MIPLearn.jl
**MIPLearn** is an extensible open-source framework for solving discrete optimization problems using a combination of Mixed-Integer Linear Programming (MIP) and Machine Learning (ML). See the [main repository](https://github.com/ANL-CEEESA/MIPLearn) for more information. This repository holds an experimental Julia interface for the package.
[miplearn]: https://github.com/ANL-CEEESA/MIPLearn
## 1. Usage
### 1.1 Installation
To use MIPLearn.jl, the first step is to [install the Julia programming language on your machine](https://julialang.org/). After Julia is installed, launch the Julia console, type `]` to switch to package manager mode, then run:
```
(@v1.6) pkg> add MIPLearn@0.2
```
This command should also automatically install all the required Python dependencies. To test that the package has been correctly installed, run (in package manager mode):
```
(@v1.6) pkg> test MIPLearn
```
If you find any issues installing the package, please do not hesitate to [open an issue](https://github.com/ANL-CEEESA/MIPLearn.jl/issues).
### 1.2 Describing instances
```julia
using JuMP
using MIPLearn
# Create problem data
weights = [1.0, 2.0, 3.0]
prices = [5.0, 6.0, 7.0]
capacity = 3.0
# Create standard JuMP model
model = Model()
n = length(weights)
@variable(model, x[1:n], Bin)
@objective(model, Max, sum(x[i] * prices[i] for i in 1:n))
@constraint(model, c1, sum(x[i] * weights[i] for i in 1:n) <= capacity)
# Add ML information
@feature(model, [5.0])
@feature(c1, [1.0, 2.0, 3.0])
@category(c1, "c1")
for i in 1:n
@feature(x[i], [weights[i]; prices[i]])
@category(x[i], "type-$i")
end
instance = JuMPInstance(model)
```
### 1.3 Solving instances and training
```julia
using MIPLearn
using Cbc
# Create training and test instances
training_instances = [...]
test_instances = [...]
# Create solver
solver = LearningSolver(Cbc.Optimizer)
# Solve training instances
for instance in train_instances
solve!(solver, instance)
end
# Train ML models
fit!(solver, training_instances)
# Solve test instances
for instance in test_instances
solve!(solver, instance)
end
```
### 1.4 Saving and loading solver state
```julia
using MIPLearn
using Cbc
# Solve training instances
training_instances = [...]
solver = LearningSolver(Cbc.Optimizer)
for instance in training_instances
solve!(solver, instance)
end
# Train ML models
fit!(solver, training_instances)
# Save trained solver to disk
save("solver.bin", solver)
# Application restarts...
# Load trained solver from disk
solver = load_solver("solver.bin")
# Solve additional instances
test_instances = [...]
for instance in test_instances
solve!(solver, instance)
end
```
### 1.5 Solving instances from disk
In all examples above, we have assumed that instances are available as `JuMPInstance` objects, stored in memory. When problem instances are very large, or when there is a large number of problem instances, this approach may require an excessive amount of memory. To reduce memory requirements, MIPLearn.jl can also operate on instances that are stored on disk, through the `FileInstance` class, as the next example illustrates.
```julia
using MIPLearn
using JuMP
using Cbc
# Create a large number of problem instances
for i in 1:600
# Build JuMP model
model = Model()
@variable(...)
@objective(...)
@constraint(...)
# Add ML features and categories
@feature(...)
@category(...)
# Save instances to a file
instance = JuMPInstance(m)
save("instance-$i.bin", instance)
end
# Initialize training and test instances
training_instances = [FileInstance("instance-$i.bin") for i in 1:500]
test_instances = [FileInstance("instance-$i.bin") for i in 501:600]
# Initialize solver
solver = LearningSolver(Cbc.Optimizer)
# Solve training instances. Files are modified in-place, and at most one
# file is loaded to memory at a time.
for instance in training_instances
solve!(solver, instance)
end
# Train ML models
fit!(solver, training_instances)
# Solve test instances
for instance in test_instances
solve!(solver, instance)
end
```
### 1.6 Solving training instances in parallel
In many situations, instances can be solved in parallel to accelerate the training process. MIPLearn.jl provides the method `parallel_solve!(solver, instances)` to easily achieve this.
First, launch Julia in multi-process mode:
```
julia --procs 4
```
Then call `parallel_solve!` as follows:
```julia
@everywhere using MIPLearn
@everywhere using Cbc
# Initialize training and test instances
training_instances = [...]
test_instances = [...]
# Initialize the solver
solver = LearningSolver(Cbc.Optimizer)
# Solve training instances in parallel. The number of instances solved
# simultaneously is the same as the `--procs` argument provided to Julia.
parallel_solve!(solver, training_instances)
# Train machine learning models
fit!(solver, training_instances)
# Solve test instances in parallel
parallel_solve!(solver, test_instances)
```
**NOTE:** Only `FileInstance` instances are currently supported.
## 2. Customization
### 2.1 Selecting solver components
```julia
using MIPLearn
solver = LearningSolver(
Cbc.Optimizer,
components=[
PrimalSolutionComponent(...),
ObjectiveValueComponent(...),
]
)
```
### 2.2 Adjusting component aggressiveness
```julia
using MIPLearn
solver = LearningSolver(
Cbc.Optimizer,
components=[
PrimalSolutionComponent(
threshold=MinPrecisionThreshold(0.95),
),
]
)
```
### 2.3 Evaluating component performance
TODO
### 2.4 Using customized ML classifiers and regressors
TODO
## 3. Acknowledgments
* Based upon work supported by **Laboratory Directed Research and Development** (LDRD) funding from Argonne National Laboratory, provided by the Director, Office of Science, of the U.S. Department of Energy under Contract No. DE-AC02-06CH11357.
* Based upon work supported by the **U.S. Department of Energy Advanced Grid Modeling Program** under Grant DE-OE0000875.
## 4. Citing MIPLearn
If you use MIPLearn in your research (either the solver or the included problem generators), we kindly request that you cite the package as follows:
* **Alinson S. Xavier, Feng Qiu.** *MIPLearn: An Extensible Framework for Learning-Enhanced Optimization*. Zenodo (2020). DOI: [10.5281/zenodo.4287567](https://doi.org/10.5281/zenodo.4287567)
If you use MIPLearn in the field of power systems optimization, we kindly request that you cite the reference below, in which the main techniques implemented in MIPLearn were first developed:
* **Alinson S. Xavier, Feng Qiu, Shabbir Ahmed.** *Learning to Solve Large-Scale Unit Commitment Problems.* INFORMS Journal on Computing (2020). DOI: [10.1287/ijoc.2020.0976](https://doi.org/10.1287/ijoc.2020.0976)
## 5. License
Released under the modified BSD license. See `LICENSE` for more details.