I'm studying the problem of learning the predator hunter, and many papers suggest interesting ways to learn a policy, They all feel that they already know that the graph structure describes actions in the domain. For example, Dietetrich describes a complex graph of functions and sub-functions for a simple taxi domain, but how this article was not discovered. How do you learn the hierarchy of this article, and not just the policy?
The graphic in the Maxacu of Detective is manually created, it is considered to be a function for the system designer, In the same way that a representation is coming with place and reward functions.
Depending on what you are trying to achieve, you want to disintegrate automatically, learn state space, relevant attributes, or transfer experience to more complex people than simple tasks.
I suggest you start reading such a paper that refers to the MAXQ you linked. Without knowing what you want to achieve, I can not be very compliant (and I'm not really on top of current RL research), but you get relevant ideas in the work of Luo, Bell and Embedded You can. Letter by McCullum or Madan; Howley
Comments
Post a Comment