bi-level.
org
Tutorial:
Bilevel Optimization
Without Tears
Jesús Mejía and Efrén Mezura-Montes
Artificial Intelligence Research Center
University of Veracruz, MEXICO
Optimization and Methods from Outline
origin of Bilevel Classical to
Optimization (BO) Metaheuristics
Final comments
Bilevel Solution
Introduction Applications Conclusions
Optimization Strategies
Practical examples to BO for solving interest
understand the formal problems
definition and its
properties bi-level.org 2
Introduction
3
Local optimum
Objective function
maximum
Optimization
101
Minimum
Rao, S. S. (2019). Engineering optimization: theory and practice.
John Wiley & Sons. bi-level.org 4
Minimization
“Find a best element (with regard to
some criterion) from a set of
available alternatives”
Optimization
101
Criterion
Best element
Set of
alternatives
bi-level.org 5
Minimization
“Find an argument that minimizes an
objective function”
Optimization
101
Objective
Optimum function
Search space
bi-level.org 6
Minimization
Optimization
101
bi-level.org 7
Minimization
Optimization
Example
bi-level.org 8
Minimize
where
Optimization and y is a constant equal to 1
Homework
bi-level.org 9
Minimize
where
Optimization and y is a constant equal to 1
Homework Solution:
bi-level.org 10
1934 A Leader-follower game was introduced
1973 Formulation for bilevel optimization
1977 Introduce term “Bilevel Programming”
Origin 1988 BO problems are NP-hard!
1992 BO problems are strongly NP-hard!
Of Bilevel Optimization
1994 First Genetic Algorithm for BO
2000-Present High interest from EC
community
Sinha, A. et al (2017). A review on bilevel optimization: from classical to
evolutionary approaches and applications. IEEE Transactions on Evolutionary
Computation. bi-level.org 11
Bilevel Optimization
12
“I am expert in Multi-objective
Bilevel optimization and my trusted friend
told me that BO is a bi-objective
Optimization optimization”
“Bilevel optimization is like
Wrong ideas on BO optimizing 2D functions”
“Nobody needs BO”
bi-level.org 13
Bilevel “An upper level authority takes a
Optimization decision subject to an optimal
response from a lower level
Description authority”
Zhang, G., et al (2015). Multi-level decision making. New York: Springer.
bi-level.org 14
“A guy finds the best way to escape,
subject to the optimal path (to the
Bilevel guy’s position) planned by the
chickens”
Optimization
Dummy example
bi-level.org 15
“A guy finds the best way to escape,
subject to the optimal path (to the
Bilevel guy’s position) planned by the
chickens”
Optimization Best leader Objective
Best follower
solution
function
solution
Dummy example
Possible solutions
bi-level.org 16
“A guy finds the best way to escape,
subject to the optimal path (to the
Bilevel guy’s position) planned by the
chickens”
Optimization Best follower Objective
Current leader
solution
function
solution
Dummy example
Possible solutions
bi-level.org 17
Bilevel Optimization: Nested Scheme
1. Leader takes a decision x
2. The follower uses the leader's
decision to take the best decision
based on f
3. The leader evaluates both x, y to
evaluate F
Zhang, G., et al (2015). Multi-level decision making. New York: Springer. 18
Bilevel Optimization: Formal Definition
Minimize: Leader /
Upper Level
Subject to:
Follower /
Lower Level
Dempe, Stephan. Foundations of bilevel programming. Springer Science & Business Media, 2002. 19
Bilevel Optimization: Solutions
Feasible solutions Optimal Feasible solutions
Dempe, Stephan. Foundations of bilevel programming. Springer Science & Business Media, 2002. 20
Bilevel Optimization: Practical Example
Minimize:
Subject to:
21
Bilevel Optimization: Practical Example
Step 1: Choose said
Step 2: Solve:
Step 3: Evaluate:
22
Bilevel Optimization: Practical Example
UL & LL functions Leader performance
s.t. best follower decision
Best follower
decision
Feasible solutions Optimum solution
23
True leader
performance
Follower
Bilevel takes the
best
Optimization Leader takes
decision
a decision
Practical Example
bi-level.org 24
Bilevel Optimization: General Definition
Minimize:
Subject to:
Dempe, Stephan. Foundations of bilevel programming. Springer Science & Business Media, 2002.
