102
MGDA applied to compressible aerodynamics J.-A. Désidéri Foreword MGDA Pareto-stationarity Descent direction Main results Practicalities Mathematical test-cases Aircraft-Wing Aerodynamics Automobile Cooling System Design Problem description and numerical tools Software framework Numerical results Domain-Partitioning Model Problem Model problem Quasi-Newton Method Basic MGDA MGDA-II MGDA-III Bird of paradise Algorithm Properties MGDA-IV Conclusions Cooperation and Competition MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA) FOR MULTIOBJECTIVE OPTIMIZATION Jean-Antoine Désidéri INRIA Project-Team OPALE Sophia Antipolis Méditerranée Centre http://www-sop.inria.fr/opale ISMP 2012 21st International Synposium on Mathematical Programming, Berlin, August 19-24, 2012 PDE-Constrained Optimization and Multi-Level/Multi-Grid Methods ECCOMAS 2012 European Congress on Computational Methods in Applied Sciences and Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1 / 102

MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MULTIPLE GRADIENT DESCENTALGORITHM (MGDA)

FOR MULTIOBJECTIVE OPTIMIZATION

Jean-Antoine DésidériINRIA Project-Team OPALE

Sophia Antipolis Méditerranée Centrehttp://www-sop.inria.fr/opale

ISMP 201221st International Synposium on Mathematical Programming,

Berlin, August 19-24, 2012

PDE-Constrained Optimization and Multi-Level/Multi-Grid Methods

ECCOMAS 2012European Congress on Computational Methods in Applied Sciences and

Engineering, Vienna (Austria), September 10-14, 2012

TO16: Computational inverse problems and optimization

1 / 102

Page 2: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

2 / 102

Page 3: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Foreword: Pareto optimality• Design vector:

Y ∈Υ⊆ RN

• Objective functions to be minimized (reduced):

Ji (Y ) (i = 1, . . . ,n)

Dominance in efficiency(see e.g. K. Miettinen: Nonlinear Multiobjective Optimization, Kluwer Academic Publishers)

Design-point Y 1 dominates design-point Y 2 in efficiency, Y 1 Y 2,iff

Ji (Y 1)≤ Ji (Y 2) (∀i = 1, . . . ,n)

and at least one inequality is strict. (relationship of partial order)

The designer’s Holy Grail: the Pareto setset of "non-dominated design-points" or "Pareto-optimal solutions"

The Pareto frontimage of the Pareto set in the objective function space(bounds the domain of attainable performance)

3 / 102

Page 4: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Pareto front identificationFavorable situation: continuous and convex Pareto frontGradient-based optimization can be used efficiently, if number n ofobjective functions is not too large

Non-convex or discontinuous Pareto fronts do existEvolutionary strategies or GA’s most commonly-used forrobustness: NSGA-II (Deb), PAES

From: A Fast and Elitist Multiobjective Ge-netic Algorithm: NSGA-II, K. Deb, A. Pratap,S. Agarwal, and T. Meyarivan, IEEE Trans-actions in Evolutionary Computation, Vol. 6,No. 2, April 2002.

Nondominated solutions with NSGA-II on KUR (testcase by F. Kursawe, 1990)

n = 3 f1(x) =n−1

∑i=1

−10exp

−√

x2i + x2

i+1

5

f2(x) =n

∑i=1

(|xi |

45 + 5sinx3

i

)4 / 102

Page 5: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Our general objective

Construct a robust gradient-based method for Paretofront identification

Note : in shape optimization, N is often the dimension of a parameterization, meant to be large

as the dimension of a discretization usually is; thus, often, N n. However, the following

theoretical results also apply to cases where n > N which can be the case of other situations in

engineering optimization.

5 / 102

Page 6: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

6 / 102

Page 7: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Constructing a gradient-basedcooperative algorithm

The basic question :Knowing the gradients, ∇Ji (Y 0), of n criteria, assumed to besmooth functions of the design vector Y ∈ RN , at a given startingdesign-point Y 0, can we define a nonzero vector ω ∈ RN , in thedirection of which the Fréchet derivatives of all criteria have thesame sign, (

∇Ji (Y 0),ω)≥ 0 (∀i = 1,2, ...,n)?

Answer:Yes, if Y 0 is not Pareto-optimal!

Then, −ω is locally a descent direction common to allcriteria.

7 / 102

Page 8: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Notion of Pareto-stationarityClassically, for smooth real-valued functions of Nvariables

Optimality (+ regularity) =⇒ Stationarity

Now, for smooth Rn-valued functions of N variables,which stationarity requirement is to be enforced forPareto-optimality?

Proposition 1: Pareto Stationarity at a design point Y 0

in an open domain B in which the objective-functions are smoothand convex, if there exists a convex combination of the gradientsequal to 0:

∃α = αi(i=1,,n) s.t.n

∑i=1

αi∇Ji (Y 0) = 0 , αi ≥ 0 (∀i) ,n

∑i=1

αi = 1

Then (to be established subsequently):

Pareto-optimality (+ regularity) =⇒ Pareto-stationarity

8 / 102

Page 9: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

General principle

Proposition 2: Minimum-norm element in the convex hullLet ui(i=1,...,n) be a family of n vectors in RN , and U the so-called convexhull of this family, i.e. the following set of vectors :

U =

u ∈ RN / u =

n

∑i=1

αi ui ; αi ≥ 0 (∀i) ;n

∑i=1

αi = 1

Then, U admits a unique element of minimal norm, ω, and the followingholds :

∀u ∈ U : (u,ω)≥ ‖ω‖2

in which (u,v) denotes the usual scalar product of the vectors u and v .

9 / 102

Page 10: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Existence and uniqueness of ω

Existence⇐= U closed. Uniqueness⇐= U convex.To establish uniqueness, let ω1 and ω2 be two realisations of the minimum:

‖ω1‖= ‖ω2‖= Argminu∈U ‖u‖

Then, since U is convex, ∀ε ∈ [0,1], ω1 + εr12 ∈ U (r12 = ω2−ω1) and:

‖ω1 + εr12‖ ≥ ‖ω1‖

Square both sides, and remplace by scalar products:

∀ε ∈ [0,1] : (ω1 + εr12,ω1 + εr12)− (ω1,ω1)≥ 0

Then∀ε ∈ [0,1] : 2ε(r12,ω1) + ε

2(r12, r12)≥ 0

As ε→ 0+, this condition requires that (r12,ω1)≥ 0; but then, for ε = 1:

‖ω2‖2−‖ω1‖2 = 2(r12,ω1) + (r12, r12) > 0

unless r12 = 0, i.e. ω2 = ω1.

