Decision-focused learning implementations

[gdalle]

Based on https://arxiv.org/abs/2307.13565

Features

	Julia	PyTorch	JAX
Analytical differentiation	DiffOpt, ImplicitDifferentiation		jaxopt
Regularization	InferOpt, DifferentiableFrankWolfe
Perturbation	InferOpt	PyEPO
Black box	InferOpt	PyEPO
Losses	InferOpt	PyEPO
Contrastive		PyEPO
Specific problems	InferOpt	PyEPO

Ease of use

	InferOpt	PyEPO
Math prog	JuMP	Pyomo, Gurobipy, COPT
Arbitrary solvers	Yes	Yes
Solution caching	No	Yes
Deep learning	Zygote	PyTorch

Why Julia?

Find an example where it is essential to stay in the same language between optimizer and neural network:

• JuMP is better than Pyomo
• GPU-compatible combinatorial solver?

Sell modularity first with #110