This project is carried out as part of the Optimization Methods course at MIT.
We develop a framework to be leveraged in the scenario where the inputs to an optimization problem are fueled by a machine learning model that is undergoing active learning. In such context, it is possible to leverage duality information to select a batch of to-be-labelled datapoints with awareness of the impact on the optimization problem, which is considered the end goal. We prove the added value of such method and we validate the approach empirically by testing it on a specific task. As a holistic approach in the scenario, we compare the hybrid use of three active learning methodologies, as well as the combination of only the feature based and the optimization based approaches. Finally, we propose extensions and possible future work to generalize the framework.
Check out the project report with technical details and results.