A Strategy for Adaptive, Modular Game AI
Tiilikainen, Toni (2025)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025060520621
https://urn.fi/URN:NBN:fi:amk-2025060520621
Tiivistelmä
The objective of this thesis was to explore design strategies for adaptive and modular game AI and to implement a prototype that avoids common issues associated with increased adaptiveness. Adaptive AI systems are a growing interest in game development due to their potential to provide dynamic and responsive behavior. However, developing such systems poses unique design and integration challenges.
A broad technology review covering over ten established game AI architectures was conducted, categorized into state-based, learning-based, and support systems. Each architecture was analyzed in terms of adaptability, modularity, and ease of implementation to identify potential trade-offs and integration challenges.
Based on this review, several side effects of increased adaptability were identified, including design complexity, overfitting, and unpredictability. In order to address these issues, a hybrid model was proposed, combining behavior trees with utility-based decision-making. The architecture emphasized modular design through separate behavior and evaluation modules, and incorporated a weight-based learning mechanism to enable online behavioral adaptation during runtime. A shared blackboard facilitated integration and coordination between components.
Based on this model, a prototype implementation was developed in a controlled simulation, allowing agents to adapt their behavior to environmental changes dynamically. Evaluation results showed that the system successfully adapted during runtime without sacrificing modularity or introducing significant overhead.
While limited in complexity, the implementation serves as a functional baseline for building maintainable and extensible adaptive AI. Future development could involve integrating planning systems such as GOAP for increased expressiveness or further modularizing the evaluation process to support more sophisticated learning architectures.
A broad technology review covering over ten established game AI architectures was conducted, categorized into state-based, learning-based, and support systems. Each architecture was analyzed in terms of adaptability, modularity, and ease of implementation to identify potential trade-offs and integration challenges.
Based on this review, several side effects of increased adaptability were identified, including design complexity, overfitting, and unpredictability. In order to address these issues, a hybrid model was proposed, combining behavior trees with utility-based decision-making. The architecture emphasized modular design through separate behavior and evaluation modules, and incorporated a weight-based learning mechanism to enable online behavioral adaptation during runtime. A shared blackboard facilitated integration and coordination between components.
Based on this model, a prototype implementation was developed in a controlled simulation, allowing agents to adapt their behavior to environmental changes dynamically. Evaluation results showed that the system successfully adapted during runtime without sacrificing modularity or introducing significant overhead.
While limited in complexity, the implementation serves as a functional baseline for building maintainable and extensible adaptive AI. Future development could involve integrating planning systems such as GOAP for increased expressiveness or further modularizing the evaluation process to support more sophisticated learning architectures.