Remark: The identification of the element ω is equivalent to the constrained minimization of aquadratic form in Rn , and not RN .

10 / 102

Page 11: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

First consequence

∀u ∈ U : (u,ω)≥ ‖ω‖2

Proof:Let u ∈ U, and r = u−ω.∀ε ∈ [0,1], ω + εr ∈ U (convexity); hence:

‖ω + εr‖2−‖ω‖2 = 2ε(r ,ω) + ε2(r , r)≥ 0

and this requires that:

(r ,ω) = (u−ω,ω)≥ 0

11 / 102

Page 12: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Proof of Pareto-stationarityat Pareto-optimal design-point Y 0 ∈ B (open ball) ⊆ RN

Hyp : Ji (Y ) (∀i ≤ n) smooth and convex in B ; ui = ∇Ji (Y 0) (∀i ≤ n)

Without loss of generality, assume Ji (Y 0) = 0 (∀i). Then:

Y 0 = ArgminY

Jn(Y ) subject to: Ji (Y )≤ 0 (∀i ≤ n−1) (1)

Let Un−1 be the convex hull of the gradients u1,u2, . . . ,un−1 and ωn−1 = Argminu∈Un−1‖u‖.

The vector ωn−1 exists, is unique, and such that:(ui ,ωn−1

)≥ ‖ωn−1‖2 (∀i ≤ n−1).

Two possible situations:1. ωn−1 = 0, and the objective-functions J1,J2, . . . ,Jn−1 satisfy the Pareto stationarity

condition at Y = Y 0. A fortiori, the condition is also satisfied by the whole set ofobjective-functions.

2. Otherwise ωn−1 6= 0. Let ji (ε) = Ji (Y 0− εωn−1) (i = 1, . . . ,n−1) so that ji (0) = 0 andj ′i (0) =−

(ui ,ωn−1

)≤−‖ωn−1‖2 < 0, and for sufficiently small strictly-positive ε:

ji (ε) = Ji (Y 0− εωn−1) < 0 (∀i ≤ n−1)

and Slater’s qualification condition is satisfied for (1). Thus, the LagrangianL = Jn(Y ) + ∑

n−1i=1 λi Ji (Y ) is stationary w.r.t. Y at Y = Y 0:

un +n−1

∑i=1

λi ui = 0

in which λi > 0 (∀i ≤ n−1) since equality constraints hold Ji (Y 0) = 0 (KKT condition).Normalizing this equation results in the Pareto-stationarity condition.

12 / 102

Page 13: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

RN Affine and Vector StructuresNotation

Usually indistinct

When distinction necessary, a dotted symbol is used foran affine subset/subspace

In particular for the convex hull Uthe set of points vectors of a given origin O and equipollent to u ∈ Upoint to.

Rigorously speaking, the term "convex hull" applies to U.

13 / 102

Page 14: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

RN Affine and Vector StructuresConvex hulls

A case where O⊥ /∈ U

U (affine)

U(vector)

A2

O

O⊥u1

u2

u3

u ∈ U

14 / 102

Page 15: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Second consequenceNote that ∀u ∈ U:

u−un =n

∑i=1

αiui −

(n

∑i=1

αi

)un =

n−1

∑i=1

αiun,i (un,i = ui −un)

Thus U⊆ An−1 (or in affine space, U⊆ An−1),where An−1 is a set of vectors pointing to an affine sub-space An−1

of dimension at most n−1.

Define the orthogonal projection O⊥ of O onto An−1

Since−−→OO⊥ ⊥ An−1 ⊇ U:

O⊥ ∈ U⇐⇒ ω =−−→OO⊥

In this case:∀u ∈ U : (u,ω) = const. = ‖ω‖2

15 / 102

Page 16: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Second consequence

ProofSince O⊥ is the orthogonal projection of O onto An−1,

−−→OO⊥ ⊥ An−1, and in particular

−−→OO⊥ ⊥ U.

If additionally O⊥ ∈ U, vector−−→OO⊥ is the minimum-norm element of U: ω =

−−→OO⊥.

In this case, ω can be calculated by ignoring the inequality constraints, ∀i, αi ≥ 0, that areautomatically satisfied by the solution.

Hence, vector ω realizes the minimum of the quadratic form ‖ω‖2 = ‖∑ni=1 αi ui‖2 uniquely

subject to the equality constraint: ∑ni=1 αi = 1.

The Lagrangian is formed with a unique multiplier λ:

L(α,λ) = 12

(n

∑i=1

αi ui ,n

∑i=1

αi ui

)+ λ

(n

∑i=1

αi −1

)

The stationarity condition requires that:

∀i, ∂L∂αi

= (ui ,ω) + λ = 0 =⇒ (ui ,ω) =−λ = const. = ‖ω‖2

By convex combination, the result extends straightforwardly to the whole U.

16 / 102

Page 17: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Application to the case ofgradients

Suppose ui = ∇Ji(Y 0) (∀i); then:

• either ω 6= 0, and all criteria admit positive Fréchet derivativesin the direction of ω (all equal if ω belongs to the interior U)

• or ω = 0, and the current design point Y 0 is Pareto stationary:

∃αii=1,2,...,n (αi ≥ 0,∀i;n

∑i=1

αi = 1) so thatn

∑i=1

αiui = 0

17 / 102

Page 18: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Proposition 3: Common Descent DirectionBy virtue of Propositions 1 and 2 in the particular case where

ui = J ′i =∇Ji(Y 0)

Si

(Si : user-supplied scaling constant; Si > 0), two situations are possible atY = Y 0 :

• either ω = 0, and the design point Y 0 is Pareto-stationary (orPareto-optimal);

• or ω 6= 0, and −ω is a descent direction common to all criteriaJi(x)(i=1,...,n); additionally, if ω⊥ U, the scalar product (u,ω)

(u ∈ U), and the Frechet derivatives (ui ,ω) are all equal to ‖ω‖2.

