Using trained machine learning predictors in Gurobi

Xavier Nodet
Development Manager
February 2023
ROADEF
Gurobi
Machine Learning
Using trained machine learning predictors in Gurobi
© 2023 Gurobi Optimization, LLC. All Rights Reserved

Agenda
Motivating example
Optimizing avocado shipments and prices
Gurobi Machine Learning
An open-source Python package
Gurobi 10.0
Performance improvements for models with
ML predictor constraints
© 2023 Gurobi Optimization, LLC. All Rights Reserved | 2

Optimizing avocado shipments and prices
Motivating example

• Selling avocados in the US
• Market is split in regions 𝑅
• Total supply 𝑆
• Maximize profit:
• (sales – shipping costs – unsold penalty)
with given
• shipping costs 𝑐! and waste penalty 𝑤
• We decide about the prices 𝑝!
• and estimate the demands d" from the prices
Motivating Example: Price Optimization
https://youtu.be/AJRP9pPBx6s

• Historical data of avocado sales from
Hass Avocado Board (HAB) available
on Kaggle and HAB website
• Features correlated to demand:
price, region, year, peak season
• Linear regression gives reasonably
good prediction of demand with those:
𝑑 = 𝑔(𝑝, 𝑟, 𝑦𝑒𝑎𝑟, 𝑠𝑒𝑎𝑠𝑜𝑛)
• In the case of linear regression, 𝑔 is an
affine function
𝑑 = 𝜙# 𝑝, 𝑟, 𝑦𝑒𝑎𝑟, 𝑠𝑒𝑎𝑠𝑜𝑛 + 𝜙$
Motivating Example: Estimating Demand

Motivating Example: Retrieving ML results
coef_dict = model.fit().params.to_dict()
d = {r: (coef_dict['Intercept’] +
coef_dict['price’] * p[r] +
coef_dict['C(region)[T.%s]'%r] +
coef_dict['year_index’] * (year - 2015) +
coef_dict['peak’] * peak_or_not)
for r in R}
m.addConstrs((s[r] <= d[r] for r in R))
• Retrieve results from linear regression model
• Build linear expression defining expected demand from price
• Upper bound on sales is expected demand

• Indices
• r ∈ 𝑅, the set of regions
• Variables
• p" selling price per unit
• d" demand
• u total unsold products
Motivating Example: Optimization Model
max ∑! 𝑝! − 𝑐! 𝑑! − 𝑤 ∗ 𝑢 (maximize revenue)
𝑠. 𝑡.
∑! 𝑑! + 𝑢 = 𝑆 (allocate shipments)
𝑑! = 𝑔 𝑝!, 𝑟, 𝑦𝑒𝑎𝑟, 𝑠𝑒𝑎𝑠𝑜𝑛 for r ∈ 𝑅 (define demand with regression model)
• Indices
• r ∈ 𝑅, the set of regions
• Constants
• 𝑐! transportation costs
• 𝑤 penalty for waste
• 𝑆 total supply
• Year and season of interest

Motivating Example: What if?
max ∑! 𝑝! − 𝑐! 𝑑! − 𝑤 ∗ 𝑢 (maximize revenue)
𝑠. 𝑡.
∑! 𝑑! + 𝑢 = 𝑆 (allocate shipments)
𝑑! = 𝑔 𝑝!, 𝑟, 𝑦𝑒𝑎𝑟, 𝑠𝑒𝑎𝑠𝑜𝑛 for r ∈ 𝑅 (define demand with regression model)
• With 𝑔 an affine function, the resulting model is a non-convex QP
• Solved fast with Gurobi
• But what if we need a more accurate prediction with a more complex regression:
• Decision tree, Neural network, …

1. Simplify the process of importing a trained machine learning model built with a
popular ML package into an optimization model.
2. Improve algorithmic performance to enable the optimization model to explore a
sizable space of solutions that satisfy the relationships between variables captured
with the ML model.
Motivating Example: Goals

