06061.debkalyanmoy.slides1
TRANSCRIPT
Dagstuhl Seminar, 5-10 February, 2006
1
Functional Decomposition of NSGA-II and Functional Decomposition of NSGA-II and Various Problem-Solving StrategiesVarious Problem-Solving Strategies
Kalyanmoy DebKalyanmoy DebProfessor of Mechanical EngineeringProfessor of Mechanical EngineeringIndian Institute of Technology KanpurIndian Institute of Technology KanpurDirector, Kanpur Genetic Algorithms Director, Kanpur Genetic Algorithms
Laboratory (KanGAL)Laboratory (KanGAL)Email: [email protected]: [email protected]
http://www.iitk.ac.in/kangal/deb.htmhttp://www.iitk.ac.in/kangal/deb.htm
Dagstuhl Seminar, 5-10 February, 2006
2
Overview Essentials of multi-objective
optimization NSGA-II platform Different multi-objective problem-
solving tasks Omni-optimizer
Degeneracy to various single and multi-objective tasks
Conclusions
Dagstuhl Seminar, 5-10 February, 2006
3
Multi-Objective Optimization:Multi-Objective Optimization:Handling multiple conflicting objectivesHandling multiple conflicting objectives
We often face themWe often face them
Dagstuhl Seminar, 5-10 February, 2006
4
Which Solutions are Optimal?Which Solutions are Optimal? Relates to the concept Relates to the concept
of dominationof domination xx(1)(1) dominates x dominates x(2)(2), if , if
xx(1)(1) is no worse than x is no worse than x(2)(2) in in all objectivesall objectives
xx(1)(1) is strictly better than is strictly better than xx(2)(2) in at least one in at least one objectiveobjective
Examples: Examples: 3 dominates 23 dominates 2 3 does not dominate 53 does not dominate 5
Dagstuhl Seminar, 5-10 February, 2006
5
Pareto-Optimal SolutionsPareto-Optimal Solutions P’=Non-dominated(P) P’=Non-dominated(P)
Solutions which are Solutions which are not dominated by any not dominated by any
member of the set Pmember of the set P O(N log N) O(N log N) algorithms existalgorithms exist Pareto-Optimal set Pareto-Optimal set = Non-dominated(S)= Non-dominated(S) A number of solutions A number of solutions
are optimalare optimal
Dagstuhl Seminar, 5-10 February, 2006
6
Pareto-Optimal Fronts Depends on
the type of objectives
Definition of domination takes care of possibilities
Always on the boundary of feasible region
Dagstuhl Seminar, 5-10 February, 2006
7
Local Versus Global Pareto-Optimal Fronts Local Pareto-optimal Front: Domination check is
restricted within a neighborhood (in decision space) of P
Dagstuhl Seminar, 5-10 February, 2006
8
Some Terminologies Ideal point (z*)
nonexistent, lower bound on Pareto-optimal set
Utopian point (z**) nonexistent
Nadir point (znad) Upper bound on
Pareto-optimal set Normalization:
Dagstuhl Seminar, 5-10 February, 2006
9
Differences with Single-Objective Optimization One optimum versus multiple optima Requires search and decision-making Two spaces of interest, instead of one
Dagstuhl Seminar, 5-10 February, 2006
10
Ideal Multi-Objective Ideal Multi-Objective OptimizationOptimizationStep 1 :Step 1 : Find a set of Pareto-Find a set of Pareto-optimal solutionsoptimal solutions
Step 2Step 2 :: Choose one from Choose one from the setthe set
Dagstuhl Seminar, 5-10 February, 2006
11
Two Goals in Ideal Multi-Objective Two Goals in Ideal Multi-Objective OptimizationOptimization
Converge to the Converge to the Pareto-optimal frontPareto-optimal front
Maintain as diverse a Maintain as diverse a distribution as distribution as possiblepossible
Dagstuhl Seminar, 5-10 February, 2006
12
Elitist Non-dominated Sorting Elitist Non-dominated Sorting Genetic Algorithm (NSGA-II)Genetic Algorithm (NSGA-II)
NSGA-II can NSGA-II can extract Pareto-extract Pareto-optimal frontieroptimal frontierAlso find a well-Also find a well-distributed set of distributed set of solutionssolutionsiSIGHT and iSIGHT and modeFrontier modeFrontier adopted NSGA-IIadopted NSGA-II
Fast-Breaking Paper in Engineering by ISI Web of Science (Feb’04)Fast-Breaking Paper in Engineering by ISI Web of Science (Feb’04)
Dagstuhl Seminar, 5-10 February, 2006
13
Functional Decomposition Convergence:
Emphasize non-dominated solutions
Diversity: Prefer less-crowded solutions
Elite-preservation For ensuring convergence
properties
Dagstuhl Seminar, 5-10 February, 2006
14
An Iteration of NSGA-IIAn Iteration of NSGA-IIEl
ite-p
rese
rvat
ion
Elite
-pre
serv
atio
n
