Developing AI for Angry Birds with PDDL+

Angry Birds is a popular mobile game where players use a slingshot to launch birds at structures to knock over pigs. This game presents a challenge for artificial intelligence (AI) because it involves making decisions based on the physics of motion and collision. This study looks at how a specific type of planning and searching can help create an AI agent that plays Angry Birds effectively.

#The Challenge of Angry Birds

The game is known for requiring players to think about their moves carefully. Different Levels require different strategies. Some levels are easy while others are very hard, making the game a good candidate for developing AI solutions. Many versions of Angry Birds have been shown to be complex, requiring advanced reasoning skills that are tough for AI to replicate.

#Why Use PDDL+?

To Model the game, we use a language called PDDL+. This language is designed to work with both simple and complex changes in games and other systems. One key aspect of PDDL+ is its ability to manage different types of events happening at the same time, like a bird flying through the air while structures are collapsing.

#How We Created the Model

Our first step was to create a model of Angry Birds that accurately reflects how the game operates. We had to decide how best to represent the different elements of the game, such as the birds, pigs, and blocks. Each of these elements has its own features, like how much damage they can take or how they move.

We focused on modeling how the birds fly and interact with other objects. By simplifying some parts of the game, we could create a model that runs efficiently. We decided to represent the slingshot indirectly by noting where birds are launched from without making it an object in the model.

#Game Dynamics in the Model

The model describes how birds are launched, how they fly, and how they interact with pigs and blocks. The key to success in Angry Birds is making the right choices about how to launch the birds. The model simulates the launch of a bird by considering its speed and angle, which are critical for hitting targets.

#Events and Interactions

In the game, many interactions happen when birds hit pigs or blocks. When a bird collides with a pig, several outcomes can occur. The bird can destroy the pig directly or cause other objects to fall and hit the pig. To capture this complexity, we need to model various events that occur during these interactions.

For example, when a bird hits the ground, it can bounce back up. This was modeled by incorporating rules that dictate how birds behave on interaction with surfaces. Similarly, we modeled the effects of explosions, such as when a bird hits a TNT crate. The explosion can cause damage to nearby objects, which the model also needs to capture.

#Creating a Game-Playing Agent

To turn our model into an actionable AI agent, we had to create a game-playing agent called Hydra. This agent reads the game state and translates it into a format that the planner can understand. From there, Hydra generates a plan to solve each level by determining how to launch birds effectively.

#Search Techniques

The AI uses various search techniques to find the best moves. The agent can employ different algorithms, including breadth-first search and depth-first search. The challenge is to keep the search efficient while identifying the best possible moves.

We also developed specific strategies to improve how the agent Searches for solutions. One approach focuses on score, which considers how many pigs and blocks can be destroyed with each move. Another method looks at how close the bird is to a pig and how it moves toward it.

#Experimental Evaluation

To test how well Hydra performs, we compared it to other Agents in various Angry Birds levels. We established benchmark levels that range from simple to complex, varying in the number of objects and required strategies. By doing this, we can see how Hydra performs against other established AI agents.

#Results

The results of our evaluation showed that Hydra performed better than some previous agents in simpler levels. However, it faced challenges in the more complex levels, where numerous objects and interactions made it harder for the AI to make the best moves. In those levels, established agents that were designed to handle specific strategies performed better.

#Conclusion

In this study, we demonstrated how using PDDL+ can capture the dynamics of a challenging game like Angry Birds. The combination of effective modeling and efficient search techniques shows promise for developing AI that can tackle similar problems in the future.

While Hydra has had successes, there is still room for improvement, especially in handling complex levels. Future work will focus on developing better strategies and enhancing the model to include additional features of Angry Birds. This may involve including special powers of some birds and varying pig types.

By continuing this research, we aim to improve the performance of AI agents in games and other real-world scenarios, making them more effective at solving complex problems and adapting to new challenges.