Policy Gradient with PyTorch

Hugging Face · 4 years ago · 45

Hugging Face published Unit 5 of its free Deep Reinforcement Learning course, introducing policy gradient methods as an alternative to value-based approaches like Deep Q-Learning. The unit covers Reinforce (Monte Carlo Policy Gradient), which optimizes a policy directly using gradient ascent to increase the probability of rewarding actions, implemented in PyTorch. Students implement and test agents on CartPole-v1, PixelCopter, and Pong environments, with results comparable on a public leaderboard.