DualSDDP implements the Dual SDDP algorithm for multistage stochastic linear problems with stagewise independent uncertainty, allowing for polyhedral risk measures such as Expectation-AV@R. The theory is developped in https://arxiv.org/abs/2107.10930.

It also implements, for comparison:

• A primal SDDP algorithm
  • with outer approximations of the value function, and
  • stochastic sampling in the scenario tree.
• a deterministic DDP algorithm
  • with inner and outer approximations for the value function, and
  • worst-gap selection in the scenario tree.

• Create a module M, of type MSLBO, containing the problem data;
• Determine the number of stages of the problem;
• Choose the E-AV@R parameters and create builder functions (primal or dual) using mk_primal_avar and mk_copersp_avar (respectively);
• Set the LP solver for subproblems;
• Set the initial value;
• Set the number of iterations of the algorithm;
• call primalsolve, dualsolve or problem_child_solve.

The examples directory contains both "one shot" examples (in subdirectories), as well as Config.jl batches and configs.

The first time you use the examples directory, you'll have to instantiate it. The fastest way of doing so is using Pkg commands (in the examples directory):

pkg> activate .
pkg> dev ..

This will use the Project.toml we provide, then the current checkout for the DualSDDP package, and finally resolve all packages.

In so doing, it will also require that the Gurobi package is installed, which was used for some of the examples. If that is not desired, one can remove it from the Packages.toml file, or simply issue pkg> rm Gurobi.