Using Aspect-Oriented Programming to Record DRL Agents' Data

3 days ago

Playtesting is crucial for game development to find bugs, evaluate UX, balance the game, and assess fun.
PEAK is a game engine with integrated deep reinforcement learning (DRL) agents for automated testing of 2D platformers.
The workflow involves designing a level, training DRL agents with specific playstyles, and collecting data to identify balance issues.
The project focuses on data collection, aggregation, and visualization, using the Observer Pattern and Aspect-Oriented Programming (AOP).
Aspect-Oriented Programming (AOP) is used to separate data collection (aspects) from core game code, with pointcuts specifying join points.
Decorators in Python are initially used to track data, but they require editing core game files.
Config files (YAML) are introduced to inject recorder methods into game methods via reflection, avoiding code changes.
A simple Streamlit dashboard is created for data visualization, though current insights are limited.
Future work includes defining useful metrics (e.g., completion rate, mean completion time) and thresholds via config files, with challenges in measurement.

Hasty Briefsbeta