Value Propagation Networks
Authors: Nantas Nardelli, Gabriel Synnaeve, Zeming Lin, Pushmeet Kohli, Philip H. S. Torr, Nicolas Usunier
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate on static and dynamic configurations of Maze Base grid-worlds, with randomly generated environments of several different sizes, and on a Star Craft navigation scenario, with more complex dynamics, and pixels as input. 4 Experiments |
| Researcher Affiliation | Collaboration | Nantas Nardelli University of Oxford nantas@robots.ox.ac.uk Gabriel Synnaeve Facebook AI Research gab@fb.com Zeming Lin Facebook AI Research zlin@fb.com Pushmeet Kohli Google Deep Mind pushmeet@google.com Philip H. S. Torr University of Oxford philip.torr@eng.ox.ac.uk Nicolas Usunier Facebook AI Research usunier@fb.com |
| Pseudocode | No | The paper describes mathematical equations and formulas for the models and training updates, but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | All the models and agent code was implemented in Py Torch (Paszke et al., 2017), and will be made available upon acceptance together with the environments. |
| Open Datasets | Yes | We use Maze Base (Sukhbaatar et al., 2015) to generate the configurations of our world and the agent interface for both training and testing phases. Additionally we also evaluate our trained agents on maps uniformly sampled from the 16 16 dataset originally used by Tamar et al. (2016) |
| Dataset Splits | No | The paper describes procedures for generating environments for training and testing (e.g., using Maze Base, generating maps of different sizes), and mentions training on 8x8 maps and evaluating on 32x32 maps, but it does not specify explicit percentages or sample counts for training, validation, and test splits of a predefined dataset. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running the experiments. |
| Software Dependencies | No | All the models and agent code was implemented in Py Torch (Paszke et al., 2017). The paper mentions PyTorch and Torch Craft but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | The capping constant (C = 10 in our experiments)... In all our experiments we used RMSProp rather than plain SGD, with relative weights λ = η = 100.0η . We also used a learning rate to 0.001 and mini-batch size of 128, with learning updates set at a frequency of 32 steps. |