J. Luo et al. / Journal of Information & Computational Science 10:5 (2013) 1303–1313 1305
for metho d is a key problem for HTN planning [12]. As the typical Abstract-HTN algorithm
[1] showed in Fig. 1, it’s adopted to ensure the complete decomposition of the HTN and return
a decomposition tree. The program in rectangle
1
○ is the constraint consistency detection.
All the available decomposition methods for task node u compose the set active , which may
contain multiple elements. The non-deterministical choosing operation in rectangle
2
○ should
not only make the amount of decomposition trees of the HTN increase exponentially with the
number of elements in set active [11], but also bring some backtracking operations for further
decompositionas the compound tasks in the abstract level are always not independent and the
high-level method selection may cause the failure decomposition of the low-level compound tasks.
Abstract-HTN (s, U
′
, C
′
, O, M)
If (U, C) can be shown to have no solution
then return failure
else if U is primitive then
if (U, C) has a solution then
non-deterministically let π be any such solution
return π
else return failure
else
choose a no-primitive task node u ∈ U
active ← {m ∈ M |task(m) is unifiable with t
u
1
○
if active = ∅ then
non-deterministically choose any m ∈ active
2
○
σ ← an mgu for m and t
u
that renames all variables of m
(U
′
, C
′
) ← δ(σ(U, C), σ(u), σ(m))
(U
′
, C
′
) ← apply-critic (U
′
, C
′
)
return Abstract-HTN (s, U
′
, C
′
, O, M)
else return failure
Fig. 1: Abstract-HTN program
In order to search the desired solution, knowledge learning and Heuristic method are introduced
into the HTN planners [10, 13, 14, 17]. However, because those algorithms always offer preferable
decomposition strategy based on current planning states, a large number of backtracking oper-
ations are needed. SONG et al. [16] adopt a segmented backtracking method, which can build
a number of plans quickly by utilizing pieces of action-sequence produced during search process.
This algorithm can provide powerful supporting to optimum plan generation and plan evaluation
in Ordered Task Decomposition (OTD)-based [21, 22] HTN planning. As to reduce the search
space, Reiko Tsuneto et al. [11] preserve all possible decomposition strategies in an AND/OR
graph, and prune the invalid branches with domain knowledge to dwindle the search space.
All the strategies mentioned above are searching the valid solution. However, the solution may
contain some redundant actions or be profuse in resource consumption. Alexandru Horia Brie et
al. [18] try to use messy genes to solve the optimum executable COAs search problem for HTN
planning. It focuses on mining the optimum combination of primitive tasks to achieve the best
effects without taking the mission decomposition process into account. In other words, we don’t