#66 – Michael Cohen on Input Tampering in Advanced RL Agents
Hear This Idea
00:00
The Problem With Gaussian Processes in Machine Learning
Michael Cohen: I've been trying to do some pessimistic reinforcement learning in practice, having done some theory on that. And it's tricky. So I've been using Gaussian processes for the pessimistic reinforcement learners model of the world. He says as you add more and more and more points, nothing breaks. With a Gaussian process, it's technically an infinite dimensional multivariate Gaussian.
Play episode from 02:27
Transcript
