Differentiation of Blackbox Combinatorial Solvers
Authors: Marin Vlastelica Pogančić, Anselm Paulus, Vit Musil, Georg Martius, Michal Rolinek
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we experimentally validate a proof of concept: that architectures containing exact blackbox solvers (with backward pass provided by Algo. 1) can be trained by standard methods. To that end, we solve three synthetic tasks as listed in Tab. 1. These tasks are designed to mimic practical examples from Section 2 and solving them anticipates a two-stage process: 1) extract suitable features from raw input, 2) solve a combinatorial problem over the features. |
| Researcher Affiliation | Academia | Marin Vlastelica1 , Anselm Paulus1 , V ıt Musil2, Georg Martius1, Michal Rol ınek1 1 Max-Planck-Institute for Intelligent Systems, T ubingen, Germany 2 Universit a degli Studi di Firenze, Italy |
| Pseudocode | Yes | Algorithm 1 Forward and Backward Pass |
| Open Source Code | Yes | The code is available at https://github.com/martius-lab/blackbox-backprop. |
| Open Datasets | Yes | The training dataset for problem SP(k) consists of 10000 examples of randomly generated images of terrain maps from the Warcraft II tileset (Guyomarch, 2017). The dataset consists of randomly generated grids of MNIST digits that are sampled from a subset of 1000 digits of the full MNIST dataset. |
| Dataset Splits | No | The paper reports Train % and Test % in tables but does not provide specific details on the dataset splits (e.g., explicit percentages, absolute counts for each split, or a clear splitting methodology for reproducibility). |
| Hardware Specification | No | All models train in under two hours on a single machine with 1 GPU and no more than 24 utilized CPU cores. This is not specific enough (e.g., GPU model, CPU type/speed). |
| Software Dependencies | No | The paper mentions software like Adam optimizer, ResNet18, Gurobi MIP solver, Blossom V, and Dijkstra's algorithm, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | Optimization was carried out via Adam optimizer (Kingma & Ba, 2014) with scheduled learning rate drops dividing the learning rate by 10 at epochs 30 and 40. Table 5: Experimental setup for Warcraft Shortest Path. k Optimizer(LR) Architecture Epochs Batch Size λ 12, 18, 24, 30 Adam(5 10 4) subset of Res Net18 50 70 20 |