized; And aiming at the node to be optimized, triggering a structure searching process, selecting a group of candidate substructures from a candidate neural node library, performing simulation evaluation on a verification set formed by historical data, and replacing the node to be optimized with the candidate substructures with optimal evaluation performance.
  7. 7. The game AI-oriented nodal data driven decision-making method of claim 1, wherein said neural sense node is configured to receive game environment state data and output continuous values of condition satisfaction, said neural action node being configured to output executable actions and parameterized instructions.
  8. 8. The game AI-oriented nodal data-driven decision-making method of claim 7, wherein the neural-aware nodes include one or more of visual-aware subnodes, numerical-state-aware subnodes and auditory/event-aware subnodes.
  9. 9. The game AI-oriented nodal data-driven decision-making method of claim 7, wherein said neural-aware node and said neural-action node share an underlying feature extraction network based on game state encoding.
  10. 10. A game AI-oriented nodal data-driven decision system for implementing the game AI-oriented nodal data-driven decision method of any of claims 1-9, comprising: The node library construction module is used for constructing a neural node library of the game AI, and comprises a neural sensing node and a neural action node; The decision logic structure assembling module is used for assembling a graphic decision logic structure in a mode of graphic node connection based on the neural node library; the decision logic structure executing module is used for traversing and executing the decision logic structure based on the current game environment state data in the game running process, generating and executing an AI control instruction, and recording executing process data to obtain a differentiable decision track; The result feedback module is used for calculating a performance rewarding value according to game result feedback generated after the AI control command is executed; and the decision logic structure optimization module is used for updating the internal network parameters of the neural nodes in the neural node library by utilizing the performance rewarding value and the differentiable decision track and optimizing the decision logic structure.

Description

Game AI-oriented node data driving decision-making method and system Technical Field The invention relates to the technical field of electronic games, in particular to a node data driving decision method and system for a game AI. Background In recent years, with the increasing complexity of electronic games, there has been a higher demand for the level of intelligence of non-player characters (NPCs). Conventional game AI typically relies on state machines, behavioral trees, or scripted hard coded logic, which, while advantageous in terms of controllability, tend to appear stiff and difficult to adapt when faced with an open, dynamic game environment. The developer needs to put a great deal of effort to manually design and debug the numerous rules and state transitions, which not only is inefficient, but also the finally generated AI behavior has obvious bottlenecks in diversity and strain capacity. Meanwhile, technologies such as deep reinforcement learning and the like are remarkably developed in the field of general artificial intelligence, and the technology provides new possibility for constructing a more flexible and powerful game AI through the characteristics of interaction with the environment and self-optimization in a data-driven manner. However, directly applying an end-to-end depth model to a game presents many challenges, including the model's decision process being "black box" and difficult to understand and debug, the training process being unstable, requiring massive trial and error, the generated strategy lacking a definite logical structure and difficult to incorporate into the high-level intent of the game designer. Disclosure of Invention The application provides a node data driving decision method and a node data driving decision system for a game AI, which solve the technical problems that the traditional game AI decision logic relies on manual writing rules, is difficult to adapt to a complex dynamic environment, cannot realize end-to-end training and is difficult to automatically optimize according to game feedback. In a first aspect of the present application, there is provided a game AI-oriented nodal data-driven decision-making method comprising: The method comprises the steps of constructing a neural node library of a game AI, constructing a graph-shaped decision logic structure based on the neural node library in a mode of connecting graphical nodes, traversing and executing the decision logic structure based on current game environment state data in the game operation process, generating and executing an AI control instruction, recording execution process data to obtain a differentiable decision track, feeding back a game result generated after the AI control instruction is executed, calculating a performance rewarding value, and updating internal network parameters of the neural nodes in the neural node library by utilizing the performance rewarding value and the differentiable decision track to optimize the decision logic structure. In a second aspect of the present application, there is provided a game AI-oriented, nodal data-driven decision-making system, the system comprising: The node library construction module is used for constructing a neural node library of a game AI and comprises a neural sensing node and a neural action node, the decision logic structure assembly module is used for assembling a graphic decision logic structure based on the neural node library in a graphical node connection mode, the decision logic structure execution module is used for traversing and executing the decision logic structure based on current game environment state data in the game operation process, generating and executing an AI control instruction and recording execution process data to obtain a differentiable decision track, the result feedback module is used for calculating a performance rewarding value according to game result feedback generated after the AI control instruction is executed, and the decision logic structure optimization module is used for updating internal network parameters of the neural nodes in the neural node library and optimizing the decision logic structure by utilizing the performance rewarding value and the differentiable decision track. One or more technical schemes provided by the application have at least the following technical effects or advantages: Firstly, a node library comprising two types of neural nodes of perception and action is established in advance, and an editable graph decision structure is built by utilizing a visual node connection mode. When the game runs, the decision structure is traversed and executed node by node according to the real-time environment state, a corresponding AI control instruction is output, and meanwhile, the calculation process of each step is completely recorded as a differentiable track for learning. And then, calculating performance rewards according to the results generated by the instructions in the game, and