GenCO: Generating Diverse Designs with Combinatorial Constraints
Authors: Aaron M Ferber, Arman Zharmagambetov, Taoan Huang, Bistra Dilkina, Yuandong Tian
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of our approach on a variety of generative combinatorial tasks, including game level generation, map creation for path planning, and photonic device design, consistently demonstrating its capability to yield diverse, high-quality solutions that verifiably adhere to user-specified combinatorial properties. We test Gen CO on three applications: Zelda game level design, Warcraft path planning map generation, and diverse inverse photonic device design. Settings are summarized in Table 2, and all settings involve combinatorial optimization, making the application of deep generative models nontrivial. In these experiments, Gen CO significantly outperforms the baselines, efficiently finding diverse, realistic solutions that obey combinatorial properties, paving the way for combining combinatorial optimization with deep generative models. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science, University of Southern California, Los Angeles, CA, USA 2Department of Computer Science, Cornell University, Ithaca, NY, USA 3AI at Meta (FAIR), Menlo Park, CA, USA. |
| Pseudocode | Yes | Algorithm 1 Gen CO; Algorithm 2 Gen CO Penalized generator training for GANs.; Algorithm 3 Gen CO in the constrained generator setting; Algorithm 4 Constrained Generator Training for VQVAE |
| Open Source Code | No | The paper mentions adapting a dataset from (Guyomarch, 2017) and provides a GitHub link for that third-party resource, but it does not contain a statement from the authors about releasing their own source code for the Gen CO framework described in the paper. |
| Open Datasets | Yes | We use the same 50 Zelda levels as (Zhang et al., 2020); We use the same dataset as in (Poganˇci c et al., 2020) with DCGAN architecture adapted from (Zhang et al., 2020)... The dataset used for training in the Shortest Path problem with k = 12 comprises 10,000 randomly generated terrain maps from the Warcraft II tileset (Poganˇci c et al., 2020) (adapted from (Guyomarch, 2017)). |
| Dataset Splits | No | The paper mentions training data and test sets (implicitly through evaluation), but does not explicitly describe a validation dataset split or strategy. |
| Hardware Specification | No | The paper does not specify the hardware details (e.g., CPU, GPU models, memory, or specific cloud resources) used for running the experiments. |
| Software Dependencies | No | The paper mentions various software components and solvers such as 'Gurobi (ILP)', 'Gurobi (LP)', 'SCIP', 'DCGAN architecture', 'WGAN', and 'VQVAE'. However, it does not provide specific version numbers for these software dependencies, which would be necessary for full reproducibility. |
| Experiment Setup | Yes | We employ Res Net as the mapping f( ) from the above Algorithm A.1, which transforms an image of a map into a 12 12 grid representation of a weighted directed graph: f : ℜ96 96 3 ℜ12 12. The first five layers of Res Net18 are pre-trained (75 epochs, Adam optimizer with lr=5e 4) using the dataset from (Poganˇci c et al., 2020)... We employ similar DCGAN architecture taken from (Zhang ets al., 2020) (see fig. 3 therein). Input to generator is 128 dimensional vector sampled from Gaussian noise centered around 0 and with a std of 1. Generator consists of five (256 128 64 32 16) blocks of transposed convolutional layers, each with 3 3 kernel sizes and batch normalization layers in between. Discriminator follows by mirroring the same architecture in reverse fashion. The entire structure is trained using the WGAN algorithm, as described in(Zhang et al., 2020). |