A groundbreaking methodological advance is embedding mathematical programming problems as layers in neural networks. Frameworks like cvxpylayers allow backpropagation through convex optimization problems, enabling end-to-end learning of model parameters. Hot applications include:
Methodological shift: The modeller now co-designs the predictive model and the prescriptive model, blurring the line between data science and operations research.
As mathematical programming models affect hiring, lending, policing, and healthcare, modellers must now justify decisions — not just optimize. This has sparked a methodological hot spot: Explainable Optimization. modelling in mathematical programming methodol hot
In mathematical programming, sparsity (ensuring a document only belongs to a few topics) is handled via norm regularization.
The Optimization Program: $$ \min_W, H | X - WH |_F^2 + \lambda_1 |W|_1 + \lambda_2 |H|_1 $$ JuMP’s decomposition libraries).
This is a Penalty Method. The $L_1$ norm ($|.|_1$) induces sparsity. This formulation is mathematically equivalent to the automatic relevance determination in Bayesian models but is solved using gradient descent or proximal gradient methods (e.g., ISTA/FISTA algorithms).
Instead of assuming distributions, modellers: Example: In energy systems
Example: In energy systems, historical renewable generation data shapes an ambiguity set, ensuring solutions are feasible for likely scenarios without over-conservatism.
Instead of modelling the whole system, modellers now design problems amenable to:
Hot twist: These are no longer just algorithms but are built into modelling languages (e.g., Pyomo’s GDP, JuMP’s decomposition libraries).