From: Anders Sandberg (asa@nada.kth.se)
Date: Fri Jun 08 2001 - 05:27:38 MDT
torsdagen den 7 juni 2001 19:50 Eliezer S. Yudkowsky wrote:
> Would you people please, please, PLEASE get into the habit of
> distinguishing between subgoals and supergoals?
>
> Subgoal content: Events or states whose desirability is contingent on
> their predicted outcomes.
> Supergoal content: Events or states with "inherent" desirability, i.e.
> desirability not contingent on predicted outcome.
Well, in my reinforcement learning framework subgoals are simply whatever
local goals the policy function sets up ("go through that door", "construct a
terraforming device") as a result of the value maximization (the supergoal).
Some goals are the result of a primary reinforcement signal, others due to
secondary signals. I don't see any contradiction between this perspective and
yours.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:08:01 MST