• Search Research Projects
  • Search Researchers
  • How to Use
  1. Back to project page

2016 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

Research Project

Project/Area Number 25330253
Research InstitutionThe University of Tokyo

Principal Investigator

福永 ALEX  東京大学, 大学院総合文化研究科, 准教授 (90452002)

Project Period (FY) 2013-04-01 – 2018-03-31
Keywords人工知能
Outline of Annual Research Achievements

A significant technical issue when executing parallel portfolio algorithms is the reduction of communications overhead among processors involved in the parallel computation.
One well-known approach to work distribution among processors is hash-based work distribution, in which work units are assigned to processors according to a hash value (signature) computed based on features of the work unit. In earlier work, we investigated Zobrist hashing, which achieve good load balancing at the cost of high communications overhead.
This year, we developed Abstract Zobrist Hashing, a new method for distributing work among processors which significantly reduces communications overhead compared to Zobrist hashing. This method computes hash signatures based on "abstract features", i.e., projectsions of a state rather than the state itself. We evaluated Abstract Zobrist Hashing on several classes of benchmark problem domains: sliding tiles puzzles (16-puzzle and 24-puzle), multiple sequence alignment, and a set of domain-independent planning instances from the International Planning Competition. We showed that compared to previous work distribution methods, Abstract Zobrist Hashing achieved significantly better parallel efficiency.

Current Status of Research Progress
Current Status of Research Progress

2: Research has progressed on the whole more than it was originally planned.

Reason

In this project, we have investigated the following algorithmic issues related to parallel algorithm portfolios:
We investigated randomized portfolios models for evolutionary evolution, specifically a Randomized Heterogeneous Model which randomly assigns random control parameters for distributed evolutionary algorithms on each processor.
We also investigated a portfolio based approach for domain-independent classical problem which uses multiple integer/linear programming models, each with a different set of constraints is generated for a given planning problem instance.
We have also investigated methods for efficiently distributed workloads among parallel processors which reduce communications overhead.

Strategy for Future Research Activity

We will continue investigating parallel workload distribution mechanisms. This will extend the work on Abstract Zobrist Hashing which was perfomed this year. Note that this research project was originally scheduled to be completed in fiscal year H28, and was extended into fiscal year H29 because a conference trip to present results of this project which was planned to be funded by this project was cancelled due to illness. Thus, next year, work on this project will be combined and integrated with its successor project on heterogeneous parallel search algorithms.

Causes of Carryover

Originally, funds from this research project were planned to be used in order to attend a conference held in February, 2017. This trip was cancelled due to illness, and therefore, discussions regarding this project which were planned during that conference needed to be postponed until the next fiscal year (H29). Therefore, an extension was requested.

Expenditure Plan for Carryover Budget

The plan is to use the funds to travel to a conference during fiscal year H29 in order to present results related to this research project as well as engage in discussion of this research project with international researchers.

  • Research Products

    (2 results)

All 2016

All Journal Article (2 results) (of which Int'l Joint Research: 1 results,  Peer Reviewed: 2 results,  Open Access: 2 results)

  • [Journal Article] Abstract Zobrist Hashing: An Efficient Work Distribution Method for Parallel Best-First Search2016

    • Author(s)
      Jinnai Yuu, Alex Fukunaga
    • Journal Title

      Proceedings of 30th AAAI Conference on Artificial Intelligence (AAAI-2016)

      Volume: 1 Pages: 717-723

    • Peer Reviewed / Open Access / Int'l Joint Research
  • [Journal Article] Automated Creation of Work Distribution Functions for Parallel Best-First Search2016

    • Author(s)
      Jinnai Yuu, Alex Fukunaga
    • Journal Title

      Proceedings of International Conference on Automated Planning and Scheduling

      Volume: 1 Pages: 184-192

    • Peer Reviewed / Open Access

URL: 

Published: 2018-01-16  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi