TLDR: This research introduces a new automated game testing method using a hierarchical AI system. A high-level Bayesian Optimization module efficiently identifies key areas in a game level for an agent to explore, while a low-level learned agent navigates. A novel, scalable grid map model addresses data limitations of traditional optimization. This approach significantly improves map coverage and bug detection efficiency compared to random exploration.
Automated game testing is a critical but challenging aspect of game development. With the vast complexity of modern games, relying solely on human testers to find every potential bug is prohibitively expensive and time-consuming. From collision issues and performance drops to game balance problems, the sheer number of possible game states makes comprehensive manual testing nearly impossible.
Traditionally, game testing has involved manually scripted bots or, more recently, agents developed using machine learning techniques like Reinforcement Learning (RL) and Imitation Learning (IL). While these methods have shown promise, RL can be computationally intensive and require significant iteration, especially when game levels change. IL, while less expensive, relies on expert demonstration data, and agents trained to “play well” might not be the best at uncovering bugs.
A new research paper introduces an innovative approach to automated game testing that significantly improves efficiency and bug detection. The work, titled “Bayesian Optimization-based Search for Agent Control in Automated Game Testing,” proposes a hierarchical and modular system that combines the power of learned AI agents with a sophisticated search manager based on Bayesian Optimization (BO).
The core idea is to uncouple the decision-making process into two main modules: a high-level module and a low-level module. The high-level module, powered by Bayesian Optimization, acts as a “search manager.” Its role is to intelligently determine the most informative locations within a game level for an agent to explore, based on the information gathered so far. This is crucial because BO is excellent at efficiently finding optimal points in complex, unknown functions, making it ideal for navigating a game level to maximize bug discovery or map coverage.
The low-level module, on the other hand, consists of an already trained AI agent (using IL or RL in controlled environments) whose job is to take the target coordinates provided by the high-level module and navigate the game character to those locations. This modularity offers significant advantages: if game mechanics change, only the low-level module needs retraining, while the high-level BO module remains effective. Similarly, if the goal is to find a different type of bug, only the high-level module needs adaptation, allowing the low-level agent to be reused.
A key innovation in this system is the development of a game testing-specific surrogate model for Bayesian Optimization, built on a grid map representation. Traditional BO models, such as Gaussian Processes, struggle with the massive datasets generated during game testing, leading to scalability issues. The proposed grid map model, however, maintains a constant size regardless of the amount of data collected, making it highly efficient and scalable. It can also easily incorporate game level information like NavMesh data, which helps in discarding physically meaningless regions from exploration.
This grid map model also allows for efficient computation of uncertainty, a vital component for BO. By blending observed samples with their similarity to unvisited areas, it creates a smoothed representation of the game level’s “interest metric” (e.g., performance drops, visit frequency). The system then uses an acquisition function, specifically the Lower Confidence Bound (LCB), to balance between exploiting known areas of interest and exploring unknown, high-uncertainty regions. This ensures a thorough and intelligent exploration of the game map.
Furthermore, the low-level module incorporates an adaptive exploration mechanism. While it primarily follows the learned policy to reach target coordinates, it also introduces random exploratory actions in areas of high uncertainty. This means the agent is more likely to “make mistakes” or run into objects in unexplored or uncertain areas, which is highly beneficial for uncovering general issues with game entities that a perfectly trained agent might otherwise avoid. This adaptive exploration helps maximize the probability of finding bugs and reduces the cost of exploration by focusing efforts where they are most needed.
Experiments conducted in both Unity and Unreal Engine game levels demonstrated the effectiveness of this approach. Compared to a baseline system using uniform random target selection and no uncertainty-driven exploration, the proposed system achieved significantly better map coverage and a more even exploration distribution. For instance, in tests involving a “ghost wall” bug, the complete system not only covered more ground but also traversed the buggy area more frequently, proving its ability to detect subtle issues. The results indicate a substantial improvement in map coverage and a higher likelihood of discovering bugs, ultimately reducing testing costs.
Also Read:
- Landmarks Enhance Monte Carlo Planning for Uncertain Environments
- Holistic Explainable AI: A New Framework for End-to-End AI Transparency
This work represents a significant step forward in automated game testing, offering a robust, flexible, and scalable solution for efficiently exploring game environments and identifying potential issues. For more technical details, you can refer to the full research paper available here.
