I built a Flappy Bird game that runs a Q-Learning agent on an ESP32, with a small OLED display and a push button. The project lets me see the agent train and also play the game manually.
I developed a Flappy Bird with a Reinforcement Learning (Q-Learning) agent running on an ESP32 microcontroller, with a 0.96-inch OLED display and a push button.
The agent uses an epsilon-greedy strategy balancing exploration and exploitation.
The OLED displays its training stats every time the AI loses.
In the images, the agent has completed 22,033 training episodes.
In the images, the agent's best score is 12.
In the images, the agent's exploration rate is 0.04.
Watching an agent learn through thousands of attempts on a tiny OLED screen makes the concepts much more swallowable than studying reinforcement learning equations alone.
I can switch to manual mode and play the game using the button.
I have no idea why I'm getting into microcontrollers.