Automatic Loss Function Search for Predict-Then-Optimize Problems with Strong Ranking Property
Authors: Boshi Wang, Jialin Yi, Hang Dong, Bo Qiao, Chuan Luo, Qingwei Lin
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed APOC method is applied to three predict-then-optimize problems with real-world datasets to validate its performance. Those three PTO problems contain optimization problems with different representative combinatorial structures, i.e., ranking, knapsack, and shortest path problem. Empirically, our experimental results demonstrate that significant performance gains can be achieved against some previously proposed algorithms by finding a better loss for the prediction phase. |
| Researcher Affiliation | Collaboration | Boshi Wang1 , Jialin Yi2 , Hang Dong3, Bo Qiao3, Chuan Luo3 , Qingwei Lin3 1The Ohio State University, United States 2London School of Economics and Political Science, United Kingdom 3Microsoft Research, China |
| Pseudocode | Yes | Algorithm 1: Automatic Prediction and Optimization Connector (APOC) |
| Open Source Code | Yes | Source code We have made the source code for the proposed APOC available in the following repository: https://github.com/Microsoft/Auto Pred Opt Connector . |
| Open Datasets | Yes | Energy Cost-Aware Scheduling This dataset contains two years of historical energy price data from the day-head market of the Irish Single Electricity Market Operator (SEM-O)...(Simonis et al.) Knapsack Formulation of Portfolio Optimization We use the historical daily price and volume data from 2004 to 2017 from partition SP500 of the Quandl WIKI dataset (QUANDL, 2020). Shortest Path We use the processed data from Mandi & Guns (2020) where the goal is to find the shortest path of the given source-destination pairs on a Twitter ego network (Mc Auley & Leskovec, 2012). |
| Dataset Splits | Yes | For each experiment, we use the same prediction model architecture and train/validation/test splits across different methods for fair comparisons. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for the experiments. It only describes the prediction model architecture and optimizers used. |
| Software Dependencies | No | The paper mentions 'Adam optimizer', 'Gurobi optimizer', 'CVXPY', and 'QPTH' but does not provide specific version numbers for any of these software dependencies, which are necessary for full reproducibility. |
| Experiment Setup | Yes | For Energy Cost-Aware Scheduling, the prediction model is instantiated to be a 4-layer neural network (hidden layer sizes of 500-300 with Re LU as activation function) with learning rate 5 10 5 and γ = 5.0. For Shortest Path, the prediction model is a 4-layer network (hidden layer sizes of 100-100 with Re LU as activation function) with learning rate 5 10 4 and γ = 3.0. For Portfolio Optimization, we use a 5-layer neural network (hidden layer sizes of 100-100-100 with Re LU as activation function) with learning rate 5 10 5, γ = 3.0. |