Lex Fridman Podcast Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics
4 snips
Mar 12, 2019 
Leslie Kaelbling, a leading professor of computer science at MIT and expert in reinforcement learning, discusses her journey in AI, sparked by 'Gödel, Escher, Bach.' She explores the connection between philosophy and computer science, and the cyclical evolution of artificial intelligence. The conversation dives into challenges in symbolic reasoning, the quest for optimal solutions in robotics, and the complexities of machine perception. Kaelbling also critiques academic publishing, advocating for open-access to enhance research in the rapidly changing landscape of AI.
AI Snips
Chapters
Books
Transcript
Episode notes
Belief Space vs. State Space
- Belief space focuses on managing an agent's beliefs about the world, represented as probability distributions.
- Controlling beliefs allows for deliberate information gathering and accounts for uncertainty in actions.
Hierarchical Planning and Abstractions
- Hierarchical planning tackles long-horizon problems by abstracting state space and time, breaking tasks into sub-goals.
- Predicting sub-goal difficulty is crucial for planning, even without detailed information.
Human Life Beyond Planning
- Human life is too complex to be modeled solely as a planning problem.
- Intelligence requires a combination of reasoning styles, representations, and learning approaches.
