Learning Generalized Relational Heuristic Networks for Model-Agnostic Planning

Authors: Rushang Karia, Siddharth Srivastava8064-8073

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

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Empirical Evaluation We implemented GHN learning and tested the learned GHNs with various search algorithms (referred to as GHN/algorithm in the remainder of this section). Our implementation1 uses Pyperplan, a popular Python-based platform for implementing and evaluating planning algorithms (Alkhazraji et al. 2020). Summary of observations Our results indicate that even though they do not use action models, (a) GHNs are competitive when compared against hand-coded HGFs, (b) in the absence of externally generated training data, leapfrogging is an effective self-training technique, and (c) GHNs successfully transfer to problems with more objects than those in the training data. We discuss the configuration and methods used for evaluating these hypotheses below.
Researcher Affiliation Academia Rushang Karia, Siddharth Srivastava School of Computing, Informatics and Decision Systems Engineering Arizona State University, Tempe, AZ 85281, USA {Rushang.Karia, siddharths}@asu.edu
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Code available at https://github.com/AAIR-lab/GHN
Open Datasets Yes We generated problems randomly from problem generators used by organizers of the IPC (Fawcett et al. 2011).
Dataset Splits No The paper discusses training and test sets, but does not explicitly mention a separate validation set or provide details on its split (e.g., percentages or counts).
Hardware Specification Yes We ran our experiments on Agave compute instances provided by Arizona State University. Each compute node is configured with an Intel Xeon E5-2680 v4 CPU composed of 28 cores and 128GB of RAM.
Software Dependencies No The paper mentions software like Pyperplan and Keras, but does not provide specific version numbers for them.
Experiment Setup Yes Our optimization algorithm was the Keras (Chollet et al. 2015) implementation of RMSProp (Hinton, Srivastava, and Swersky 2012) configured with a learning rate, η = 0.001 and ˆϵ = 1e 3. GHNs were trained for 100 epochs using a batch size of 32. categorical cross entropy, binary cross entropy and mean absolute error were the loss minimization functions for the NNA, NN1,...,Amax, and NNlen layers respectively.