Flappy Bird AI

Overview

This project is a dependency-free Python port of the Flappy Bird neuroevolution demo from a larger portfolio: a flock of agents each controlled by a small neural network learns to fly through gaps by selection, mutation, and elitism across generations. There is no graphics in this repo version—the same physics, collisions, scoring, and evolution run headlessly so the AI logic is easy to read and run from the command line.

How It Works

• Neural network — Each bird has a tiny feedforward network: 5 normalized inputs (vertical position, velocity, gap bounds, distance to pipe), 6 hidden units with tanh, and 1 output with a logistic sigmoid. If the output is greater than 0.5, the bird flaps (same rule as the original implementation).
• Genetic algorithm — The population is fixed size (POP_SIZE). When all birds die, the next generation is built from the top ELITE_COUNT brains (by fitness). Each new bird gets a mutated copy of an elite (random Gaussian-style perturbations on weights with fixed mutation rate and scale).
• Fitness — Birds are ranked by pipes_passed * 1000 + frames_alive (survival score ticks up each step while alive). Higher is better; this matches the sort key used in the source demo.
• World — Pipes spawn on a fixed frame interval, move left at constant speed, and kill birds outside the gap when horizontally overlapping. Passing a pipe increments that bird’s pipe count and a global pass counter.

Features

• Population-based neuroevolution with elitism and weight mutation
• Alpha-free small MLP forward pass (tanh + sigmoid)
• Headless simulation class (FlappySimulation) mirroring the original update loop
• No third-party dependencies (standard library only)

Usage

python flappy_bird_ai.py

To experiment in code, import FlappySimulation, Bird, Brain, or constants from flappy_bird_ai and call sim.reset_world(), sim.make_flock(), and sim.step(sim_speed=1) in your own loop.

Example

The built-in script seeds RNG for reproducibility, creates a simulation, runs a fixed number of inner physics steps (step with sim_speed=1), and prints progress every 20,000 steps: current generation, alive count, pipe pass score, and best composite fitness seen so far. This demonstrates evolution advancing without a canvas.

Why I Built This

It is a compact reference for embodied decision-making under noise: cheap function approximators (neural nets), parallel search over policies (populations), and iterative improvement (evolution)—ideas that connect to RL, neuroevolution, and interactive systems more broadly.