MGDA : substitute vector ω to the single-criterion gradientin the steepest-descent method.

Proposition 4: ConvergenceCertain normalization provisions being made, in RN , the MultipleGradient Descent Algorithm (MGDA) converges to aPareto-stationary point.

18 / 102

Page 19: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Convergence proof

Define the criteria to be positive (possibly apply an exp-transform), and ∞ at ∞ (possibly add anexploding term outside of the working ball):

∀i, Ji (Y )→ ∞ as ‖Y‖→ ∞.

Assume these criteria to be continuous.

If the iteration stops in a finite number of steps, a Pareto-stationary point has been reached.

Otherwise, the iteration continues indefinitely, generating an infinite sequence of design points,Y k. The corresponding sequences of criteria are infinite, positive and monotone decreasing.They are bounded. Hence, the sequence of iterates, Y k, is itself bounded and it admits asubsequence converging to say Y ?.Necessarily, Y ? is a Pareto-stationary point. To establish this, assume instead that ω?, whichcorresponds to Y ∗, is nonzero. Then for each criterion, there exists a stepsize ρ for which, thevariation is finite. These criteria are in finite number. Hence, the smallest ρ wlll cause a finitevariation to all criteria, and this is in contradiction with the fact that only infinitely-small variationsof the criteria are realized from Y ?.

19 / 102

Page 20: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Remark 1 : practicaldetermination of vector ω

in the convex hull

Problem to be solved in U⊂ RN

ω = Argminu∈U ‖u‖

U =

u ∈ RN / u =

n

∑i=1

αi ui ; αi ≥ 0 (∀i) ;n

∑i=1

αi = 1

Usually, but not necessarily : N ≥ n.

ParameterizationLet

αi = σ2i (i = 1, ...,n)

to satisfy trivially the inequality constraints (αi ≥ 0 ,∀i), and transform the equality constraint,∑

ni=1 αi = 1 into

n

∑i=1

σ2i = 1⇐⇒ σ = (σ1,σ2, ...,σn) ∈ Sn

where Sn is the unit sphere of Rn and precisely not RN .

20 / 102

Page 21: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Determining vector ω (cont’d)The sphere is easily parameterizedusing trigonometric functions of n−1 independent arcs φ1,φ2, ...,φn−1:

σ1 = cosφ1 . cosφ2 . cosφ3 . ... . cosφn−1σ2 = sinφ1 . cosφ2 . cosφ3 . ... . cosφn−1σ3 = 1 . sinφ2 . cosφ3 . ... . cosφn−1

.

.

....

σn−1 = 1 . 1 . ... . sinφn−2 . cosφn−1σn = 1 . 1 . ... . 1 . sinφn−1

(Consider only : φi ∈ [0,π/2] ,∀i and set : φ0 = π

2 .)

Let ci = cos2 φi (i = 1, ...,n)and get:

α1 = c1 . c2 . c3 . ... . cn−1α2 = (1− c1) . c2 . c3 . ... . cn−1α3 = 1 . (1− c2) . c3 . ... . cn−1

.

.

....

αn−1 = 1 . 1 . ... . (1− cn−2) . cn−1αn = 1 . 1 . ... . 1 . (1− cn−1)

(c0 = 0, and ci ∈ [0,1] for all i ≥ 1).

21 / 102

Page 22: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Determining vector ω (end)

The convex hull is thus parameterized in

[0,1]n−1 (n : number of criteria)

independently of N (dimension of design space).

For example, with n = 4 criteria :α1 = c1c2c3

α2 = (1− c1)c2c3

α3 = (1− c2)c3

α4 = (1− c3)

(c1,c2,c3) ∈ [0,1]3

22 / 102

Page 23: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Remark 2 : appropriatenormalization of the gradients

may reveal to be essential : ui = ∇Ji (Y 0)/SiCase n = 2

• Without gradient normalizationThen, ω =

−−→OO⊥ unless the angle < u1,u2 > is acute, and the norms ||u1||, ||u2|| are very

different; in that case, ω is equal the one of smaller norm.

u1u2 ω =−−→OO⊥

u1

u2

ω =−−→OO⊥

u1u2 = ω−−→OO⊥

• With normalization: The equality ω =−−→OO⊥ holds automatically =⇒

EQUAL FRÉCHET DERIVATIVES

General CaseIF the vectors ui are NOT NORMALIZED, those of smallernorms are more influential to determine the direction of vector ω

23 / 102

Page 24: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Normalizations being examinedRecall

ui = J ′i = ∇Ji (Y 0)/Si (Si > 0) =⇒ ω

δY =−ρω =⇒ δJi = (∇Ji ,δY ) =−ρSi (ui ,ω)

(ui ,ω) = const. if ω does not lie on the “edge” of convex hull

• Standard :

ui =∇Ji (Y 0)

‖∇Ji (Y 0)‖• Equal logorithmic variations (whenever ω is not on edge) :

ui =∇Ji (Y 0)

Ji (Y 0)

• Newton-inspired when limJi = 0 :

ui =Ji

‖∇Ji (Y 0)‖2 ∇Ji (Y 0)

• Newton-inspired when limJi 6= 0 :

ui =max

(J(k−1)

i − J(k)i ,δ

)‖∇Ji (Y 0)‖2 ∇Ji (Y 0) , k : iteration no. , δ > 0 (small)

24 / 102

Page 25: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Is standard normalizationsufficient to guarantee that

ω =−−→OO⊥?

