Reinforcement Learning Applications

Speaker

Carol Cuesta-Lazaro, IAS/Flatiron

A hands-on reinforcement learning tutorial that builds policy-gradient methods from scratch with LunarLander as the running environment, progressing from vanilla REINFORCE to variance-reduced policy gradients and actor-critic learning.

Topics Covered

• Using LunarLander to connect the agent-environment loop to code
• Implementing a policy network and sampling actions with PyTorch
• Training vanilla REINFORCE from trajectory-level returns
• Reducing variance with reward-to-go, discounting, and normalized advantages
• Building actor-critic methods with a learned value-function baseline
• Comparing learning curves across REINFORCE, improved REINFORCE, and actor-critic

Meeting Connection

Join the Meeting