Abstract
This paper provides a comprehensive overview of research from the Machine Learning and Perception Lab at the University of Manchester (MLP@UoM) focused on applying machine learning (ML) to the video game development lifecycle. By moving away from handcrafted rules toward systematic, learning-based approaches, the lab addresses challenges in automated level generation, educational game design, and robust AI agent perception.
Introduction
Traditional video game development is a laborious and costly process. Developers often rely on handcrafted rules for procedural content generation (PCG) and hardcoded AI behaviors, leading to several bottlenecks. These include the rigidity of search-based content generation, the difficulty of mapping educational material to gameplay in serious games, the slow pace of real-time difficulty adjustment, and the brittleness of standard Reinforcement Learning (RL) agents that learn from raw pixels. MLP@UoM researchers aim to overcome these limitations by integrating ML models that encode developer knowledge and adapt to player behavior dynamically.
Methods
The research at MLP@UoM introduces several novel frameworks designed to streamline development and enhance player experience:
Learning-Based Procedural Content Generation (LBPCG)
The LBPCG framework utilizes a three-stage process—Development, Public Test, and Adaptive—to generate personalized content. It uses ML models to capture developer intent and public consensus before providing individualized levels to target players.
Constructive Primitives (CPs)
Instead of generating entire levels at once, the system decomposes them into short, fixed-size segments known as Constructive Primitives. By employing active learning and clustering, the framework minimizes the manual effort required for developers to rate content, allowing the AI to learn quality standards efficiently.
Object Embedding and Skill Prediction
To improve AI robustness, the Object Embedding Network (OEN) allows agents to perceive game states as coherent objects rather than raw pixels. Additionally, the lab utilizes Random Forest classifiers to predict player skill levels based on raw input events (mouse/keyboard) rather than just game scores.
Results and Key Insights
The lab's research has yielded several significant findings:
- Rapid Skill Capture: A player's skill level can be predicted with reasonable accuracy from just 30 seconds of raw input data, enabling faster dynamic difficulty adjustment.
- Robustness to Visual Noise: Moving from pixel-based to object-based RL (OEN) provides agents with "immunity" to background changes, ensuring they focus on relevant game properties.
- Personalization and Education: The LBPCG framework successfully generated quality, personalized levels for various skill tiers. In educational studies using Chem Dungeon, automated mapping of chemistry knowledge to gameplay resulted in measurable learning improvements.
Conclusion and Future Work
While significant strides have been made, challenges remain. Future research aims to reduce the data burden on developers through autonomous learnable agents for auto-testing. Additionally, the lab seeks to refine adaptive strategies in educational games to maintain player "flow" and develop AI agents that exhibit human-like playstyles, mimicking the quirks and mistakes of real players rather than just optimizing for high scores.
Summary updated