(yielding equal Fréchet derivatives)

No, if n > 2 :

A2

u1

u2 u3O⊥

−−→OO⊥ /∈ U

0

Unit Sphere

25 / 102

Page 26: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Experimenting MGDAFrom Adrien ZERBINATI, on-going doctoral thesis

Basic testcase : minimize the functions

f (x ,y) = 4x2 + y2 + xy g(x ,y) = (x−1)2 + 3(y−1)2

26 / 102

Page 27: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Basic testcaseConvergence from different initial design points

DESIGN SPACE FUNCTION SPACE

(0.5,2)

(1.5,2.5)

(-1.5,-2.5)

27 / 102

Page 28: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Fonseca testcase

Minimize 2 fonctions of 3 variables :

f1,2(x1,x2,x3) = 1−exp

(−

3

∑i=1

(xi ±

1√3

)2)

28 / 102

Page 29: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Fonseca testcase (cont’d)Convergence from initial design points over a sphere

DESIGN SPACE FUNCTION SPACE

Continuous but nonconvex front

29 / 102

Page 30: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Fonseca testcase (end)Compare MGDA with Pareto Archived Evolution Strategy

(PAES)

DESIGN SPACE FUNCTION SPACE

−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 −1

0

1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

x2

x1

PAES 2x50 & MGDA 6x12

x3

MGDA iterate

PAES generate

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1PAES 2x50 & MGDA 6x12

MGDA iterate

PAES generate

30 / 102

Page 31: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

31 / 102

Page 32: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Wing-shape optimizationWave-drag minimization in conjunction with

lift maximization

Metamodel-assisted MGDA (Adrien Zerbinati)

• Wing-shape geometryextruded from airfoil geometry (initial airfoil : NACA0012), 10 parameters

• Two-point/two-objective optimization

1 Subsonic Eulerian flow : M∞ = 0.3, α = 8o , for CL maximization2 Transonic Eulerian flow : M∞ = 0.83, α = 2o , for CD

minimization

• Algorithm - Initial database of 40 design points evolves as follows :

1 Evaluate each new point by two 3D Eulerian flow computations2 Construct surrogate models for CL and CD (using the entire

dataset), and perform MGDA to convergence3 Enrich the database with new points, if appropriate

After 6 loops, the database is made of 227 design points(554 flow computations)

32 / 102

Page 33: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Wing-shape optimization2

Convergence of dataset

6 loops

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045−0.85

−0.8

−0.75

−0.7

−0.65

−0.6

−0.55

−0.5

−0.45

drag coefficient

−lif

t coeffic

ient

initial database

1st step

2nd step

3rd step

4th step

5th step

6th step

DRAG (transonic)

-LIFT(subsonic)

33 / 102

Page 34: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Wing-shape optimization3

Low-drag quasi-Pareto-optimal shape

SUBSONIC TRANSONIC

M∞ = 0.3, α = 8o M∞ = 0.83, α = 2o

34 / 102

Page 35: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Wing-shape optimization4

High-lift quasi-Pareto-optimal shape

SUBSONIC TRANSONIC

M∞ = 0.3, α = 8o M∞ = 0.83, α = 2o

35 / 102

Page 36: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Wing-shape optimization5

Intermediate quasi-Pareto-optimal shape

SUBSONIC TRANSONIC

M∞ = 0.3, α = 8o M∞ = 0.83, α = 2o

36 / 102

Page 37: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

37 / 102

Page 38: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Description of the test case :

In-house CFD Solver(Num3sis) :

• Compressible Navier-Stokes(FV/FE)

• Parallel computing (MPI)

• CPU cost : 4h on 32 cores

• CAD and mesh by GMSH

Test case :• Mach number : 0.1

• Reynolds number : Re ≈ 2800

• Two-criterion shapeoptimization

• Fine mesh (some 700,000points)

38 / 102

Page 39: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

First objective function

Velocity Variance

• Umean =1

Vtot∑

d(cel,Po )≤ε

u(cel)∗Vol(cel)

• σ2vel,x =

1Vtot

∑d(cel,Po )≤ε

(Umean,x −ux (cel))2 ∗Vol(cel)

• σ2 = σ2vel,x + σ2

vel,y + σ2vel,z

39 / 102

Page 40: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Second objective function

Pressure loss• pk = pmean for inlet (k = i), or outlet (k = o)

• uk = ‖u‖mean for in- or out-let

• ∆p = pi −po +ρi u2

i

2− ρou2

o

2

40 / 102

Page 41: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Design parameters

41 / 102

Page 42: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA on surrogate model

42 / 102

Page 43: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Software framework

43 / 102

Page 44: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Identifying the Pareto set

3

21

Nominal PointInitial database3rd step6th step9th stepInitial Pareto frontFinal Pareto front

Pres

sure

loss

0,5

1

1,5

2

2,5

Velocity variance0,6 0,8 1 1,2 1,4

Figure: MGDA metamodel assisted step by step evolution. Initial and finalnon-dominated sets. 223 solver calls.

44 / 102

Page 45: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Velocity Magnitude Streamlines

1: BEST VELOCITY VARIANCE 2 : NOMINAL SHAPE

SHAPE COMPARISON 3 : BEST PRESSURE LOSS

45 / 102

Page 46: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Velocity magnitude in a sectionjust after the bend :

1: BEST VELOCITY VARIANCE 2 : NOMINAL SHAPE

SECTION POSITION 3 : BEST PRESSURE LOSS

46 / 102

Page 47: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Velocity Magnitude in outputsection :

1 : BEST VELOCITY VARIANCE 2 : NOMINAL SHAPE

SECTION POSITION 3 : BEST PRESSURE LOSS

47 / 102

Page 48: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

NOMINAL versus LOWESTVELOCITY VARIANCE

48 / 102

Page 49: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

NOMINAL versus LOWESTPRESSURE LOSS

49 / 102

Page 50: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

LOWEST VELOCITYVARIANCE versus LOWEST

PRESSURE LOSS

50 / 102

Page 51: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

51 / 102

Page 52: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Dirichlet problem

Discrete solution by standard 2nd-order finite-differencemethod over 40×40 qradrangular mesh

−∆u = f over Ω = [−1,1]× [−1,1]

u = 0 (Γ = ∂Ω)

-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 -1-0.8

-0.6-0.4

-0.2 0

0.2 0.4

0.6 0.8

1

-0.5

0

0.5

1

1.5

u_h

u_h

’fort.20’ 1.3 1.1 0.9 0.7 0.5 0.3 0.1 0

-0.1 -0.2 -0.3 -0.4 -0.5

x

y

u_h

u_h contour lines 1.3 1.1 0.9 0.7 0.5 0.3 0.1 0

-0.1 -0.2 -0.3 -0.4 -0.5

-1 -0.5 0 0.5 1

x

-1

-0.5

0

0.5

1

y

52 / 102

Page 53: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Domain partitioningFunction value controls at 4 interfaces

yielding 4 sub-domain Dirichlet problems with 2controlled interfaces each

53 / 102

Page 54: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Interface jumps

Formal expression

• over γ1 (0≤ x ≤ 1 ; y = 0):

s1(x) =∂u∂y

(x ,0+)− ∂u∂y

(x ,0−) =

[∂u1

∂y− ∂u4

∂y

](x ,0);