ConvergenceConvergence Diversity-maintenanceDiversity-maintenance
Dagstuhl Seminar, 5-10 February, 2006
15
NSGA-II: Crowding DistanceNSGA-II: Crowding Distance
Overall Complexity Overall Complexity O(O(N logN logM-1M-1NN))
Diversity is maintainedDiversity is maintained
Improve diversity byImprove diversity by• k-mean clusteringk-mean clustering• Euclidean distance Euclidean distance measuremeasure• Other techniquesOther techniques
Dagstuhl Seminar, 5-10 February, 2006
16
Simulation on ZDT1Simulation on ZDT1
Dagstuhl Seminar, 5-10 February, 2006
17
Simulation on ZDT3Simulation on ZDT3
Dagstuhl Seminar, 5-10 February, 2006
18
Changing Dominance Relation Alter the meaning of Pareto-optimal
points Constrained optimization (Fonseca and
Fleming, 1996, Deb et al., 2000) Cone dominance (guided dominance,
Branke et al., 2000) Distributed EMO (Deb et al., 2003) Epsilon-MOEA (Laumanns et al., 2003;
Deb et al., 2005) Robust and reliability-based EMO (Deb
and Gupta, 2005)
Dagstuhl Seminar, 5-10 February, 2006
19
Constraint-Domination PrincipleConstraint-Domination Principle
1.1. ii is feasible and is feasible and jj is not is not2.2. ii and and jj are both are both
infeasible, but infeasible, but ii has a has a smaller overall smaller overall constraint violationconstraint violation
3.3. ii and and jj are feasible and are feasible and ii dominates dominates jj
A solution A solution ii constraint-constraint-dominatesdominates a solution a solution jj, if any , if any is true:is true:
Dagstuhl Seminar, 5-10 February, 2006
20
Constrained NSGA-II Simulation Constrained NSGA-II Simulation ResultsResults
22 )( xxf MinimizeMinimize 11 )( xxf 1
22
1)(xxxf
1969
12
12
xxxx WhereWhere
5.05.05.0
0tan16cos1011
22
21
2
1122
21
xx
xxxx
MinimizeMinimize11 )( xxf
WhereWhere
Dagstuhl Seminar, 5-10 February, 2006
21
Simulation on TNKSimulation on TNK
Dagstuhl Seminar, 5-10 February, 2006
22
Simulation on CTP5Simulation on CTP5
Dagstuhl Seminar, 5-10 February, 2006
23
Cone DominanceCone DominanceUsing a DM’s Using a DM’s preference (not a preference (not a solution but a region)solution but a region)
Guided domination Guided domination principle: Biased principle: Biased niching approachniching approach
Weighted domination Weighted domination approachapproach
Dagstuhl Seminar, 5-10 February, 2006
24
Distributed Computing of Distributed Computing of Pareto-Optimal SetPareto-Optimal Set
Guided domination concept to search different parts Guided domination concept to search different parts of Pareto-optimal region of Pareto-optimal region Distributed computing of different partsDistributed computing of different parts
Dagstuhl Seminar, 5-10 February, 2006
25
Distributed computing: Distributed computing: A Three-Objective ProblemA Three-Objective Problem
Spatial computing, not temporalSpatial computing, not temporal
TheoryTheory NSGA-II SimulationsNSGA-II Simulations
Dagstuhl Seminar, 5-10 February, 2006
26
εε-MOEA: Using -MOEA: Using εε-Dominance-Dominance
EA and archive EA and archive populations evolve populations evolve One EA and one One EA and one archive member are archive member are mated mated Archive update using Archive update using εε-dominance -dominance EA update using EA update using usual dominanceusual dominance
Dagstuhl Seminar, 5-10 February, 2006
27
Comparative Study on Comparative Study on Three-Objective DTLZ ProblemsThree-Objective DTLZ Problems
Dagstuhl Seminar, 5-10 February, 2006
28
Test Problem DTLZ2Test Problem DTLZ2
Dagstuhl Seminar, 5-10 February, 2006
29
Multi-Objective Robust Multi-Objective Robust SolutionsSolutions
Not all Pareto-Not all Pareto-optimal points may optimal points may be robustbe robustA is robust, but B is A is robust, but B is notnotDecision-makers will Decision-makers will be interested in be interested in knowing robust part knowing robust part of the frontof the front
Dagstuhl Seminar, 5-10 February, 2006
30
Domination Based on Aggregate Domination Based on Aggregate FunctionsFunctions
Functions averaged over a delta-Functions averaged over a delta-neighborhodneighborhod
Alternate Strategy: (Type II Robustness)Alternate Strategy: (Type II Robustness)
Dagstuhl Seminar, 5-10 February, 2006
31
Effect of Effect of δδ-neighborhood Size-neighborhood Size Theory and NSGA-II simulationTheory and NSGA-II simulation Larger Larger δδ, more shift from original front, more shift from original front Some part is more sensitive than othersSome part is more sensitive than others
