Question about elevators

[Andrey Kolobov]

Hi Scott,

The ppddl versions of Elevators problems have actions with many
preconditions like

(when (and (not (elevator_dir_up__e0)) (not
(elevator_at_floor__e0_f0)) (elevator_at_floor__e0_f1)
(elevator_at_floor__e0_f2) (elevator_at_floor__e0_f3) (not
(person_in_elevator_going_down__e0)) (not (person_waiting_down__f1))
(person_waiting_down__f2) (not (elevator_closed__e0))) (probabilistic
1.0 (person_in_elevator_going_down__e0) ))

Aren't they requiring the elevator to be on several floors at once
(e.g., on 3 floors in the above example)? If so, these preconditions
are contradictions and the corresponding effects are superfluous. Is
there any way you could remove them?

Cheers,

Andrey

[Scott Sanner]

Hi Andrey,

> Aren't they requiring the elevator to be on several floors at once
> (e.g., on 3 floors in the above example)? If so, these preconditions
> are contradictions and the corresponding effects are superfluous. Is
> there any way you could remove them?

Thanks a lot for pointing this out and for the example.  

PPDDL does not provide an explicit mechanism for specifying illegal states, but from your example, it is obvious that I can check the partial state instantiated by the preconditions to see if they violate RDDL state constraints and then simply remove them if so.  (The translations are still correct though as is -- unreachable states are simply unreachable.)

I *will* certainly make this enhancement for the *final competition* translations, but for now I'm going to leave the test competition domains and translations as is since some competitors have already completed or are now completing their test runs.

In the meantime for the test competition, I note that if you are using any sort of data structure like a tree or decision diagram for the transition evaluation, your planner should not incur any efficiency loss (aside from initial compilation effort) from these superfluous when conditions.

Cheers,

Scott

[Scott Sanner]

Actually an addendum to this last email: checking whether partial states (given by the preconditions) violate state constraints will require something like a SAT solver in the general case (to evaluate all extensions of the partial state).  However, for certain types of constraints like mutual exclusivity that are involved in your example, perhaps I can detect these constraints separately and handle them more efficiently than with a general SAT solver.  I'll have to think about this; I'll post my final solution (if any) here when I have it.

Cheers,

Scott

[Andrey Kolobov]

Thanks!

I have another small remark; the probabilities for some conditional
effects of come actions of some problems don't add up to 1 (e.g.,
game_of_life_mdp__2.ppddl, line 299). This is probably due to some
small issues with precision or output in the translator -- the sum of
probabilities is actually extremely close to 1. This is a bit of a
pain because our planner has some error-checking code that trips up on
these artifacts.

Would you be able to fix that somehow too?

Best,

Andrey

[Scott Sanner]

For now I suggest simply checking equality for floating point numbers within 1e-10 precision.  In the next translator version I will use a more numerically stable method of normalization.

Cheers,

Scott

But in the next translation I'll tweak the generation to 

[Ashwin NR]

FYI, the SPUDD file format can input unnormalized CPTs if every variable is suffixed with the keyword "'unnormalized". I had changed RDDL2Format accordingly. The parser performs normalization. You might be able to the same modification to the PPDDL parser.

Cheers,

Aswin NR

Oregon State U.