Introduction to Reinforcement Learning#
DL4Sci Summer School 2026 · Lawrence Berkeley National Laboratory · 90 minutes
Welcome! These are the companion notes for the Intro to Reinforcement Learning lecture. The talk assumes you are comfortable with deep learning (gradients, losses, training neural networks) but assumes no prior RL background.
The one-sentence version: supervised learning learns from instruction (labels) reinforcement learning learns from evaluation (consequences). This lecture covers how that idea goes from gridworlds to the training of modern reasoning models and agents.
Quick links#
- 📊 Slides: (link here)
- 📓 Demo notebook (CartPole / gridworld): (Colab badge / link here)
- 🏫 School website: dl4sci-school.lbl.gov
Roadmap of the lecture#
| Time | Section |
|---|---|
| 0:00–0:10 | Why RL, why now |
| 0:10–0:30 | Fundamentals: MDPs, values, Bellman |
| 0:30–0:50 | Deep RL: DQN, policy gradients, PPO |
| 0:50–1:10 | RL for LLMs, reasoning, and agents |
| 1:10–1:20 | RL in science |
| 1:20–1:30 | Wrap-up, resources, Q&A |
How to use these notes#
The main text follows the lecture at lecture pace. Expandable “Deep dive” blocks contain derivations and extra material that we will skip live — open them if you want the details.