Subgoal Search For Complex Reasoning Tasks

Authors: Konrad Czechowski, Tomasz Odrzygóźdź, Marek Zbysiński, Michał Zawalski, Krzysztof Olejnik, Yuhuai Wu, Łukasz Kuciński, Piotr Miłoś

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically demonstrate the efficiency of MCTS-k Sub S and BF-k Sub S. In particular, we show that they vastly outperform their standard ( non-subgoal ) counterparts. As a testing ground, we consider three challenging domains: Sokoban, Rubik s Cube, and INT. All of them require non-trivial reasoning.
Researcher Affiliation Collaboration Konrad Czechowski University of Warsaw k.czechowski@mimuw.edu.pl Tomasz Odrzygó zd z University of Warsaw tomaszo@impan.pl Marek Zbysi nski University of Warsaw m.zbysinski@ students.mimuw.edu.pl Michał Zawalski University of Warsaw m.zawalski@uw.edu.pl Krzysztof Olejnik University of Warsaw k.olejnik3@ student.uw.edu.pl Yuhuai Wu University of Toronto, Vector Institute ywu@cs.toronto.edu Łukasz Kuci nski Polish Academy of Sciences lkucinski@impan.pl Piotr Miło s Polish Academy of Sciences, University of Oxford, deepsense.ai pmilos@impan.pl
Pseudocode Yes Algorithm 1 Best-First Subgoal Search (BF-k Sub S) [...] Algorithm 2 Low-level conditional policy [...] Algorithm 3 Subgoal generator
Open Source Code Yes We provide the code of our method and experiment settings at https://github.com/ subgoal-search/subgoal-search, and a dedicated website https://sites.google.com/ view/subgoal-search.
Open Datasets Yes 2The dataset for INT or Sokoban can be easily generated or are publicly available. For the Rubik s Cube, we use random data or simple heuristic (random data are often sufficient for robotic tasks and navigation.) ... INT [55]
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning.
Hardware Specification Yes The results were obtained on a server with an Intel Xeon E5-2630 v4 CPU and eight NVIDIA Tesla V100 GPUs.
Software Dependencies No The paper mentions software components like 'transformer architecture' and 'convolutional network' but does not specify their version numbers or the versions of any other software dependencies.
Experiment Setup Yes Table 1: BF-k Sub S hyperparameters. [...] In Table 1, we provide the values of the hyperparameters used in all experiments.