Dempe, S., et al. (2015). Bilevel programming problems. Energy Systems. Springer, Berlin.
25
Solution Strategies
26
Solution Strategies
Classical Approximate Hybrid
Mathematical Metaheuristics Metaheuristics +
programming Mathematical
Population-based
programming
● Karush-Kuhn-Tucker algorithms have been
conditions for successfully used. Usually,
single-level Karush-Kuhn-Tucker
● Evolutionary
reduction conditions are used in
Algorithms
● Branch and Bound addition to population
● Trust region ● Swarm Intelligence based algorithms for
● Among others ● Among others global optimization
Sinha, A. et al (2017). A review on bilevel optimization: from classical to evolutionary approaches and applications. IEEE bi-level.org 27
Transactions on Evolutionary Computation.
Solution Strategies Karush-Kuhn-Tucker conditions
Classical
Subject to:
Mathematical
programming
● Karush-Kuhn-Tucker
conditions for
single-level
reduction
● Branch and Bound
● Trust region
● Among others
bi-level.org 28
Solution Strategies
Approximate
Metaheuristics
Population-based The metaheuristics use
algorithms have been the nested scheme to
successfully used. sequentially optimize
● Evolutionary both levels.
Algorithms
● Swarm Intelligence
● Among others
bi-level.org 29
Talbi, E. G. (2013). Metaheuristics for bi-level optimization. Springer, Berlin, Heidelberg.
Solution Strategies UL Population
Approximate
Metaheuristics
Population-based
algorithms have been
successfully used.
LL Population
● Evolutionary
Algorithms
● Swarm Intelligence
● Among others
bi-level.org 30
Talbi, E. G. (2013). Metaheuristics for bi-level optimization. Springer, Berlin, Heidelberg.
Solution Strategies
Approximate
Metaheuristics
Population-based
algorithms have been
successfully used.
● Evolutionary
Algorithms
● Swarm Intelligence Differential Evolution (DE), Genetic Algorithm (GA), Sequential Quadratic Programming (SQP),
Interior Point Method (IP), Efficient Global Optimization (EGO).
● Among others
bi-level.org 31
Mejía-de-Dios, J. A., & Mezura-Montes, E. (2020). A surrogate-assisted metaheuristic for bilevel optimization. GECCO 2020.
Solution Strategies
Simple (naive) solution in Julia
Try it yourself online for free: bi-level.org/tutorial
bi-level.org 32
Applications
33
Applications Upper level authority that wants to
optimize the tolls revenue for a
Toll Setting Problem network of roads.
Followers are the network users
that want to optimize their
objectives (costs, time, etc).
Brotcorne, L., et al. (2001). A bilevel model for toll optimization on a
multicommodity transportation network. Transportation science.
bi-level.org 34
Applications
Automated Parameter Tuning
Leader: Chooses the parameters
Follower: Finds the hardest
instances for the target algorithm
Working paper. bi-level.org 35
Applications
Leader (principal) subcontracts a job to an
Principal-Agent Problems agent (follower). Uses an incentive scheme
that aligns the interests of the agent with the
principal.
Follower: (agent) prefers to act in his own
interests rather than those of the leader.
Cecchini, M., et al. (2013). Solving nonlinear principal-agent problems using bilevel
programming. European Journal of Operational Research.
bi-level.org 36
Conclusions
37
1. Bilevel Optimization (BO) is useful
to model problem with inherent
hierarchical structure
2. BO offers rich properties for solving
problems
3. Approximate methods can
Conclusions successfully solve real-world
problems
4. More theoretical studies are needed
5. BO problems can be hard to solve
because of the computational
complexity
38
Thank you!
Download this tutorial from:
Questions?
bi-level.org/tutorial
39
