180: Reinforcement Learning
Mar 17, 2025•1 hr 52 min•Ep. 180
Listen to this episode in Metacast mobile app
Don't just listen to podcasts. Learn from them with transcripts, summaries, and chapters for every episode. Skim, search, and bookmark insights. Learn more
Episode description
Intro topic: Grills
News/Links:
- You can’t call yourself a senior until you’ve worked on a legacy project
- Recraft might be the most powerful AI image platform I’ve ever used — here’s why
- NASA has a list of 10 rules for software development
- AMD Radeon RX 9070 XT performance estimates leaked: 42% to 66% faster than Radeon RX 7900 GRE
Book of the Show
- Patrick:
- The Player of Games (Ian M Banks)
- https://a.co/d/1ZpUhGl (non-affiliate)
- The Player of Games (Ian M Banks)
- Jason:
- Basic Roleplaying Universal Game Engine
Patreon Plug https://www.patreon.com/programmingthrowdown?ty=h
Tool of the Show
- Patrick:
- Pokemon Sword and Shield
- Jason:
- Features and Labels ( https://fal.ai )
Topic: Reinforcement Learning
- Three types of AI
- Supervised Learning
- Unsupervised Learning
- Reinforcement Learning
- Online vs Offline RL
- Optimization algorithms
- Value optimization
- SARSA
- Q-Learning
- Policy optimization
- Policy Gradients
- Actor-Critic
- Proximal Policy Optimization
- Value optimization
- Value vs Policy Optimization
- Value optimization is more intuitive (Value loss)
- Policy optimization is less intuitive at first (policy gradients)
- Converting values to policies in deep learning is difficult
- Imitation Learning
- Supervised policy learning
- Often used to bootstrap reinforcement learning
- Policy Evaluation
- Propensity scoring versus model-based
- Challenges to training RL model
- Two optimization loops
- Collecting feedback vs updating the model
- Difficult optimization target
- Policy evaluation
- Two optimization loops
- RLHF & GRPO
For the best experience, listen in Metacast app for iOS or Android