Task Generalization in Decision-Focused Learning
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: task generalization, decision-focused learning, operations research, constrained optimization, combinatorial optimization, linear programming
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Real-world optimization problems often contain uncertain parameters that must be predicted prior to solving. For example, a delivery company must make its routing decisions when the traffic conditions, and thus the road traversal times, are uncertain. The models used to predict these uncertain quantities are commonly trained in a way that is agnostic of the optimization problem and that focuses solely on predictive accuracy. However, such a prediction-focused training procedure generally does not minimize the downstream task loss of interest (e.g., the suboptimality of the roads that are selected based on the predictions). This has led to the development of decision-focused learning (DFL) methods, which specifically train the predictive model to make predictions that lead to good decisions on the considered optimization task. However, as we show in this paper, such models often generalize poorly to altered optimization tasks. For example, in the context of a routing problem, their performance may deteriorate when the destination node changes. To improve on this, we first explore how the model can be trained to generalize implicitly, by simply training it on different tasks sampled at training time. We then propose a more sophisticated approach by adding the use of explicit task representations, to enable the model to adapt its predictions better to different tasks. To this end, we represent the optimization problems as bipartite variable-constraint graphs, and train graph neural networks (GNNs) to produce informative node embeddings that are then given to the predictive model. In our experiments, we start by showing that the state of the art in DFL tends to overfit to the specific task it is trained on, and generalizes poorly to changing tasks. We then show that both of our proposed strategies significantly improve on this, with the explicit task representations generally providing an additional improvement over the implicit strategy.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8165
Loading