Game AI Decision Model Training for Adaptive Battle Personification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In fighting games, the game AI's decision-making abilities are limited by fixed behavior trees and unreliable player-generated battle data, leading to over-fitting and a poor personification effect.
Innovation Solution
Implement a reinforcement learning method to train decision models by iteratively inputting battle state data, using reward values to enhance the decision-making ability of virtual characters, and incorporating behavior trees to improve battle personification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If supervised learning is used to train game AI with player-generated battle data, then the training process is simple and fast, but the reliability of training data is poor leading to over-fitting and reduced personification effect
Solution Approach 1:
A behavior tree is introduced as an intermediary component that works in conjunction with the decision model. The behavior tree provides reliable, pre-defined battle policies and decision-making logic, compensating for the poor reliability of player-generated training data. This mediator ensures that the game AI maintains high personification effect while still benefiting from the fast training speed of supervised learning.
2Ease of manufacture
If fixed behavior trees are used for game AI decision-making, then the implementation is simple, but the decision-making ability is limited and lacks adaptability
Solution Approach 1:
The patent merges two different approaches: fixed behavior trees and machine learning-based decision models. The behavior tree handles simple, rule-based decisions with high reliability, while the decision model handles complex, adaptive decision-making scenarios. This combination allows the game AI to maintain implementation simplicity through the behavior tree structure while gaining enhanced adaptability and decision-making ability from the trained decision model.
3Adaptability or versatility
If player battle data is used for training, then data acquisition is easy, but the data quality is poor leading to over-fitting problems
Solution Approach 1:
The behavior tree serves as a mediator that filters and guides the use of player battle data. Instead of directly training the decision model on raw player data (which causes over-fitting), the behavior tree provides a structured framework that processes and refines the data, ensuring higher quality training inputs while maintaining the ease of data acquisition from player gameplay.
Data Source
AI summary
A decision model training method and apparatus are provided. The method includes: obtaining a first model pool; inputting, in an nth iteration process, ith battle state data into an nth decision model to obtain an action instruction outputted by the nth decision model, the ith battle state data being for representing a battle state at an ith battle time in a battle process, and the action instruction being for instructing a first client to control a first virtual character to act and return i+1th battle state data collected by the first virtual character after acting; training the nth decision model based on battle state data in the battle process to obtain an n+1th decision model, the battle state data comprising the i+1th battle state data; and adding the n+1th decision model to the first model pool.