flâneur

The Inverted Pendulum Problem with Deep Reinforcement Learning | by Saif Uddin Mahmud | Dabbler in Destress | Medium

medium.com · 2,334 words · saved by 1 readers

A professor of mine introduced me to the rather simple inverted pendulum problem — balance a stick on a moving platform, a hand let’s say. Intuition built on the physics of the “Game Engine” in our head tells us: if the stick is leaning to the left, move your hand to the left; if the stick is leaning to the right, move your hand to the right. We, humans, are exceptional at learning new novel tasks like these with very few sample points. Could a simple Arduino balance it? Control Theory is the obvious way to go, and having had prior experience tinkering with PID, my overconfident self walked straight into Steve Brunton’s excellent Control Bootcamp…only to return disillusioned. The task was not trivial and involved heavy math. Well if I couldn’t learn it, I’ll let the machines learn it themselves. I tried again with David Silver’s Reinforcement Learning class. Now Q-Learning and Policy Methods based on Markov Decision Processes are cool and all, but they still seemed unwieldy for continu

The Inverted Pendulum Problem with Deep Reinforcement Learning A look into Keras-RL and OpenAI libraries Saif Uddin Mahmud 10 min read · Mar 12, 2019 -- 3 Listen Share A professor of mine introduced me to the rather simple inverted pendulum problem — balance a stick on a moving platform, a hand let’s say. Intuition built on the physics of the “Game Engine” in our head tells us: if the stick is leaning to the left, move your hand to the left; if the stick is leaning to the right, move your hand to the right. We, humans, are exceptional at learning new novel tasks like these with very few sample

saved by

related reading