Automated decision making via optimization and machine learning

本研究项目将在逆优化、鲁棒优化和机器学习的交叉领域开发一个新的自动化决策平台。许多决策问题结合了优化和机器学习。例如，生成决策的优化模型可能使用通过机器学习预测的辅助参数。为了训练机器学习模型，我们需要参数的数据。然而，直接获得所需解决方案的数据往往更容易，也更相关。因此，我的“发现”项目的目标是为这种“预测-然后重建”范式开发新的计算工具。特别是，我将使用机器学习来预测最优解的理想特性。然后，我将开发新的逆优化方法来学习一个鲁棒优化模型的参数，该模型可以生成具有期望属性的解。从某种意义上说，框架将是自动化的，人们所需要的只是与解决方案质量相关的基本协变量；有了这些数据，机器学习，然后逆优化，然后鲁棒优化可以在没有人为干预的情况下实现。

This research program will develop a new automated decision making platform at the intersection of inverse optimization, robust optimization and machine learning. Many decision making problems combine optimization and machine learning. For example, an optimization model that generates decisions may use auxiliary parameters predicted via machine learning. To train the machine learning model, one needs data on the parameters. However, it is often easier and more relevant to obtain data on the desired solutions directly. Thus, the goal of my Discovery program is to develop new computational tools for this predict-then-reconstruct paradigm. In particular, I will use machine learning to predict desirable characteristics of an optimal solution. Then, I will develop new inverse optimization methods to learn parameters of a robust optimization model that can generate a solution with the desired properties. The framework will be automated in the sense all that one needs are basic covariates that relate to solution quality; with this data, machine learning, then inverse optimization, then robust optimization can be implemented without human intervention. The main application in this research program is radiation therapy treatment planning. Radiation therapy is one of the primary ways to treat cancer. The current process to design a treatment is inefficient, relying on manual trial-and-error effort. The challenge stems from the difficulty in determining appropriate parameters in the planning software that will produce an acceptable treatment. Using my predict-then-reconstruct approach, I will demonstrate automated treatment plan generation. A deep learning model trained on historical treatments will predict a clinically acceptable dose distribution for a patient, given only imaging data. Then, an inverse optimization model will learn parameters of a robust optimization model that will create a deliverable treatment plan with clinically desirable characteristics. My automated planning approach will simultaneously improve efficiency and process standardization without sacrificing treatment personalization. While the example application is radiation therapy, the general methodology will be broadly applicable. Optimization and machine learning are fast-growing fields that are in short supply of graduates. My trainees will develop highly employable skills that will be valued in data-driven industries like healthcare, transportation, manufacturing, finance, supply chain management, energy and defense. The application of my research to radiation therapy also has major economic potential. Radiation therapy is used to treat half of all cancer patients, but there is a looming shortage of skilled personnel to design and deliver treatments. My research helps close the demand-supply gap and allows radiation therapy to be delivered at scale, particularly important in developing countries. Thus, my research represents a made-in-Canada innovation with global impact.