Possible bug in the perseus translation of the elevators_pomdp domain

5 views
Skip to first unread message

Sylvie C.W. Ong

unread,
Mar 8, 2011, 4:58:08 PM3/8/11
to IPPC 2011
Hi Scott,

Thanks for making available the problems files for the test
competition! I took a look at some of the problem files in
test_comp.tgz and found what could possibly be a bug in how concurrent
(RDDL) actions are translated into actions in symbolic perseus.

In the file elevators_pomdp.elevator_inst_pomdp__2.sperseus, these are
two distinct actions:
close_door__e1___close_door__e0
close_door__e0___close_door__e1

My understanding is that in both cases, the action is that
close_door(e1) and close_door(e0) are executed simultaneously, without
any priority and without any sequential order. So in the perseus file
there should only be one of the above two actions, and not both. Is my
understanding correct?

thanks and best regards
sylvie

Sylvie C.W. Ong

unread,
Mar 8, 2011, 5:24:43 PM3/8/11
to IPPC 2011
Hi Scott,

Just an addendum/clarification. What I mentioned about the actions in
elevators_pomdp.elevator_inst_pomdp__2.sperseus occurs for other
combinations of actions as well, for example, these two action are
distinct in the .sperseus file:
open_door_going_down__e0___close_door__e1
close_door__e1___open_door_going_down__e0

The same thing goes for the other instances of the elevators_pomdp
domain which have concurrent actions.

It also occurs in the traffic_pomdp domain, which seems to be the only
other domain which has concurrent actions. In
traffic_pomdp.traffic_inst_pomdp___1.sperseus, these two actions are
distinct:
advance__ia2a4___advance__ia2a8
advance__ia2a8___advance__ia2a4

thanks and best regards
sylvie

Scott Sanner

unread,
Mar 8, 2011, 7:09:45 PM3/8/11
to ippc...@googlegroups.com, Sylvie C.W. Ong
Hi Sylvie,

Thanks a lot for pointing this out.

This is not quite a bug in that these are just redundant actions with identical transition functions, but these redundant actions are certainly unnecessary.  

I will filter out these actions in the next version of the translator and test_comp domains by modifying ActionGenerator.getLegalBoolActionMap(...).

===

ALL COMPETITORS: before I regenerate the test_comp domains, can those who are parsing the rddl_prefix language send me an email to let me know they are using it?  (Patrick / Thomas: I have you guys down already.)

Cheers,
Scott
Reply all
Reply to author
Forward
0 new messages