Sequential decision analytics

[Warren Powell]

Colleagues,

While I am on the topic of sharing the first chapter of my book (click here if you did not see the previous email), I would like to share another perspective, which is that "RL problems" are basically sequential decision problems:

decision, information, decision, information, ...

which spans every RL problem I have seen (including bandit problems), and some I haven't seen associated with the RL literature (e.g. stochastic gradient algorithms, where the "decision" is the stepsize).  

I have a video that presents this idea

http://tinyurl.com/sdafieldyoutube

I compare "machine learning" and "sequential decisions" to make the point that I think these problems are distinct (the video does it best - see the five layers of intelligence).  I also make the point that I think we need a field that I would call "sequential decision analytics" which stands on the shoulders of machine learning, optimization and simulation (and yes, it draws heavily on concepts familiar to the RL community, but with some extensions).  I summarize the field at the webpage:

http://tinyurl.com/sdafield

I would love feedback on this idea.

Warren

------------------------------

Warren B. Powell

Chief Analytics Officer, Optimal Dynamics

Professor Emeritus, Princeton University

http://www.castlelab.princeton.edu