Towards Clause-Learning State Space Search: Learning to Recognize Dead-Ends

Authors: Marcel Steinmetz, Joerg Hoffmann

AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that this can be quite powerful. On problems where dead-ends abound, the learning reliably reduces the search space by several orders of magnitude. and Figure 2: Empirical results. Left: Coverage. Best per-domain results highlighted in boldface. DFS-CL is our approach, α = with unlimited learning, α = 1 without learning. Other abbreviations see text. For each domain, there are 30 or 5 base instances as specified (union of Nakhost et al. s small and large test suites, for No Mystery and Rovers). For each base instance and value of W, the resource budget is set according to W. Right: Search space size for DFS-CL with learning (α = , y-axis) vs. without learning (α = 1, x-axis).
Researcher Affiliation Academia Marcel Steinmetz and J org Hoffmann Saarland University Saarbr ucken, Germany {steinmetz,hoffmann}@cs.uni-saarland.de
Pseudocode Yes Algorithm 1: Generic forward search algorithm with dead-end identification and learning. and Algorithm 2: Refining C for ˆS with recognized neighbors ˆT.
Open Source Code No The paper does not provide any specific link, explicit statement, or mention of supplementary materials for the open-source code of the methodology described. It only states, "Our implementation is in FD (Helmert 2006)," referring to a third-party planner.
Open Datasets Yes We use the benchmarks by Nakhost et al. (2012), which are especially suited as they are controlled: the minimum required budget bmin is known, and the actual budget is set to W bmin. and union of Nakhost et al. s small and large test suites, for No Mystery and Rovers).
Dataset Splits No The paper evaluates its state-space search method on planning benchmarks (problem instances) but does not describe training, validation, or test dataset splits in the typical machine learning context.
Hardware Specification Yes We use a cluster of Intel E5-2660 machines running at 2.20 GHz, with runtime (memory) limits of 30 minutes (4 GB).
Software Dependencies Yes Our implementation is in FD (Helmert 2006).
Experiment Setup Yes We break ties (i. e. order children in the DFS) using h FF (Hoffmann and Nebel 2001)). and We always start with C = C1 := {{p} | p F}... and We experimented with an input parameter α, stopping the learning when the number of counters reaches α times the number of counters for C1. We experimented with α {1, 16, 32, 64, 128, }. and As a standard satisficing planner, we run FD s greedy best-first dual-queue search with h FF and preferred operators, denoted FD-h FF .