APPLICATIONDRIVENLEARNING. JL: A HIGH-PERFORMANCE JULIA PACKAGE FOR TRAINING PREDICTIVE MODELS IN THE CONTEXT OF DECISION-MAKING

The Application-Driven Learning framework proposes an approach that integrates the training of predictive models with decision-making processes, optimizing the models specifically for the context of the application. In this article, we introduce ApplicationDrivenLearning.jl, an open-source library that, from the Julia language's optimization and machine learning ecosystems, enables flexible representation and efficient training of these models. The interface allows the user to accurately model decision-making processes, with constraints and objective functions of various natures, as well as to use predictive models of varying sizes and structures. The training can be carried out from an exact formulation of bilevel optimization or a heuristic approach using optimization methods with Nelder-Mead or gradient descent, a secondary contribution of this work. The suite dramatically lowers the barriers to access to this new technology by combining ease of use and high-performance solution methods.