Parallel Restarted Search

Authors: Andre Cire, Serdar Kadioglu, Meinolf Sellmann

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We develop a simple technique for parallelizing restarted search deterministically and demonstrate experimentally that we can achieve near-linear speed-ups in practice.
Researcher Affiliation Collaboration Andre A. Cire Carnegie Mellon University Pittsburgh, PA 15213, USA acire@andrew.cmu.edu Serdar Kadioglu Oracle America Inc. Burlington, MA 01803, USA serdar.kadioglu@oracle.com Meinolf Sellmann IBM Research Yorktown Heights, NY 10598, USA meinolf@us.ibm.com
Pseudocode No The paper describes the Luby strategy and its parallelization in narrative text and figures, but does not include formal pseudocode blocks or algorithms.
Open Source Code No The paper does not provide any explicit statements about releasing their source code or links to a code repository.
Open Datasets Yes We first evaluated our proposed approach on 3 distinct problem domains: Quasi-Group Completion (QCP), Magic Squares, and Costas Arrays (Costas 1984). Our last experiment was performed on the instances from the Minizinc Challenge using Gecode solver.
Dataset Splits No The paper mentions generating instances and using instances from the Minizinc Challenge, but does not provide specific details on train/validation/test splits, percentages, or explicit methodologies for data partitioning.
Hardware Specification Yes All experiments were conducted on Intel Xeon 2.4 GHz computers with varying number of cores as required by each run.
Software Dependencies No The paper states "We employed Gecode solver" and refers to "Gecode (Schulte, Tack, and Lagerkvist 2010)" in the text, but does not provide a specific version number for the Gecode solver or any other software dependency.
Experiment Setup Yes We employed Gecode solver with 2,000 second timeout and all experiments presented use a Luby scaling factor of 64. This solver uses a search strategy based on active failure count with a decay of 0.99 and geometric restarts (scale factor 1,000 and base 2). We use a static failure count reset at the beginning of each restart. To randomize the restarts, again, the first five branching variables are selected uniformly at random.