Locally Convex Global Loss Network for Decision-Focused Learning
Authors: Haeun Jeon, Hyunglip Bae, Minsu Park, Chanyeong Kim, Woo Chang Kim
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We confirm the effectiveness and flexibility of LCGLN by evaluating our proposed model with three stochastic decision-making problems. In the experiment section, we demonstrate the capability of LCGLN on three stochastic decision-making problems, namely inventory stock problem, budget allocation problem, and portfolio optimization problem. We show that LCGLN learns task loss well with a single surrogate loss. Table 1: The table contains normalized test regret Rtest/Rworst with standard error mean (SEM) tested on three stochastic optimization problems. |
| Researcher Affiliation | Academia | KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1: Training Predictive Model with LCGLN Lψ |
| Open Source Code | No | The paper states: "Our experiments were built on top of the public codes from previous research (Donti, Amos, and Kolter 2017; Shah et al. 2022)." This refers to code utilized from other sources, not code released by the authors for the methodology described in this paper. There is no explicit statement about releasing their own source code or a link to a repository. |
| Open Datasets | No | The paper discusses experiments on "inventory stock problem", "budget allocation problem", and "portfolio optimization problem" and mentions using "historical daily stock return data". However, it does not provide specific dataset names (e.g., a well-known benchmark like MNIST or CIFAR-10), links, DOIs, repository names, or formal citations for public access to the data used in their experiments. It references previous research for problem descriptions but not for direct dataset access. |
| Dataset Splits | No | The paper mentions "normalized test regret Rtest/Rworst" and that "For the surrogate loss models, we used 32 samples to learn the loss." and "We run 10 experiments for each setting to ensure statistical significance." While it implies the use of test sets and refers to sample sizes for training the loss model, it does not provide specific details on how the main dataset was split into training, validation, or test sets (e.g., percentages, absolute counts, random seed, or specific predefined splits with citations). |
| Hardware Specification | Yes | The experiments were performed on a Ryzen 7 5800X CPU and an RTX 3060 GPU with 64GB of RAM. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, or specific solvers/libraries and their versions). It only mentions that experiments were built "on top of the public codes from previous research" without detailing the software environment for their own implementation. |
| Experiment Setup | Yes | We used a predictive model Mθ as one hidden layer MLP and a learning rate of 0.001. 500 hidden nodes were used for the portfolio optimization and 10 for the other problems. We mirrored the predictive model exactly for the sampling model Mξ. For LCGLN, we employed a single hidden layer PICNN with two nodes per layer, a learning rate of 0.001, and a softplus activation function. ... For each global surrogate loss model employing model-based sampling, we selected the learning rate that demonstrated the best performance from {0.01, 0.05, 0.1, 0.5, 1}. We run 10 experiments for each setting to ensure statistical significance. |