Learning to play Super Mario Bros
Here is a video of a project from my lecture „Applied Reinforcement Learning“ @ LMU Munich. It shows how Mario learns to improve its behavior by reinforced sensorimotor interactions using Q-learning with eligibility traces. In the last part of the video you see how Mario can even learn to jump over gaps again and again, because we rewarded it appropriately.
Thanks to the developers of https://github.com/micheltokic/marioaifor sharing their Java source code with the community! My students and I developed a Python interface to the Java engine and wrapped it into an OpenAI Gym environment. See our code on GitHub: