3
$\begingroup$

There are many methods of exploring in a Reinforcement Learning setting but two of the most used ones are Ornstein Uhlenbeck (OU) processes and epsilon-greedy approaches. Could anyone elucidate the major advantages/disadvantages of using one over the other?

One of the things associated with OU processes is that you need to two additional parameters to bias exploration which might mean additional tuning. I'd be glad if someone could help!

$\endgroup$
2
  • $\begingroup$ Welcome to Robotics sg_robs, but I'm afraid that questions like this really aren't a good fit for a stack exchange site. We prefer practical, answerable questions based on actual problems that you face. Take a look at How to Ask and tour for more information on how stack exchange works. Also, the Robotics question checklist has good advice on how to write a good question. If you edit your question to fit our community guidelines we can reopen it for you. $\endgroup$
    – Mark Booth
    Nov 25, 2016 at 17:19
  • $\begingroup$ I'm sure this question is perfectly fine on Artificial Intelligence or Data Science SE. It's sad the question was closed here... $\endgroup$
    – Blaszard
    Aug 10, 2018 at 18:57

1 Answer 1

-1
$\begingroup$

Ornstein Uhlenbeck processes and epsilon-greedy are no antagonist but epsilon-greedy is a algorithm for finding actions for an Ornstein Uhlenbeck process. Reinforcement learning is another term for a Function Approximation, also known as metaprogramming. The idea is not to focus on the domain which can be control of a car or multi-arm-bandit problem but to recognize the data as a stochastic process. Like all blackbox optimization algorithm the problem is the huge search space. There are unlimited number of possibilites for a function to control the system. According to my knowledge, there is no example in which reinforcement learning ever solved a problem.

$\endgroup$
2
  • $\begingroup$ Although I appreciate your insight on reinforcement learning in general, your answer does not address what I have asked specifically. Additionally, to see if RL ever solved a problem, please read the vast literature out there, beginning with this paper - arxiv.org/pdf/1312.5602.pdf $\endgroup$
    – sg_robs
    Nov 24, 2016 at 1:30
  • $\begingroup$ DeepMind Technologies works with psychology not with computerscience. They are using electrodes for measuring human-performance on playing Atari games and want to reproduce these signals with stochastics models. The aim is not "optimal control", the aim is to explain what humans are thinking. $\endgroup$ Nov 24, 2016 at 12:38

Not the answer you're looking for? Browse other questions tagged or ask your own question.