• over γ2 (x = 0 ; 0≤ y ≤ 1):

s2(y) =∂u∂x

(0+,y)− ∂u∂x

(0−,y) =

[∂u1

∂x− ∂u2

∂x

](0,y);

• over γ3 (−1≤ x ≤ 0 ; y = 0):

s3(x) =∂u∂y

(x ,0+)− ∂u∂y

(x ,0−) =

[∂u2

∂y− ∂u3

∂y

](x ,0);

• over γ4 (x = 0 ; −1≤ y ≤ 0):

s4(y) =∂u∂x

(0+,y)− ∂u∂x

(0−,y) =

[∂u4

∂x− ∂u3

∂x

](0,y).

Approximationby 2nd-order one-sided finite-differences

54 / 102

Page 55: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Functionals and matchingcondition

Interface functionals

Ji =∫

γi

12 s2

i w dγi := Ji (v)

that is, explicitly:

J1 =∫ 1

0

12 s1(x)2 w(x)dx J2 =

∫ 1

0

12 s2(y)2 w(y)dy

J3 =∫ 0

−1

12 s3(x)2 w(x)dx J4 =

∫ 0

−1

12 s4(y)2 w(y)dy

(w(t) (t ∈ [0,1]) is an optional weighting function, andw(−t) = w(t).)

Matching condition

J1 = J2 = J3 = J4 = 0

55 / 102

Page 56: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Adjoint problems

Eight Dirichlet sub-problems

∆pi = 0

pi = 0

pi = siw

(Ωi )

(∂Ωi\γi )

(γi )

∆qi = 0

qi = 0

qi = si+1w

(Ωi )

(∂Ωi\γi+1)

(γi+1)

Green’s formula∫γi

si u′i nw =∫

γi

pi n v ′i +∫

γi+1

pi n v ′i+1

and ∫γi+1

si+1 u′i nw =∫

γi

qi n v ′i +∫

γi+1

qi n v ′i+1

56 / 102

Page 57: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Functional gradients

J ′1 =∫ 1

0s1(x)s′1(x)w(x)dx =

∫ 1

0s1(x)

[∂u′1∂y− ∂u′4

∂y

](x ,0)w(x)dx

=∫ 1

0

∂(p1−q4)

∂y(x ,0)v ′1(x)dx +

∫ 1

0

∂p1

∂x(0,y)v ′2(y)dy +

∫ 0

−1

∂q4

∂x(0,y)v ′4(y)dy

J ′2 =∫ 1

0s2(y)s′2(y)w(y)dy =

∫ 1

0s2(y)

[∂u′1∂x− ∂u′2

∂x

](0,y)w(y)dy

=∫ 1

0

∂q1

∂y(x ,0)v ′1(x)dx +

∫ 1

0

∂(q1−p2)

∂x(0,y)v ′2(y)dy +

∫ 0

−1

∂p2

∂y(x ,0)v ′3(x)dx

J ′3 =∫ 0

−1s3(x)s′3(x)w(x)dx =

∫ 0

−1s3(x)

[∂u′2∂y− ∂u′3

∂y

](x ,0)w(x)dx

=−∫ 1

0

∂q2

∂x(0,y)v ′2(y)dy +

∫ 0

−1

∂(q2−p3)

∂y(x ,0)v ′3(x)dx−

∫ 0

−1

∂p3

∂x(0,y)v ′4(y)dy

J ′4 =∫ 0

−1s4(y)s′4(y)w(y)dy =

∫ 0

−1s4(y)

[∂u′4∂x− ∂u′3

∂x

](0,y)w(y)dy

=−∫ 1

0

∂p4

∂y(x ,0)v ′1(x)dx−

∫ 0

−1

∂q3

∂y(x ,0)v ′3(x)dx +

∫ 0

−1

∂(p4−q3)

∂x(0,y)v ′4(y)dy

Conclusion:

J ′i =4

∑j=1

∫γj

Gi,j v ′j dγj (i = 1, ...,4) ; Gi,j : partial gradients.

57 / 102

Page 58: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Other technical details

Dirichlet sub-problemsAll 12 (4 direct, 4×2 adjoint) sub-problems are solved by directinverse (discrete sine-transform)

uh = (ΩX ⊗ΩY ) (ΛX ⊕ΛY )−1 (ΩX ⊗ΩY ) fh

Integralsare approximated by the trapezoidal rule

58 / 102

Page 59: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Reference methodQuasi-Newton Method applied to agglomerated criterion

J =4

∑i=1

Ji

Gradient

∇J =4

∑i=1

∇Ji

Iterationv (`+1) = v (`)−ρ`∇J(`)

Stepsize

δJ = ∇J(`).δv (`) =−ρ`

∥∥∥∇J(`)∥∥∥2

is set to −εJ(`) by fixing

ρ` =εJ(`)∥∥∇J(`)

∥∥2

59 / 102

Page 60: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Quasi-Newton steepest-descentε = 1

Convergence history

Asymptotic Global

1e-06

1e-05

0.0001

0.001

0.01

0.1

0 2 4 6 8 10 12 14 16 18 20

J_1J_2J_3J_4

J = SUM TOTAL

1e-09

1e-08

1e-07

1e-06

1e-05

0.0001

0.001

0.01

0.1

1

10

100

0 20 40 60 80 100 120 140 160 180 200

J_1J_2J_3J_4

J = SUM TOTAL

60 / 102

Page 61: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Quasi-Newton steepest descentε = 1

History of gradients over 200 iterations

∂J/∂v1 ∂J/∂v2

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

0 2 4 6 8 10 12 14 16 18 20

DISCRETIZED FUNCTIONAL GRADIENT OF J = SUM_i J_i W.R.T. V1

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

0 2 4 6 8 10 12 14 16 18 20

DISCRETIZED FUNCTIONAL GRADIENT OF J = SUM_i J_i W.R.T. V2

61 / 102

Page 62: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Quasi-Newton steepest descentε = 1

