2014 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 / 進化計算 |
|---|
| Outline of Annual Research Achievements |
One of the goals proposed for this project is the development of algorithm portfolios for domain-independent planning. We investigated a new integer-linear programming model of a relaxation of cost-optimal, classical, domain-independent planning. While a naive model of the delete relaxation as integer programming (IP) is impractical, our model incorporates landmarks and relevance-based constraints, resulting in an much more tractable model. We showed that our IP model outperforms the previous state-of-the-art solver for delete-free problems. We then use the LP relaxation of the IP as a heuristics for a forward search planner.
We developed a portfolio approach that combines multiple variations of this LP model and automatically selects the best model to apply to a given planning problem instance, and showed that the resulting domain-independent planner is competitive with previous, state-of-the-art approaches.
In another line of work, we continued to develop evolutionary computation algorithms for black-box optimization problems. We have been developing and improving SHADE, an adaptive, differential evolution (DE) algorithm which has been shown to be competitive with previous state-of-the-art DE variants as well as other evolutionary approaches. In addition to improving SHADE using a technique that combines restarts and population resizing, we studied a particularly difficult class of benchmark problems called hybrid objective functions, and showed that they pose a significant challenge for adaptive DE variants (including SHADE and other previous algorithms).
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The delete relaxation for classical, domain-independent planning has been widely studied because (1) it is useful as the basis for heuristic functions (i.e., lower bounds) for forward state-space search-based approaches to domain-independent planning, and (2) some planning problems can be modeled directly as delete-relaxed problems.
We showed that for many problems, the linear programming (LP) relaxation of the IP model computes the optimal cost of the delete-relaxed problem exactly. We also showed that the LP model is quite successful as the heuristic function for A*-based forward search planning.
Thus, our new integer programming model for the delete relaxation of classical, domain-independent planning is a significant advance in the state of the art. Furthermore, we also showed that a portfolio-based approach to combining multiple, LP models significantly boosts the performance of an A*-based planner using our delete relaxation model. While we have not yet parallelized the portfolio model and currently rely on a sequential, selection-based approach, we plan to develop a parallelized portfolio approach.
|
| Strategy for Future Research Activity |
In 2015, we will continue our development of portfolio-based approaches for (1) black-box optimization using evolutionary computation and (2) domain-independent planning. We will parallelize the portfolio based approach we developed for selecting among linear programming models of the delete relaxation for domain-independent planning. We will also continue to investigate evolutionary methods for black-box optimization. This year, we focused on improving the performance of SHADE and identifying difficult benchmark problems for adaptive differential evolution. We will continue to investigate adaptive differential evolution, particularly focusing on applications where a very limited number of calls to the evaluation function are possible. In 2016-2017, we will focus on investigating the effective use of parallel resources by the search algorithms that we have developed so far in this project.
|
| Causes of Carryover |
In 2014, we purchased a server/workstation for computational experiments, which cost 915408 yen. The remaining amount (33,527 yen) is therefore being carried over for the next fiscal year.
|
| Expenditure Plan for Carryover Budget |
Due to increasing demands for computational resources, in 2015, we may purchase another server for computational experiments. The amount carried over from 2014 will be used to fund part of this purchase, or for trvel costs or office equipment/supplies.
|
