Audio of [Apr 21, 2009] More on planning: Regression search as a "lifted" search, and also regression as a search in the space of belief states with an over-easy termination crterion. Discussion of actions with conditional effects. Conditional effects as an expressiveness issue (a single conditional effect operator can be compiled down to an exponential number of non-conditional effect operators); progression with conditional effects is easy, while regression becomes more involved. Regression in the presence of conditional effects also leads to partial states that are sets of CNF claues (which means we now are in full belief space). Except that the disjunction can be pushed into init state as long as the init state is complete.
Planning graph heuristics continued; h-sum, h-level and their
tradeoffs. H-relax as a superior middle-ground. Extracting relaxed
plans. PG heuristics with progression vs. regression planners. Issues
of PG
heuristics with action costs.