VAIN: Attentional Multi-agent Predictive Modeling

Authors: Yedid Hoshen

NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our method is evaluated on tasks from challenging multi-agent prediction domains: chess and soccer, and outperforms competing multi-agent approaches.
Researcher Affiliation Industry Yedid Hoshen Facebook AI Research, NYC yedidh@fb.com
Pseudocode No The paper provides a detailed schematic and descriptive steps for the VAIN architecture in Figure 1, but it is presented as a textual explanation rather than a formal, structured pseudocode block or algorithm labeled as such. For example, it says: 'Figure 1: A schematic of a single-hop VAIN: i) The agent features Fi are embedded by singleton encoder Es()...'
Open Source Code No The paper states 'All methods were implemented in Py Torch [34]', and reference [34] points to the PyTorch framework's GitHub repository. However, it does not explicitly state that the authors' own implementation code for VAIN is open-source or provide a link to their specific code repository.
Open Datasets Yes For training and evaluation of this task we downloaded 10k games from the FICS Games Dataset, an on-line repository of chess games.
Dataset Splits No The paper specifies training and evaluation sets (e.g., '9k randomly sampled games were used for training, and the remaining 1k games for evaluation'), and discusses test/train splits for soccer, but it does not explicitly mention a separate validation set used for hyperparameter tuning or model selection.
Hardware Specification Yes The chess prediction training for the MPP took several hours on a M40 GPU, other tasks had shorter training times due to smaller datasets.
Software Dependencies No The paper mentions 'All methods were implemented in Py Torch [34]' and refers to 'ADAM [35]' as the optimizer. While PyTorch is a software dependency, a specific version number for it is not provided, nor are version numbers for any other key software components or libraries.
Experiment Setup Yes The encoding and decoding functions Ec(), Es() and D() were implemented by fullyconnected neural networks with two layers, each of 256 hidden units and with Re LU activations. The encoder outputs had 128 units.