2015 Fiscal Year Research-status Report
大規模並列アルゴリズムポートフォリオの構築と実行の研究
| Project/Area Number |
25330253
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
福永 ALEX 東京大学, 総合文化研究科, 准教授 (90452002)
|
|---|
| Project Period (FY) |
2013-04-01 – 2017-03-31
|
| Keywords | 探索アルゴリズム / 自動行動計画 / 人工知能 |
|---|
| Outline of Annual Research Achievements |
We have been developing portfolio-based approaches for domain-independent planning, the problem of finding a sequence of operators which achieve a set of goals. In 2015, we continued development of an approach to cost-optimal, classical, domain-independent planning which uses a linear programming (LP) model of the delete relaxation model of the original planning problem as a heuristic function for best-first search. We analyzed the model both theoretically and empirically. We showed that a portfolio based approach which which initially generates multiple LP models of a given planning problem (where each model incorporates a different set of constraints) and automatically selects one of the candidates prior to search results in performance which is competitive with state-of-the-art domain independent planners. We analyzed the prediction accuracy of this portfolio (i.e., the fraction of problems where the best available configuration was selected), and showed that correct configuration was selected in the majority of instances.
We also began investigating methods for optimizing plans found by domain-independent planners. When the time available to search for a solution is limited, domain-independent planners use search algorithms that seek to find solutions quickly, while sacrificing the quality of the plan that is found. We investigated methods which apply a portfolio of post-processing algorithms in order to improving the quality of suboptimal plans, and showed that this approach consistently improves the quality (according to a given cost function) of suboptimal plans.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
We have successfully shown that a portfolio-based approach which combines multiple linear programming models as a heuristic function for best-first search is an effective method for domain-independent planning. Furthermore, each of the component algorithms (LP models) used in this model is generated automatically from the planning problem's formal description expressed in the standard Planning Domain Descrtion Language (PDDL). Thus, this is a successful instance of an automatically generated portfolio algorithm, which was one of the primary goals of this project.
We have also shown that a portfolio approach can be used to significantly improve the quality of plans generated by a state-of-the-art, satisficing planner.
The remaining aspect of this project which has not yet been addressed is the paralellization of automatically generated portfolios. We plan to address this next year.
|
| Strategy for Future Research Activity |
In 2016, we will continue to develop portfolio methods for domain-independent planning and other search problems, including black-box optimization, as well as methods for effectively parallelizing search algorithms.
In particular, we will investigate how multiple search algorithms which work at different levels of abstraction can be combined. For example, algorithms which attempt to quickly find satisficing (possibly suboptimal) solutions can be combined with algorithms which require more time but find higher quality (or optimal) solutions. Possibilities for combining such algorithms include either applying them in sequence or embedding them in each other (e.g., embedding a fast algorithm as a subroutine used by the slow algorithm).
In another line of work, we will investigate mechanisms for improving the efficiency of parallel search algorithms, including parallel portfolios. Recent work on parall search has focused on minimizing the amount of wasted work (i.e., unnecessary work that is performed by the parallel algorithm but not performed by an efficient serial algorithm). In many problems, an additional source of parallel overhead is the communications related overhead incurred when transferring work from one processor to another. We will investigate methods for reducing communications overhead in parallel search algorithms.
|
| Causes of Carryover |
In 2015, we purchased a workstation for computational experiments which cost 968868yen. The remaining amount of 64,659yen is being carried over for the next fiscal year.
|
| Expenditure Plan for Carryover Budget |
We plan to combine the amount carried over from 2015 with the amount received in 2016 in order to pay for travel expenses for 2016. The purpose of such travel will be for presenting our research results at international conferences.
|
