Flappy NEAT is a chaotic little testbench where neural networks learn to play a Flappy Bird–style game using a mix of genetic neuro-evolution and reinforcement learning. It’s designed to let you:
- Compare human, algorithmic, neuro-evolved, and RL-trained performance
- Seamlessly switch from evolution to RL fine-tuning once a score threshold is hit
- Experiment with unique genetic operators and learning tricks you've probably never seen before
It’s a research toy, a performance benchmark, and a testbed for weird ideas—all in one.
- 🚀 Alpha-selection: A novel genetic operator to drive better diversity
- 🔥 Neuron heat dynamics: RL weight updates are modulated by how “hot” neurons are
- 🧬 Custom crossover/inversion/mutation logic
- 🎩 Interactive human mode for score comparison
- 🤖 Algorithmic solver mode (perfect agent)
- 🧪 Experimental PPO-inspired RL strategy
- 🔀 Hybrid GA → RL transition based on fitness threshold
- 🛠️ Live UI with parameter tweaking
- 📚 Help system built into the UI with all information needed to understand every parameter
- 🔄 Save/load for agents and populations
- 🌍 Neural network visualization and weight matrix viewer
- 🕛 Scenario replay support for post-mortem analysis
This program was built using Delphi 7, which means:
⚠️ You will need a Windows XP environment with Delphi 7 installed to compile and run this project.
We recommend spinning up a virtual machine for this purpose. Nostalgia points are included for free.
Train a population to solve the level using evolution, then transition into RL for fine-tuning behavior. Compare its score to your own flapping skills, or the perfect algorithmic solver. Discover if your AlphaSort operator beats standard crossover methods. Watch neuron "heat maps" influence updates. Or just... create a mutant bird that flaps with pure chaos.
Hint: Try to use test INI files provided in the repo.