DecisionProgramming.jl
is a Julia package for solving multi-stage decision problems under uncertainty, modeled using influence diagrams. Internally, it relies on mathematical optimization. Decision models can be embedded within other optimization models. We designed the package as JuMP extension. We have also developed a Python interface, which is available here.
The Decision Programming framework is described in this publication. If you found the framework useful in your work, we kindly ask you to cite the following publication (pdf):
@article{Salo_et_al-2022,
title = {Decision programming for mixed-integer multi-stage optimization under uncertainty},
journal = {European Journal of Operational Research},
volume = {299},
number = {2},
pages = {550-565},
year = {2022},
issn = {0377-2217},
doi = {https://doi.org/10.1016/j.ejor.2021.12.013},
url = {https://www.sciencedirect.com/science/article/pii/S0377221721010201},
author = {Ahti Salo and Juho Andelmin and Fabricio Oliveira},
keywords = {Decision analysis, Influence diagrams, Decision trees, Contingent portfolio programming, Stochastic programming}
}
If you use the rooted junction tree models, we kindly ask you to cite the following publication (pdf):
@article{Parmentier_et_al-2020,
title = {Integer programming on the junction tree polytope for influence diagrams},
journal = {INFORMS Journal on Optimization},
volume = {2},
number = {3},
pages = {209--228},
year = {2020},
issn = {0377-2217},
doi = {https://doi.org/10.1287/ijoo.2019.0036},
url = {https://pubsonline.informs.org/doi/epdf/10.1287/ijoo.2019.0036},
author = {Parmentier, Axel and Cohen, Victor and Lecl{\`e}re, Vincent and Obozinski, Guillaume and Salmon, Joseph},
keywords = {Influence diagrams, Partially observed Markov decision processes, probabilistic graphical models, Linear programming}
}
We can create an influence diagram as follows:
using DecisionProgramming
diagram = InfluenceDiagram()
add_node!(diagram, DecisionNode("A", [], ["a", "b"]))
add_node!(diagram, ChanceNode("B", ["A"], ["x", "y"]))
add_node!(diagram, ChanceNode("C", ["A"], ["v", "w"]))
add_node!(diagram, DecisionNode("D", ["B", "C"], ["k", "l"]))
add_node!(diagram, ValueNode("V", ["D"]))
generate_arcs!(diagram)
add_probabilities!(diagram, "B", [0.4 0.6; 0.6 0.4])
add_probabilities!(diagram, "C", [0.7 0.3; 0.3 0.7])
add_utilities!(diagram, "V", [1.5, 1.7])
Using the influence diagram, we generate the model as follows:
using JuMP
model, z, variables = generate_model(diagram, model_type="RJT")
We can optimize the model using MILP solver.
using HiGHS
optimizer = optimizer_with_attributes(
() -> HiGHS.Optimizer()
)
set_optimizer(model, optimizer)
optimize!(model)
Finally, we extract the decision strategy from the decision variables.
Z = DecisionStrategy(z)
See the documentation for more detailed examples.
DecisionProgramming.jl
is registered. You can add it using the command:
pkg> add DecisionProgramming
To run examples and develop and solve decision models, you have to install JuMP and a solver capable of solving mixed-integer linear programs (MILP). JuMP documentation contains a list of available solvers.
pkg> add JuMP
We recommend using the HiGHS solver, which is an efficient open-source solver.
pkg> add HiGHS
Now you are ready to use decision programming.
Using the package manager, add DecisionProgramming.jl
package for development using the command:
pkg> develop DecisionProgramming
If you have already cloned DecisionProgramming
from GitHub, you can use the command:
pkg> develop .
Inside DecisionProgramming
directory, run tests using the commands:
pkg> activate .
(DecisionProgramming) pkg> test
You can find more instruction on how to install packages for development at Julia's Pkg documentation.