MetaModel is a powerful design pattern that makes Gurobi modelling more efficient because it allows Operational Researchers to analyze models without worrying about manually logging progress and results.
- Easily take snapshots of the state of your model at any time
- Recreate any state of your model without having to store multiple versions of the model
- Works with any optimization solver that exposes an API
MetaModels came out of years working interactively with very large optimization models. These models were so large that it was prohibitively expensive to store multiple versions of the model anywhere. At the same time, these models took a very long time to generate. That is, it was much more efficient to generate a base model and then make changes to it interactively than it was to generate different models for each experiment. This means that I spent a lot of time analyzing and scripting in my Python interpreter.
A MetaModel is a thin object wrapper around an optimization model. It has a constructor that takes an optimization model as it's main argument, and a few simple methods. These methods allow you to add modules to your MetaModel, apply functions to your MetaModel, and store/load snapshots of your MetaModel. I use MetaModels now in my daily work because they make my workflow more efficient.
The example I'm going to show you is in Python using a Gurobi model. Find it in the meta_model directory as forest.lp.
forest.lp is a harvest scheduling model that seeks the optimal assignment of harvest schedules to cut-blocks. Our forest management model will run for 10 periods with each period representing 10 years.
The objective is to maximize the total volume of timber harvested.
The variables are:
-
x[i,j] - the number of hectares of cut-block, i, to be managed under schedule j, where j is a sequence of harvests. Each cut-block, i, has both hardwood and softwood timber species, belongs to either the north or south region, and has an initial age.
-
harv[s,r,t] - the volume of timber harvested from cut-blocks of species s (softwood or hardwood), in region, r (north or south), in period t.
-
age[r,t] - the area of the forest that is at least 60 years old in region, r, and period, t.
The constraints are:
- gub(i) - assignment problem constraints that say that no more than the area of each cut-block can be assigned to harvest schedules.
- harv(s,r,t) - inventory constraints that say that the harv[s,r,t] variables equal the volume harvested from cut-blocks of species, s, belonging to region, r, in period, t.
- age(r,t) - inventory constraints that say that the age[r,t] variables equal the sum of cut-blocks that are at least age 60 in period, t, and belong to region, r.
- env(r, t) - environmental constraints that say that at least 20% of the forest in each region, r has to be age 60 or greater after period 5.
# We'll create a MetaModel by passing the constructor
# the filename of an optimization model.
# The constructor opens the model using the Gurobi API and attaches
# it as an attribute of the MetaModel.
>>> from meta_model import MetaModel
>>> mm = MetaModel("forest.lp")Now there are some simple modifications that we want to perform on forest.lp so I've created a small Python module containing functions to perform those modifications. This module is called analysis_functions.py.
>>> # We'll add the analysis_functions module to the MetaModel
>>> mm.add_module("analysis_functions")
>>>
>>> # adding a description to a MetaModel makes it easy
>>> # to organize snapshots.
>>> mm.description = "Baseline model"
>>>
>>> # First we solve the model to establish a baseline.
>>> # We call a function from the analysis_functions module by passing
>>> # the module name and function name to the meta_function.
>>> mm.meta_function("analysis_functions.solve")
Optimize a model with 160 rows, 1945 columns and 10172 nonzeros
Coefficient statistics:
Matrix range [1e+00, 2e+02]
Objective range [1e+00, 1e+00]
Bounds range [0e+00, 0e+00]
RHS range [1e+00, 1e+01]
Presolve removed 50 rows and 1102 columns
Presolve time: 0.06s
Presolved: 110 rows, 843 columns, 2770 nonzeros
Iteration Objective Primal Inf. Dual Inf. Time
0 4.1153473e+04 1.100000e+01 0.000000e+00 0s
20 4.0724706e+04 0.000000e+00 0.000000e+00 0s
Solved in 20 iterations and 0.08 seconds
Optimal objective 4.072470571e+04
Snapshot saved as forest_2017330_0.jsonWhen we called meta_function("analysis_functions.solve") a few things happened. First the MetaModel looked to see that analysis_functions is one of its modules, and that solve is a function of analysis_functions. Then it passes the MetaModel to analysis_functions.solve, and solve operates on the MetaModel.
You'll notice that at the end of the optimization Snapshot saved as forest_2017330_0.json was printed to screen. This means that analysis_functions.solve calls the MetaModel's take_snapshot method, which serializes the current state of the model for easy recreating in the future. The serialized model state is stored at forest_2017330_0.json. This name isn't as intimidating as it looks. It's just the name of the model, concatenated with today's date, and the version of the model that's being serialized, in this case 0.
