The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Off-Line, Off-Policy RL for Real-World Decision Making at Facebook - #448
Jan 18, 2021
In this discussion, Jason Gauci, a Software Engineering Manager at Facebook AI, dives into the complexities of their Re-Agent reinforcement learning platform. He highlights its role in real-world decision-making, including user engagement strategies for Facebook notifications. The conversation explores counterfactual causality and safety in social network decision-making. Jason also shares insights on differentiating online/offline training models, emphasizing the impact of reinforcement learning on small businesses and the future of AI in eCommerce.
AI Snips
Chapters
Transcript
Episode notes
Early Programming
- Jason Gauci's interest in computers began with an Atari 400/800 found at a garage sale.
- The computer only had BASIC, sparking his curiosity and leading him to create an arm wrestling game.
Games as Research Platforms
- Games like chess and Go are valuable for decision-making research.
- They distill complex real-world scenarios, like battles, into their core strategic essence.
Go-Playing AI
- Jason Gauci's PhD research involved a Go-playing AI that learned on smaller boards.
- It progressively scaled to larger boards, similar to how humans learn.