History of gradients over 200 iterations

∂J/∂v3 ∂J/∂v4

-0.6

-0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0 2 4 6 8 10 12 14 16 18 20

DISCRETIZED FUNCTIONAL GRADIENT OF J = SUM_i J_i W.R.T. V3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

0 2 4 6 8 10 12 14 16 18 20

DISCRETIZED FUNCTIONAL GRADIENT OF J = SUM_i J_i W.R.T. V4

62 / 102

Page 63: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Quasi-Newton steepest descentε = 1

Discrete solution

-1-0.5

0 0.5

1 -1

-0.5

0

0.5

1

-0.5

0

0.5

1

1.5

u_h

u_h

u_h(x,y) 1.3 1.1 0.9 0.7 0.5 0.3 0.1 0

-0.1 -0.2 -0.3 -0.4 -0.5

x

y

u_h

u_h contour lines 1.3 1.1 0.9 0.7 0.5 0.3 0.1 0

-0.1 -0.2 -0.3 -0.4 -0.5

-1 -0.5 0 0.5 1

x

-1

-0.5

0

0.5

1

y

63 / 102

Page 64: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Basic MGDA

Practical determination of minimum-norm element ω

ω =4

∑i=1

αiui

ui = ∇Ji (Y 0)

using the following parameterization of the convex hull:α1 = c1c2c3

α2 = (1− c1)c2c3

α3 = (1− c2)c3

α4 = (1− c3)

and(c1,c2,c3) ∈ [0,1]3

are discretized by step of 0.01. Best set of coefficients retained.

64 / 102

Page 65: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Basic MGDAStepsize

Iteration

v (`+1) = v (`)−ρ`ω(`)

Stepsize

δJ = ∇J(`).δv (`) =−ρ` ∇J(`).ω

is set to −εJ(`) by fixing

ρ` =εJ(`)

∇J(`).ω

In practice: ε = 1.

65 / 102

Page 66: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Basic MGDAε = 1

Asymptotic Convergence

Unscaled Scaledui = ∇Ji ui = ∇Ji

Ji

66 / 102

Page 67: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Basic MGDAε = 1

Global Convergence

Unscaled Scaledui = ∇Ji ui = ∇Ji

Ji

67 / 102

Page 68: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

First conclusions

Basic MGDA• Scaling essential

• Somewhat deceiving convergence

Who is to blame?• the insufficiently accurate determination of ω;

• the non-optimality of the scaling of gradients;

• the non-optimality of the step-size, the parameter ε beingmaintained equal to 1 throughout;

• the large dimension of the design space, here 76 (4 interfacesassociated with 19 d.o.f.’s).

68 / 102

Page 69: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-II : a direct algorithmValid when gradients are linearly independent

User-supplied scaling factors Si(i=1,,n), and scaledgradients

J ′i =∇Ji (Y 0)

Si

(Si > 0; e.g. Si = Ji for logarithmic gradients).

Perform Gram-Schmidt with special normalization

• Set u1 = J ′1

• For i = 2, . . . ,n, set: ui =J ′i −∑k<i ci,k uk

Aiwhere:

∀k < i : ci,k =

(J ′i ,uk

)(uk ,uk

) , and:

Ai =

1−∑k<i

ci,k if nonzero

εi otherwise (εi arbitrary, but small)69 / 102

Page 70: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Consequences

Element ω always in the interior of convex hull

ω =n

∑i=1

αiui αi =1

‖ui‖2∑

nj=1

1

‖uj‖2

=1

1 + ∑j 6=i‖ui‖2

‖uj‖2

< 1

This implies equal projections of ω onto uk

∀k :(uk ,ω

)= αk ‖uk‖2 = ‖ω‖2

and finally: (J ′i ,ω

)= ‖ω‖2 (∀i)

(with a possibly-modified scale S′i = (1 + εi )Si ), that is:the same positive constant.

EQUAL FRÉCHET DERIVATIVESFORMED WITH SCALED GRADIENTS

70 / 102

Page 71: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Geometrical demonstration

u1

u2O

u3

O⊥

u ∈ U

Given J ′i = ∇Ji (Y 0)/Si , perform G.-S. process

to compute ui = [J ′i −∑k<i ci,k uk ]/Ai ; then:

O⊥ ∈ U =⇒ ω =−−→OO⊥ =⇒(

u1,ω)

=(u2,ω

)=(u3,ω

)= ‖ω‖2

J ′i = Aiui + ∑k<i ci,k uk =⇒(J ′i ,ω

)=(Ai + ∑

k<ici,k)

︸ ︷︷ ︸= 1 by

normalization

thru Ai

‖ω‖2 = ‖ω‖2 = const.

71 / 102

Page 72: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-II b: automatic rescaleDo not let constant Ai become negative; instead define

Ai = 1−∑k<i

ci,k

only if this number is strictly-positive. Otherwise, use modified scale:

S′i =

(∑k<i

ci,k

)Si

(“automatic rescale”), so that:

c′i,k =

(∑k<i

ci,k

)−1

ci,k , ∑k<i

c′i,k = 1 ,

and set Ai = εi , for some small εi .

Same formal conclusion: the Fréchet derivatives are equal; but the

value is much larger, and (at least) one criterion has been rescaled

72 / 102

Page 73: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Some open questions

• ωII = ω?

• Yes, if n = 2 and angle obtuse

• No, in general.

• Is the following implication

limωII = 0 =⇒ limω = 0

true? (assuming the limit is the result of the MGDA iteration)

In which order should we perform the Gram-Schmidtprocess?

• Etc

73 / 102

Page 74: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-IIε = 1

Global convergence

Unscaled Scaledui = ∇Ji ui = ∇Ji

Ji

74 / 102

Page 75: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-II bε = 1

Global convergence

Unscaled Scaledui = ∇Ji ui = ∇Ji

Ji

automatic rescale on automatic rescale on

75 / 102

Page 76: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-II : Recap

Convergence over 500 iterations

Unscaled Scaledui = ∇Ji ui = ∇Ji

Ji

automatic rescale off automatic rescale on

76 / 102

Page 77: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

77 / 102

Page 78: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-IIIBird of paradise (Strelitzia reginae)

78 / 102

Page 79: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-IIIBird of paradise (Strelitzia reginae)

79 / 102

Page 80: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-IIIBird of paradise (Strelitzia reginae)

u1

u2

ω

80 / 102

Page 81: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-IIIBird of paradise (Strelitzia reginae)

ω

81 / 102

Page 82: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-III - definition1

Start from scaled gradients, J ′i , and duplicates, gi,J ′i =

∇Ji (Y 0)

Si(Si : user-supplied scale)

and proceed in 3 steps: A,B and C.

