SamanKhamesian/Hybrid-Self-Attention-NEAT ? reverse-engineered prompt

Build me this Hybrid Self Attention NEAT project in Python so I can reproduce the paper’s results and experiment with it on Atari style raw pixel inputs.

I want a working training flow that runs from a simple command, uses the default experiment configs, and saves the trained model so it can be tuned afterward. Please make sure it can run both on a normal machine and on a headless server, with the server setup handled cleanly if a display is needed. It should be easy to change things like population size and number of iterations without digging through a lot of code.

If anything about the original setup or dependencies is unclear, look up the current docs online if you need to. Also make the code easy to read and keep the main scripts straightforward so I can train first, then run tuning on the saved model.