An open-source Python package

• Open source python package:
• https://github.com/Gurobi/gurobi-machinelearning
• https://gurobi-machinelearning.readthedocs.io/
• Apache License 2.0
• Initial release 1.0.0 last November
• Version 1.1.0 recently released
• Experimental package
• Not supported through usual Gurobi processes

• Linear/Logistic/PLS regression
• Decision trees
• Neural network with ReLU activation
• Random Forests
• Gradient Boosting
• Preprocessing:
• Simple scaling
• Polynomial features of degree 2
• Column transformers
• pipelines to combine them
• Dense layers
• ReLU layers
• Object Oriented, functional or
sequential
• Dense layers
• ReLU layers
• Only torch.nn.Sequential models
Known regression models

Example: Define and train ML model
...
lin_reg = make_pipeline(columntransformer, LinearRegression())
lin_reg.fit(X_train, y_train)

• Variables
• 𝑝! selling price per unit
• 𝑑! demand
• 𝑢 total unsold products
Example: Creating the variables
import gurobipy_pandas as gppd
data = pd.concat(...)
m = gp.Model("Avocado_Price_Allocation")
p = gppd.add_vars(m, data, lb=0.0, ub=2.0)
d = gppd.add_vars(m, data)
u = m.addVar()
m.setObjective(((p – c) * d).sum() - w * u, GRB.MAXIMIZE)
m.addConstr(d.sum() + u == S)

Example: Adding regression constraints
fixed = pd.DataFrame(
data={
"year": 2020,
"peak": 1,
"region": regions,
},
index=regions)
feats = pd.concat(
[fixed, p],
axis=1)
𝑑! = 𝑔 𝑝!, 𝑟, 𝑦𝑒𝑎𝑟, 𝑠𝑒𝑎𝑠𝑜𝑛 for r ∈ 𝑅.

Example: Adding Regression Constraints
from gurobi_ml import add_predictor_constr
pred_constr = add_predictor_constr(m, lin_reg, feats, d)
pred_constr.print_stats()

Example: Optimizing
m.Params.NonConvex = 2
m.optimize()

Example: Solution

• Function add_predictor_constr creates the formulation for regression
model and returns a modeling object.
• If input of add_predictor_constr has several rows, it introduces one sub-
model for each row
• Query statistics about modeling object, remove it, query solution after solve and
error in solutions.
• Models for logistic regression use a piecewise linear approximation and can have
modeling error (controlled by parameters).
• Models for decision tree can also introduce small errors at threshold values of
node splitting (can be controlled).
Remarks

Performance improvements for models with ML
predictor constraints
Gurobi 10.0

Gurobi 9.5 vs Gurobi 10.0
• 478 instances
• 10.000 seconds time limit
• Intel(R) Xeon(R) CPU E3-1240 CPUs, 4 cores, 4
threads
• Solved means 0.01% gap reached
• Models not solved by any version were
excluded
• Models incorporating Neural
Networks
• Performance gains come mostly
from Optimization Based Bound
Tightening (OBBT)

Conclusion

• Gurobi Machine Learning
https://github.com/Gurobi/gurobi-machinelearning
• This is experimental
• Input very welcome
• Performance for models with Neural Networks in
Gurobi 10
Dangers and Pitfalls
• ML models we can hope to handle is still limited (e.g. no categorical features)
• Methodological issues:
• How to decide which prediction model to use?
• How to make sure that optimization doesn’t misuse results of the predictor?
Conclusion

For more details…
https://www.gurobi.com/events/
using-trained-machine-learning-
predictors-in-gurobi

For more information: gurobi.com
Xavier Nodet
Development Manager
xavier.nodet@gurobi.com
Thank You https://bit.ly/xnodet-slides

Using trained machine learning predictors in Gurobi

More Related Content

What's hot (20)

Similar to Using trained machine learning predictors in Gurobi (17)

Recently uploaded (20)

Using trained machine learning predictors in Gurobi