Abstract
Games have been an important tool for motivating undergraduate students majoring in computer science and engineering. However, it is difficult to build an entire game for education from scratch, because the task requires high-level programming skills and expertise to understand the graphics and physics. Recently, there have been many different game artificial intelligence (AI) competitions, ranging from board games to the state-of-the-art video games. The competitions have been designed such that participants develop their own AI module on top of public/commercial games. In this paper, we report our experiences using the Angry Birds AI Competition for such a project-based course. In the course, teams of students consider computer vision, strategic decision-making, resource management, and bug-free coding for their outcome. To promote understanding of game contents generation and extensive testing on the generalization abilities of the student's AI program, we developed software to help them create user-created levels. Students actively participated in the project and the final outcome was comparable with that of successful entries in the 2013 International Angry Birds AI Competition.
Introduction: The Motivation-Complexity Trade-off
Teaching Artificial Intelligence (AI) to undergraduates often involves a difficult balance between student engagement and technical complexity. While modern video games are highly motivating, the overhead of building a game engine—including physics, graphics, and user interfaces—is often prohibitive for a single-semester course.
The authors address this by leveraging the International Angry Birds AI Competition. Unlike discrete-state board games like Chess, Angry Birds presents a continuous 2D physics environment with hidden information and diverse agent types (birds) with unique abilities. This platform forces students to move beyond hard-coded logic toward generalized AI capable of succeeding on previously unseen levels.
Methodology: The Five Pillars of Game AI
The course was structured around five core technical challenges that students had to solve to build a successful autonomous agent:
- Computer Vision: Students improved upon basic APIs to implement pixel-level analysis. This was necessary to identify specific objects, such as tilted ice blocks or ground textures, that standard libraries often missed.
- Structural Analysis: Agents were required to analyze the integrity of buildings to identify "weak points." This involved mapping the right bird type to the most vulnerable material (e.g., using blue birds for ice structures).
- Time Management: In a competitive setting, agents are limited to three minutes per level. Students optimized their code to ensure agents did not "freeze" while waiting for physics simulations to settle.
- Strategic Decision-Making: Logic evolved from simple level completion to score maximization, requiring agents to destroy maximum objects with minimum bird usage.
- Handling Uncertainty: Because the underlying physics engine (Box2D) can produce non-deterministic results for identical shots, agents had to be robust enough to adapt to failure.
To prevent "overfitting" to standard levels, the authors provided a Custom Level Creation Tool. This allowed students to design their own challenges, forcing their peers' agents to demonstrate true generalization.
Evaluation and Results
The project utilized three datasets: the first 21 levels of "Poached Eggs" for training, 13 custom student-designed levels, and 10 levels from the 2013 IJCAI Competition for final blind testing.
The results demonstrated significant technical and pedagogical success:
- World-Class Performance: The top student team, "Plan A+," achieved a score that would have ranked them 5th globally in the actual 2013 International Angry Birds AI Competition.
- High Engagement: Student surveys reported a satisfaction score of 4.58/5, with students noting that the project-based approach significantly improved their understanding of program design compared to traditional lectures.
- Technical Maturity: Students identified "attack strategy" as the most difficult hurdle, indicating deep engagement with high-level AI logic rather than just low-level coding.
Conclusion
The use of the Angry Birds AI Competition platform effectively bridges the gap between theoretical AI and practical application. By tackling vision, strategy, and optimization in a familiar yet complex environment, students gain a realistic understanding of AI development. Future work will focus on refining parallelized optimization platforms to support more computationally intensive methods, such as Deep Reinforcement Learning, in undergraduate curricula.
Summary updated