Risk Bounds and Calibration for a Smart Predict-then-Optimize Method

Authors: Heyuan Liu, Paul Grigas

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments to empirically demonstrate the strength of the SPO+ surrogate, as compared to standard ℓ1 and squared ℓ2 prediction error losses, on portfolio allocation and cost-sensitive multi-class classification problems.
Researcher Affiliation Academia Heyuan Liu University of California, Berkeley Berkeley, CA 94720 heyuan_liu@berkeley.edu Paul Grigas University of California, Berkeley Berkeley, CA 94720 pgrigas@berkeley.edu
Pseudocode No The paper describes mathematical formulations and theoretical results but does not include any pseudocode or algorithm blocks.
Open Source Code Yes 3. If you ran experiments... (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Please refer Appendix D.
Open Datasets No We present computational results of synthetic dataset experiments wherein we empirically examine the performance of the SPO+ loss function... In our simulations, the relationship between the true cost vector c and its auxiliary feature vector x is given by c = φdeg(Bx) ϵ, where φdeg is a polynomial kernel mapping of degree deg, B is a fixed weight matrix, and ϵ is a multiplicative noise term. The features are generated from a standard multivariate normal distribution, we consider d = 50 assets, and further details of the synthetic data generation process are provided in Appendix D.
Dataset Splits No We set the size of the test set to 10000. ... Figure 2 shows a detailed comparison between these alternative SPO+ surrogates as we vary the training set size. (No explicit mention of a validation set split.)
Hardware Specification Yes All experiments were run on a single machine with a 3.0 GHz Intel Xeon W-2145 CPU and 128 GB of RAM.
Software Dependencies No We use Pytorch [Paszke et al., 2019] for all of our experiments. (No version number provided for PyTorch, nor for Adam optimizer or other potential dependencies).
Experiment Setup Yes We focus on two classes of prediction models: (i) linear models, and (ii) two-layer neural networks with 256 neurons in the hidden layer. ... For all loss functions, we use the Adam method of Kingma and Ba [2015] to train the parameters of the prediction models. ... The learning rate is initialized at 1e-3 and decays by 0.5 every 20 epochs. For both linear models and neural networks, we train for 100 epochs.