NfgTransformer: Equivariant Representation Learning for Normal-form Games
Authors: Siqi Liu, Luke Marris, Georgios Piliouras, Ian Gemp, Nicolas Heess
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The goal of our empirical studies is three-fold. First, we compare the Nfg Transformer to baseline architectures to demonstrate improved performance in diverse downstream tasks on synthetic and standard games. Second, we vary the model hyper-parameters and observe how they affect performance. We show in particular that some tasks require larger action embedding sizes while others benefit from more rounds of iterative refinement. Lastly, we study how the model learned to solve certain tasks by interpreting the sequence of learned attention masks in a controlled setting. Our results reveal that the solution found by the model reflects elements of intuitive solutions to NEsolving in games. |
| Researcher Affiliation | Collaboration | Google Deep Mind, }University College London {liusiqi,marris,gpil,imgemp,heess}@google.com |
| Pseudocode | No | The paper describes the model architecture and its components using figures and textual descriptions, but it does not include any explicitly labeled pseudocode block or algorithm. |
| Open Source Code | Yes | 1The model is open-sourced at https://github.com/google-deepmind/nfg_transformer. |
| Open Datasets | No | The paper mentions generating synthetic games by following 'Marris et al. (2022)' and using 'DISC games (Balduzzi et al., 2019)', which are citations to papers describing game generation or types. However, it does not provide concrete access information (e.g., a specific link, DOI, or repository) for the exact datasets used in their experiments. |
| Dataset Splits | No | The paper discusses training and evaluating models on synthetic games and DISC games, but it does not provide specific details on how the datasets were split into training, validation, or test sets (e.g., percentages, sample counts, or explicit standard splits). |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware used for running experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper describes the model architecture and training process but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For Nfg Transformer variants (Ours), we annotate each variant with corresponding hyper-parameters (K, A and D as shown in Figure 1). We optimised an Nfg Transformer network as in Section 5.1, but focused on 10 10 games and removed the action-toaction self-attention (i.e. A = 0) for a more concise visualisation. Instead of attending to multiple pieces of information in parallel, each attention layer is implemented by a single attention head (H = 1). |