The analysis_functions.solve function that I've written also writes variable solution values to a csv file, forest_2017330_0_sol.csv. This enables us to go back to solutions later and analyze them using Excel without having to load a MetaModel.
Now we'll make changes to the model and solve it again.
>>> # Shorten the length of the
>>> # planning horizon by one period.
>>> mm.description = "Reduce length of planning horizon to 9"
>>> mm.meta_function("analysis_functions.remove_last_period")
>>> mm.meta_function("analysis_functions.solve")
Optimize a model with 152 rows, 1939 columns and 8706 nonzeros
Coefficient statistics:
Matrix range [1e+00, 2e+02]
Objective range [1e+00, 1e+00]
Bounds range [0e+00, 0e+00]
RHS range [1e+00, 1e+01]
Presolve removed 44 rows and 1370 columns
Presolve time: 0.00s
Presolved: 108 rows, 569 columns, 1682 nonzeros
Iteration Objective Primal Inf. Dual Inf. Time
0 3.8129707e+04 5.000000e+00 0.000000e+00 0s
7 3.8025779e+04 0.000000e+00 0.000000e+00 0s
Solved in 7 iterations and 0.00 seconds
Optimal objective 3.802577891e+04
Snapshot saved as forest_2017330_1.jsonHere when we passed analysis_functions.remove_last_period to the MetaModel, the function was applied to the MetaModel, reducing the length of the planning horizon from 10 periods to 9 periods, and then a record of the function being applied was stored to the MetaModel. This means that when we called solve and a snapshot was taken, a record of analysis_functions.remove_last_period was stored in forest_2017330_1.json.
Now we'll show how to call a function that takes arguments.
>>> # Change the objective function of the model so that
>>> # it maximizes ecosystem condition instead
>>> # of harvest volume.
>>> mm.description = "Maximize Ecosystem condition"
>>>
>>> # Set harvet variable objective coefficients to zero
>>> mm.meta_function("analysis_functions.zero_objective_coeffs")
>>> # Set age variable objective coeffs to 1
>>> mm.meta_function("analysis_functions.set_variables_attr,
args=("obj", 1, "age"))
>>> mm.meta_function("analysis_functions.solve")
Optimize a model with 152 rows, 1939 columns and 8706 nonzeros
Coefficient statistics:
Matrix range [1e+00, 2e+02]
Objective range [1e+00, 1e+00]
Bounds range [0e+00, 0e+00]
RHS range [1e+00, 1e+01]
Iteration Objective Primal Inf. Dual Inf. Time
0 5.4000000e+32 9.131251e+32 5.400000e+02 0s
202 3.7500000e+02 0.000000e+00 0.000000e+00 0s
Solved in 202 iterations and 0.00 seconds
Optimal objective 3.750000000e+02
Snapshot saved as forest_2017330_2.jsonHere we changed the objective function of the model from maximize harvest volume to maximize the amount of old forest. We did this by first zeroing all the objective coefficients and then calling analysis_functions.set_variables_attrand passing it the arguments "obj", 1, and "age". Functions called with meta_function must have JSON serializable arguments. Note that keyword arguments could be passed in a similar fashion by using the kwargs parameter.
Let's say that we come back tomorrow and want to pick up where we left off.
>>> from meta_model import MetaModel
>>> mm = MetaModel(json_file="forest_2017330_2.json")
>>> print mm.description
Maximize Ecosystem condition
>>> mm.meta_function("analysis_functions.solve")
Optimize a model with 152 rows, 1939 columns and 8706 nonzeros
Coefficient statistics:
Matrix range [1e+00, 2e+02]
Objective range [1e+00, 1e+00]
Bounds range [0e+00, 0e+00]
RHS range [1e+00, 1e+01]
Presolve removed 44 rows and 1394 columns
Presolve time: 0.00s
Presolved: 108 rows, 545 columns, 1658 nonzeros
Iteration Objective Primal Inf. Dual Inf. Time
0 1.1141210e+03 2.035147e+03 0.000000e+00 0s
112 3.7500000e+02 0.000000e+00 0.000000e+00 0s
Solved in 112 iterations and 0.00 seconds
Optimal objective 3.750000000e+02
Snapshot saved as forest_2017331_3.jsonAnd we're exactly where we left off yesterday. When we passed the snapshot location to the MetaModel constructor, the original Gurobi model was loaded into the MetaModel, and the functions that we had applied to the model were applied again in the proper order, so that we've recovered exactly the model state from yesterday.
This tutorial has provided an introduction to the MetaModel design pattern. It has shown how to create, modify, and load a MetaModel. MetaModels can make Gurobi modelling more efficient because the modeller can focus on anlaysis instead of tracking and logging their progress.