On the Robustness of Decision-Focused Learning (2311.16487v4)
Abstract: Decision-Focused Learning (DFL) is an emerging learning paradigm that tackles the task of training a ML model to predict missing parameters of an incomplete optimization problem, where the missing parameters are predicted. DFL trains an ML model in an end-to-end system, by integrating the prediction and optimization tasks, providing better alignment of the training and testing objectives. DFL has shown a lot of promise and holds the capacity to revolutionize decision-making in many real-world applications. However, very little is known about the performance of these models under adversarial attacks. We adopt ten unique DFL methods and benchmark their performance under two distinctly focused attacks adapted towards the Predict-then-Optimize problem setting. Our study proposes the hypothesis that the robustness of a model is highly correlated with its ability to find predictions that lead to optimal decisions without deviating from the ground-truth label. Furthermore, we provide insight into how to target the models that violate this condition and show how these models respond differently depending on the achieved optimality at the end of their training cycles.
- Differentiable Convex Optimization Layers. CoRR, abs/1910.12430.
- Learning with Fenchel-Young Losses. arXiv:1901.02324.
- The Perils of Learning Before Optimizing. arXiv:2106.10349.
- An Analysis of Adversarial Attacks and Defenses on Autonomous Driving Models. arXiv:2002.02175.
- Smart ”Predict, then Optimize”. arXiv:1710.08005.
- Adversarial Regression for Detecting Attacks in Cyber-Physical Systems. arXiv:1804.11022.
- Explaining and Harnessing Adversarial Examples. arXiv:1412.6572.
- An Adversarial Attacker for Neural Networks in Regression Problems. In IJCAI Workshop on Artificial Intelligence Safety (AI Safety). Montreal/Virtual, Canada.
- Gurobi Optimization, LLC. 2023. Gurobi Optimizer Reference Manual.
- guyomarch. 2017. War2Edit. https://github.com/war2/war2edit.
- Properties of Energy-Price Forecasts for Scheduling. In Milano, M., ed., Principles and Practice of Constraint Programming, 957–972. Berlin, Heidelberg: Springer Berlin Heidelberg. ISBN 978-3-642-33558-7.
- Characterizing and Improving the Robustness of Predict-Then-Optimize Frameworks.
- Adam: A Method for Stochastic Optimization. arXiv:1412.6980.
- An Exploration of Poisoning Attacks on Data-Based Decision Making. In Fang, F.; Xu, H.; and Hayel, Y., eds., Decision and Game Theory for Security, 231–252. Cham: Springer International Publishing. ISBN 978-3-031-26369-9.
- Adversarial examples in the physical world. arXiv:1607.02533.
- Decision-Focused Learning: Through the Lens of Learning to Rank. arXiv:2112.03609.
- Smart Predict-and-Optimize for Hard Combinatorial Optimization Problems. CoRR, abs/1911.10092.
- Interior Point Solving for LP-based prediction+optimisation. CoRR, abs/2010.13943.
- Decision-Focused Learning: Foundations, State of the Art, Benchmark and Future Opportunities. arXiv:2307.13565.
- White-Box Target Attack for EEG-Based BCI Regression Problems. arXiv:1911.04606.
- Contrastive Losses and Solution Caching for Predict-and-Optimize. arXiv:2011.05354.
- Adversarial Attacks, Regression, and Numerical Stability Regularization. arXiv:1812.02885.
- Implicit MLE: Backpropagating Through Discrete Exponential Family Distributions. CoRR, abs/2106.01798.
- Differentiation of Blackbox Combinatorial Solvers. CoRR, abs/1912.02175.
- PyTorch: An Imperative Style, High-Performance Deep Learning Library. CoRR, abs/1912.01703.
- OR-Tools.
- Adversarial Regression with Multiple Learners. arXiv:1806.02256.
- Melding the Data-Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization. arXiv:1809.05504.