A: Initialization

• Set

u1 := g1 = J ′k / k = Argmaxi minj

(J ′j ,J

′i

)(J ′i ,J

′i

)(justified afterwards)

• Set n×n lower-triangular array c = ci,j (i ≥ j) to 0.

• Set, conservatively, I := n(expected number of computed orthogonal basis vectors).

• Assign some appropriate value to a cut-off constant a :0≤ a < 1.

(Note: main diagonal of array c is to contain cumulative row-sums.)82 / 102

Page 83: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-III - definition2

B: Main G.-S. loop; for i = 2,3, . . . , (at most) n, do:

1. Calculate the i−1st column of coefficients:

cj,i−1 =

(gj ,ui−1

)(ui−1,ui−1

) (∀j = i, . . . ,n)

and update the cumulative row-sums:

cj,j := cj,j + cj,j−1 = ∑k<i

cj,k (∀j = i, . . . ,n)

2. Test:• If the following condition is satisfied

cj,j > a (∀j = i, . . . ,n)

set I := i−1, and interrupt the Gram-Schmidt process(go to 3.).

• Otherwise, compute next orthogonal vector ui as follows:

83 / 102

Page 84: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-III - definition3

- Identify index ` = Argminj cj,j / i ≤ j ≤ n(Note that c`,` ≤ a < 1.)

- Permute information associated with i and ` :g-vectors gi g`, rows i and ` of array c and cumulative row-sums ci,i c`,`.

- Set Ai = 1− ci,i ≥ 1−a > 0 (ci,i = former-c`,` ≤ a), and calculate

ui =gi −∑k<i ci,k uk

Ai

(in which gi = former-g`, ci,k = former-c`,k ).

- If ui 6= 0, return to 1. with incremented i; otherwise:

gi = ∑k<i

ci,k uk = ∑k<i

c′i,k gk

where the c′i,k are calculated by backward substitution.Then, if c′i,k ≤ 0 (∀k < i):

Pareto-stationarity detected: STOP MGDA iteration;otherwise (ambiguous exceptional case):

STOP G.-S. process; compute original ω and go to C.

84 / 102

Page 85: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-III - definition4

3. Calculate ω as the minimum-norm element in the convex hullof u1,u2, . . . ,uI:

ω =I

∑i=1

αiui 6= 0

αi =1

‖ui‖2∑

Ij=1

1

‖uj‖2

=1

1 + ∑j 6=i‖ui‖2

‖uj‖2

(Note that here all computed ui 6= 0, and 0 < αi < 1.)

C: If ‖ω‖< TOL, STOP MGDA iteration; otherwise,perform descent step and return to B.

85 / 102

Page 86: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

MGDA-III - consequencesIf I = n :MGDA-III = MGDA-II b + automatic ordering in Gram-Schmidtprocess making ’rescale’ unnecessary since, by construction

∀i : Ai ≥ 1−a > 0

Otherwise, I < n (incomplete Gram-Schmidt process):

• First I Fréchet derivatives:(gi ,ω

)=(ui ,ω

)= ‖ω‖2 > 0 (∀i = 1, . . . , I)

• Subsequent ones (i > I):

ω =I

∑j=1

αjuj gi =I

∑k=1

ci,k uk + vi vi ⊥ u1,u2, . . . ,uI

⇓(gi ,ω

)=

I

∑k=1

ci,k(uk ,ω

)=

I

∑k=1

ci,k ‖ω‖2 = ci,i ‖ω‖2 > a‖ω‖2 > 086 / 102

Page 87: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Choice of g1Justification

At initialization, we have set

u1 := g1 = J ′k / k = Argmaxi minj

(J ′j ,J

′i

)(J ′i ,J

′i

)This was equivalent to maximizing c2,1 = c2,2, that is,maximizing the least cumulative row-sum, at firstestimation.

(Recall that the favorable situation is when all cumulative row-sums are positive, or > a.)

87 / 102

Page 88: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Expected benefits

Automatic ordering and rescale

When gradients exhibit a general trend, ω is found infewer steps, and has a larger norm

Larger ‖ω‖ implies larger directional derivatives, andgreater efficiency of descent step(

gi ,ω

‖ω‖)

= ‖ω‖ or a‖ω‖

88 / 102

Page 89: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

89 / 102

Page 90: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Adressing the question ofscaling

When Hessians are known• For objective function Ji (Y ) alone, Newton’s iteration writes

Y 1 = Y 0−pi Hipi = ∇Ji (Y 0)

• Split pi into orthogonal components

pi = qi + ri

where: qi =

(pi ,∇Ji (Y 0)

)‖∇Ji (Y 0)‖2 ∇Ji (Y 0) and ri ⊥ ∇Ji (Y 0)

• Define scaled gradientJ ′i = qi

• Apply MGDA-III

Equivalently, set scaling factor as follows: Si =‖∇Ji (Y 0)‖2(pi ,∇Ji (Y 0)

)90 / 102

Page 91: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

VariantMGDA-IV b

Use BFGS iterative estimates in place of exact Hessians

k : MGDA iteration index

(∀i = 1, . . . ,n) H(0)i = Id

H(k+1)i = H(k)

i −1

s(k)T H(k)i s(k)

H(k)i s(k)s(k)T

H(k)T

i +1

z(k)T

i s(k)z(k)

i z(k)T

i

in which:

s(k) = Y (k+1)−Y (k)

z(k)i = ∇Ji (Y (k+1))−∇Ji (Y (k))

91 / 102

Page 92: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Recommended Step-Size

Ji (Y 0−ρω) = Ji (Y 0)−ρ(∇Ji (Y 0),ω

)+ 1

2 ρ2(Hiω,ω

)+ . . . .

|δJi | := Ji (Y 0)− Ji (Y 0−ρω) := Aiρ− 12 Biρ

2 ,

where: Ai = Siai , Bi = Sibi , and:

ai =(J ′i ,ω

)bi =

(Hiω,ω

)/Si .

