search

ramunter / masters-thesis

0 stars 0 forks source link

Bayesian Regression in RL #2

Closed ramunter closed 5 years ago

ramunter commented 5 years ago

I was looking into using Bayesian regression*BR) to solve the corridor problem. However, it seems to a bit complicated since we have an online learning problem. Given a dataset there seems to be an analytical solution to BR. However, this is not the case when learning online.

I'm having problems finding any "finished" tool for doing this, so I think I'll have to do it myself and am currently reading a bit up on TF.probability.

So my question is basically a sanity check. Do you agree that this is the right way to go or have I overlooked something simpler?

This seems like a good place to learn about tf probability if you are interested)

thigm85 commented 5 years ago

You are write in your assessment. Online learning in Bayesian setting is not trivial. As discussed today, I suggest start easy with conjugate models.

This might be useful for you: https://www.cs.ubc.ca/~murphyk/Papers/bayesGauss.pdf

thigm85 commented 5 years ago

Forgot to mention, you should also look for Kalman Filter.