Clean and complete the current MBCHAC implementation
The current MBCHAC implementation is, in fact, HAC as it does neither use curiosity, nor background planning. Fix this and make an MBCHAC implementation from it.
Along these lines, it may make sense to implement an abstract model class that is general for all kinds of hierarchical methods. This could be stored in a different ideas_baselines
subdirectory and MBCHAC and other HRL methods could inherit from it. The result would probably be a much cleaner and reduced code of the rather large MBCHAC class.
-
(re-)implement HAC -
Check performance and make sure it is at least as good as with HAC -
RE-implement forward model -
Re-implement curiosity based on prediction error -
Re-implement background planning etc.
Edited by Manfred Eppe