ep5 - Sean Meyn: Markov chains, networks, reinforcement learning, beekeeping and jazz
inControl
00:00
The Main Challenges With Traditional Performance Guarantees in Machine Learning Algorithms
I challenge you all in computer science to just take a five-state Markov chain and look at the sample path average and send me a finite end bound on the error between your estimate of the mean and the mean, okay? And many will say, oh, it's a special gap. No, it's not. And I, please, let's go read Peter Glenn's book and his paper and Houghton's Bounds for Markov chains. The bounds are worthless, you know. And then that's a simple thing of just averaging a bunch of random variables. When you look at some monster algorithm, like Q learning, you know, whatever, it's so much more
Transcript
Play full episode
