ADMOL: A Generic Framework for Active Decision Making within Online Learning

This proposal aims to develop a novel and versatile online learning framework called ADMOL, which integrates active decision-making within an online learning model to solve problem instances at each time step in a cost-efficient manner.

ADMOL is designed for real-world scenarios where both utility and information acquisition cost are crucial considerations. Its efficiency is derived from two key aspects.

Firstly, it achieves faster learning by computing an appropriate decision policy using minimal streaming data or time steps.

Secondly, it reduces the cost incurred for decision-making on each problem instance, resulting in more cost-effective decisions. An example is the online medical diagnosis of patients arriving sequentially.

Given that medical tests can be expensive, the objective is to accurately determine the disease for each patient with minimal tests, serving as the decision-making process.ADMOL is organized into four work packages. Work package 1 focuses on active decision-making at each round or time step.

Work package 2 entails setting up the entire framework with a focus on online learning and exploring unknown or partially known parameters.

Work package 3 studies the dependencies among tests, and finally, work package 4 establishes strong theoretical guarantees on the framework´s performance through regret analysis.