We approach the problem of finding plans based on multiple optimization criteria from what would seem an unlikely direction: find one valid plan as quickly as possible, then stream essentially all plans that improve on the current best plan, searching over incrementally longer length plans. This approach would be computationally prohibitive for most planners, but we describe how, by using a concise trace of the search space, the PEGG planning system can quickly generate most, if not all, plans on a given length planning graph. By augmenting PEGG with a branch and bound approach the system is able to stream parallel plans that come arbitrarily close to a userspecified preference criteria based on multiple factors. We demonstrate in preliminary experiments on cost-augmented logistics domains that the system can indeed find very high quality plans based on multiple criteria over reasonable runtimes. We also discuss directions towards extending the system such that it is not restricted to Graphplan’s scheme of exhaustively searching for the shortest step-length plans first.
[1]
Subbarao Kambhampati,et al.
Exploiting Symmetry in the Planning graph via Explanation-Guided Search
,
1999,
AAAI/IAAI.
[2]
Terry L. Zimmerman,et al.
Using Available Memory to Transform Graphplan's Search
,
2003,
IJCAI.
[3]
Terry L. Zimmerman,et al.
Effective Interplay of State Space and CSP Views in a Single Planner
,
2001
.
[4]
Blai Bonet,et al.
Planning as Heuristic Search: New Results
,
1999,
ECP.
[5]
Subbarao Kambhampati,et al.
Extracting Effective and Admissible State Space Heuristics from the Planning Graph
,
2000,
AAAI/IAAI.