TADA

1
TADA

• Transition Aligned
• Domain Analysis

J. Benton and Kartik Talamadupula and Subbarao
Kambhampati
T
2
Motivation

• Multi-valued Variables

truck truck-1
locations loc-1 loc-6
(at truck-1 loc-1) false
(at truck-1 loc-2) true
( (loc truck-1) loc-2)
(at truck-1 loc-3) false
(at truck-1 loc-4) false
(at truck-1 loc-5) false
(at truck-1 loc-6) false
(and (at ?t ?loc1) (at ?t ?loc2)) ltgt ( (?loc1
?loc2))
3
Motivation

• Action Interleaving
• Durative actions may execute concurrently with
  each other need to model this to extract
  accurate heuristics

2
lstart (T,2)
lend (T,2)
1
ulstart (T,1)
ulend (T,1)
T
4
Motivation

• Lookahead during search, following a YAHSP-style
  approach

5
Deadline Goals with Utility

• Give an estimate of establishment time eT
• as against the actual deadline eD
• Model reward or penalty in terms of their
  difference
• Reward min(0, R - k (eT eD))
• R is the reward accrued for achieving the
  deadline goal.
• If achieved by time eD, full reward. Else
  diminishing reward until 0.

6
Example Problem
I
1
2
Causal Graph

G
1
2
truck
(load/unload)
package
7
Deadline Goals Example
package
deadline t20
truck
loadstart/unloadstart
2
2
unloadend
movestart
time counter
moveend
loadstart
1
t0
t10
1
P14
moveend
movestart
t22
unloadstart
t26
T
loadstart/unloadstart
loadend
4
4
lstart (T,2)
lend (T,2)
ulstart (T,1)
ulend (T,1)
10
8
mstart (1,2)
mend (1,2)
mstart (2,1)
mend (2,1)
8
FUTURE WORK

• International Planning Competition 2008
• Tracks
• Sequential
• Net-Benefit
• Temporal
• Problem decomposition for better analysis and
  more accurate estimates
• Considering a goal ordering based on causal
  dependencies in the domain

9
Summary

• Planning with multi-valued variables
• Domain Transition Graph representation
• Extended DTGs to handle durative actions
• Useful to estimate deadline establishment times
  can be used for soft deadline goals with
  diminishing utility
• Implement a lookahead strategy for search
• Try to find a satisficing solution quickly
• Make effective use of action interleaving
• Handles net-benefit and temporal problems.