Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15 - 2019

Title : Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15 Author(s): Lex Fridman Link(s) : https://www.youtube.com/watch?v=Er7Dy8rvqOc

(Temporal and spatial) Abstractions and decompositions (e.g. goal decomposition) - involves reducing the state space, and reducing the horizon. Decompositions of goals. E.g. of time abstraction - saying afternoon instead of specific time, and spatial abstraction - room of a house instead of pose information.

MDPs model uncertainty on how the future unfolds, not the present state uncertainty, while POMDPs also model the uncertainty in the current state. Optimal planning is often intractable - but this doesn't mean we should not use POMDPs in practice, since the real world can be tackled with more and more approximations.

We don't have good approximate solution concepts for very difficult problems. We need to turn engineering into science from this perspective.

Belief space - having a distribution over the space of how the world is right now. The control problem now involves knowing how the agent's actions affect the outside world and also the agent's own belief/understanding of the outside world.

Perception is harder than planning - the main problem is representation.

Story on the founding of JMLR at 45:00 - Leslie was the start of it and served as its editor-in-chief.