CA2 Agent Uses Call Stack to Revolutionize Game Testing
By leveraging internal code signals, CA2 finds bugs faster than ever.
Automated game testing has long been a bottleneck in game development, with manual testing missing edge cases and existing automated methods failing to achieve full code coverage. Prior work applied reinforcement learning (RL) but ignored internal code signals like the call stack. The paper 'CA2: Code-Aware Agent for Automated Game Testing' introduces the Code Aware Agent (CA2), which leverages call stack information to learn more effective testing strategies. CA2 receives both the current function call trace and the game state, learning to navigate to specific target functions. The researchers instrumented two environment types—state-based and image-based—with efficient call stack extraction support.
Experimental evaluation shows that CA2 consistently outperforms non-code-aware baselines that do not use call stack data. The improvement is attributed to the agent's ability to leverage code-level information, making testing more targeted and efficient. This approach bridges the gap between game state and internal program execution, offering a promising direction for automated testing in complex software systems like modern games. The work is published on arXiv and authored by Valliappan Chidambaram Adaikkappan, Vincent Martineau, Joshua Romoff, and David Meger from McGill University and Ubisoft.
- CA2 integrates call stack information with game state for more precise testing.
- Achieves consistent performance gains over RL agents that ignore code signals.
- Supports both state-based and image-based game environments.
Why It Matters
Code-aware testing reduces manual QA effort and catches edge cases traditional automated methods miss.