Learning to Initiate and Reason in Event-Driven Cascading Processes
Authors: Yuval Atzmon, Eli Meirom, Shie Mannor, Gal Chechik
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4. Experiments We compared our approach to state-of-the-art baselines, including human performance. Then, we follow with an ablation study to examine the contributions of different components in our approach. |
| Researcher Affiliation | Collaboration | 1NVIDIA Research, Israel 2Technion, Israel institute of technology 3Bar Ilan University, Israel. |
| Pseudocode | No | The paper does not contain any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or a link to its own open-source code for the methodology described. |
| Open Datasets | No | We generated a dataset with 46K scenes (we limited generation time to 80 hours), each includes 4-6 moving balls, 0-2 pins, and 4 walls and up to 5 semantic instructions ( 4.25 on average). The paper does not provide concrete access information for this dataset. |
| Dataset Splits | Yes | The data is split by unique scenes, into 470 unseen scenes for test, 69 scenes for selecting hyper-parameters (val. set), and the rest are used for training. |
| Hardware Specification | No | The paper mentions using a 'GPU' and maximizing 'GPU memory usage' but does not provide specific hardware details like GPU model, CPU, or exact memory specifications. |
| Software Dependencies | No | The paper mentions using PyTorch, NetworkX, and Isaac Gym, but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We train the model and baselines for 15 epochs. Batch size was set to 8192 to maximize the GPU memory usage. We use the Py Torch default learning rate for Adam (Kingma & Ba, 2015) (0.001). For inference, we set Nobserved to 9, the maximal tree depth to 30, we sample 106 initial states and expand 80 nodes per episode which takes 13 seconds. The GNN uses 5 layers, with a hidden state dimension of 128. |