If the scales Si are truly relevant for the objective-functions

ρ? = Argmax

ρmin

iδi .

where:

δi =|δJi |Si

= aiρ− 12 biρ

2 .

92 / 102

Page 93: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Recommended Step-Size2

I = n

ρ? = ρ

?I :=

‖ω‖2

bI,

where bI = maxi≤I bi is assumed to be positive.

I < n

ρ?II :=

a‖ω‖2

bII,

where bII = maxi>I bi is assumed to be positive.Also

ρ× =2(1−a)‖ω‖2

bI−bII

at which point the two bounding parabolas intersect:

‖ω‖2ρ− 1

2 bIρ2 = a‖ω‖2

ρ− 12 bIIρ

2 .

93 / 102

Page 94: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Recommended Step-Size3

a) I = n

||ω||2δ

ρρ?I

b) I < n, ρ× > max(ρ?I ,ρ

?II )

||ω||2

a||ω||2

δ

ρρ?I ρ?

II ρ×

c) I < n, ρ× ∈ (ρ?I ,ρ

?II )

||ω||2

a||ω||2

δ

ρρ?I ρ× ρ?

II

d) I < n, ρ× < min(ρ?I ,ρ

?II )

||ω||2

δ

ρρ?I

ρ×

ρ?II

94 / 102

Page 95: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Recommended Step-Sizeend

If I = n, ρ? = ρ?I .

If I < n, ρ? is the element of the triplet ρ?I ,ρ

?II ,ρ×

which separates the other two.

95 / 102

Page 96: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

96 / 102

Page 97: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

CONCLUSIONSTheory

Multiple-Gradient Descent Algorithm

• Notion of Pareto stationarity introduced to characterize"standard Pareto-optimal points"

• Basic MGDA : permits to identify a descent direction commonto arbitrary number of criteria, knowing the local gradients innumber at most equal to the number of design parameters

• Proof of convergence of MGDA to Pareto-stationary points

• Capability to identify the Pareto front demonstrated formathematical test-cases; non-convexity of Pareto front not aproblem

• MGDA-II : a direct procedure, faster and more accurate, whengradients are linearly independent; variant MGDA-II b (withautomatic rescale) recommended

• On-going research: scaling, preconditioning, step-sizeadjustment, what to do to extend MGDA-II to cases oflinearly-dependent gradients, . . .

97 / 102

Page 98: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

CONCLUSIONSApplications

Implementation and first experiments

• Validation with analytical test cases and comparison with aPareto capturing method (PAES)

• Extension to Meta-Model-Assisted MGDA

Aircraft-Wing Aerodynamics and AutomobileCooling-System Design

• Successful design experiments with 3D compressible flow(Euler and Navier-Stokes)

• On-going development of meta-model-assisted MGDA andtesting of 3D external aerodynamics case to be presented atECCOMAS 2012, Vienna, Austria)

98 / 102

Page 99: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

CONCLUSIONSApplications

Domain-partitioning model problemThe method works, but it is not as cost-efficient as the standardquasi-Newton method applied to the agglomerated criterion; but thetest-case was peculiar in several ways:

• Pareto front and set reduced to a singleton

• large dimension of the design space (4 criteria, 76 design variables)

• robust procedure to adjust the step-size needed

• determination of ω in the basic method should be made more accurately, perhaps usingiterative refinement

• scaling of gradients found important but never really analyzed completely (logarithmicgradients appropriate for vanishing criteria)

• general preconditioning not yet clear (on-going)

The MGDA-II direct variant found faster and more robust; withautomatic rescaling procedure on, MGDA-II b seems to exhibitquadratic convergence.

99 / 102

Page 100: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Outline1 Foreword2 MGDA

Pareto-stationarityDescent directionMain resultsPracticalitiesMathematical test-cases

3 Aircraft-Wing Aerodynamics4 Automobile Cooling System Design

Problem description and numerical toolsSoftware frameworkNumerical results

5 Domain-Partitioning Model ProblemModel problemQuasi-Newton MethodBasic MGDAMGDA-II

6 MGDA-IIIBird of paradiseAlgorithmProperties

7 MGDA-IV8 Conclusions9 Cooperation and Competition

100 / 102

Page 101: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Flirting with the Pareto frontMGDA is combined with a locally-adapted Nash game

based on a territory splitting that preserves tosecond-order the Pareto-stationarity condition

Cooperation AND Competition

• COOPERATION (MGDA) :The design point is steered to the Pareto front.

• COMPETITION (Nash game) : at start (from the Pareto front) :α1∇J1 +α2∇J2 = 0; then let : JA = α1J1 +α2J2, et JB = J2, andadapt territory-splitting to best preserve JA; then :the trajectory remains tangent to the Pareto front.

101 / 102

Page 102: MULTIPLE GRADIENT DESCENT ALGORITHM (MGDA ...Engineering, Vienna (Austria), September 10-14, 2012 TO16: Computational inverse problems and optimization 1/102 MGDA applied to compressible

MGDA applied tocompressible aerodynamics

J.-A. Désidéri

Foreword

MGDA

Pareto-stationarity

Descent direction

Main results

Practicalities

Mathematical test-cases

Aircraft-Wing Aerodynamics

Automobile Cooling SystemDesign

Problem description and numericaltools

Software framework

Numerical results

Domain-Partitioning ModelProblem

Model problem

Quasi-Newton Method

Basic MGDA

MGDA-II

MGDA-III

Bird of paradise

Algorithm

Properties

MGDA-IV

Conclusions

Cooperation and Competition

Some references

Reports from INRIA Open Archive:

• M. G. D. A., INRIA Research Report No. 6953 (2009),Revised version, October 2012(Open Archive : http://hal.inria.fr/inria-00389811)

• Related INRIA Research Reports:Nos. 7667 (2011), 7922 (2012), 7968 (2012), 8068 (2012)

Other• C. R. Acad. Sci. Paris, Ser. I 350 (2012) 313-318

102 / 102