Our last domain is Zeno Travel, based on a domain used in IPC-3.
Problem instances of
this
domain involve using airplanes to move people between cities. An
airplane requires fuel to fly. It can be flown at two different
speeds—the higher speed requiring more fuel.
Our problem instance used one aircraft, two people,
three cities and seven fuel levels. The actions of the domain are
“start-boarding”, “complete-boarding”, “start-debarking”,
“complete-debarking”, “start-refueling”,
“complete-refueling”, “start-flying”, “complete-flying”,
“start-zooming”, and “complete-zooming”.
The initial
configuration specifies the location of the plane, the initial fuel
level of the plane and the location of all people (as well as some
initializations to allow for arithmetic type operations on the
fuel-level objects). The goal configuration specifies a destination
for the plane and destinations for all people.
The noise in this domain comes from the family of “complete-*X*”
actions. Each time a “complete-*X*” action is executed it will have the
desired effect with probability 1/*k*
for some positive integer *k*
(note
that *k*
is a function of the action executed, specifically *k* = 20
for
“complete-boarding” and *k* = 30
for “complete-debarking”). If the
desired effect is not achieved then there is no effect, and this
occurs with probability 1 − (1/*k*).
This structure is meant to
represent actions with random duration. Each “durative” action
*X*
is
represented by two primitive actions “start-*X*” and
“complete-*X*”, giving *X*
a duration that is geometrically
distributed.

Ultimately, this problem presented no real challenge because we neglected to include action costs. Since actions have either standard desired effect or none at all, a planner can simple continue to execute an action until its effect is achieved, without incurring any cost.

Håkan L. S. Younes

2005-12-06