Movement Predictions in Airstriker Genesis
As I delved into the Airstriker Genesis ROM bundled within the OpenAI Retro Gym package, I’ve started thinking how an AI could learn to play this game. One of the first things I’ve thought of was that there needs to be some prediction of movement, considering the barrage of diverse objects approaching your humble spacecraft that need to be either avoided or destroyed. For an AI (or any type of intelligence really) to do this, it should be able to approximate where those objects are headed. Since there’s different types of objects (enemy ships, meteorites, etc.) behaving differently over the course of the game, the AI must learn these on the fly. I wondered how this could be achieved not through prior training, but through dynamic online learning.
So I went and developed a little prototype that does exactly that: Learn movement patterns of different objects and recognize them - all without prior exposure or training. My approach is rooted in real-time adaptation, where AI unfolds its potential without the constraints of predefined patterns.
Check out this video that shows the prototype in action:
The video provides actual gameplay alongside the system's predictive output, offering a real-time view of its evolving movement predictions. It's important to mention that during the gameplay in the video, I was manually steering the ship to showcase the prediction only. I've obviously avoided collisions. :)
Of particular interest is the system's ability to discern the behavior of different game elements. For example its prediction of those falling meteorites starts weak, even though a spaceship has already travelled the same path before. However, as the first meteorite completes its path, the predictions dramatically improve. This highlights the system's capability to differentiate between objects and predict their behaviors accordingly.
Currently, I'm focusing on improving horizontal movement detection, my next step is implementing a way to share knowledge about object's behavior between the hole system. That way, trajectory prediction learned in one location should be available instantly for the whole system.