Dagstuhl Seminar, 5-10 February, 2006
32
Effect of Effect of δδ-neighborhood Size-neighborhood Size Theory and NSGA-II simulationTheory and NSGA-II simulation Larger Larger δδ, more shift from original front, more shift from original front Some part is no more robustSome part is no more robust
Dagstuhl Seminar, 5-10 February, 2006
33
Robust Front as Robust Front as Partial Global and Partial LocalPartial Global and Partial Local
Theory:Theory:
For global frontFor global front
Dagstuhl Seminar, 5-10 February, 2006
34
Simulation Using NSGA-IISimulation Using NSGA-IISimulation:Simulation:
Dagstuhl Seminar, 5-10 February, 2006
35
Reliability-Based Optimization Deterministic
optimum often not reliable
Due to uncertainities in decision variables/problem parameters
Find the reliable solution for a specified Reliability
Dagstuhl Seminar, 5-10 February, 2006
36
Constrained Domination for Reliability Consideration Chance
constraints: P(g(x)≤0) ≥ β
β depends on chosen reliability
Prefer reliable solutions
Indicates how P-O front moves away with β
Dagstuhl Seminar, 5-10 February, 2006
37
Goal Programming Using EMO Target function
values are specified
Convert them to objectives and perform domination
Dagstuhl Seminar, 5-10 February, 2006
38
Goal Programming Using EMO Target function values
are specified Convert them to
objectives and perform domination check with them
Dagstuhl Seminar, 5-10 February, 2006
39
Preferred Diversity Find a subset of Pareto-optimal points
dictated by preference information Biased EMO (Branke and Deb, 2005) Reference-point based EMO (Deb and
Sundar, 2006) Knee-based EMO (Branke et al., 2004) Nadir point and EMO (Deb and
Chaudhuri, 2006) Multi-modal EMO (Deb and Reddy, 2003) Variable versus objective space niching
Dagstuhl Seminar, 5-10 February, 2006
40
Preference-Based EMOPreference-Based EMOEMO (NSGA-II) not efficient for many objectivesEMO (NSGA-II) not efficient for many objectives
Large number of points neededLarge number of points neededDomination-based methods are slowDomination-based methods are slow
Dagstuhl Seminar, 5-10 February, 2006
41
EMO for a Biased DistributionChoose a hyper-planeProject points on itCompute two distances: d and d’Compute D=d(d’/d)^aPoint b has small DPoint a has large D
Dagstuhl Seminar, 5-10 February, 2006
42
Biased Distribution in NSGA-IIZDT1ZDT1 ZDT2: a=100ZDT2: a=100
Dagstuhl Seminar, 5-10 February, 2006
43
Biased NSGA-II (cont.)Three-objective Problems: a=0 and a=500Three-objective Problems: a=0 and a=500
Dagstuhl Seminar, 5-10 February, 2006
44
Reference Point Based EMO Wierzbicki, 1980 A P-O solution closer to
a reference point Multiple runs Too structured
Extend for EMO Multiple reference
points in one run A distribution of
solutions around each reference point
Dagstuhl Seminar, 5-10 February, 2006
45
Reference Point Based EMO (cont.) Ranking based on
closeness to each reference point
Clearing within each niche with ε
Dagstuhl Seminar, 5-10 February, 2006
46
More Results Five-objective with
two reference points (z1-5=0.5 & z1-4=0.2, z5=0.8)
A engineering design problem with three reference points
Dagstuhl Seminar, 5-10 February, 2006
47
Knee Based EMOKnee Based EMOFind only the knee or Find only the knee or near-knee solutions near-knee solutions Knees are important Knees are important solutionssolutions
Not much motivation to Not much motivation to move out from kneesmove out from kneesA large gain for a small A large gain for a small lossloss in any pair of in any pair of objectivesobjectives
Non-convex frontNon-convex front No knee pointNo knee point Extreme solutions are Extreme solutions are
attractorsattractors
Dagstuhl Seminar, 5-10 February, 2006
48
Finding Knee SolutionsFinding Knee SolutionsBranke et al. (2004) for more detailsBranke et al. (2004) for more details
Dagstuhl Seminar, 5-10 February, 2006
49
Nadir Point and EMONadir Point and EMOImportant for Important for knowing range and knowing range and normalization of normalization of objectivesobjectivesDifficult to find Difficult to find using classical using classical methodmethod
Pay-off table Pay-off table method does not method does not workwork
Dagstuhl Seminar, 5-10 February, 2006
50
EMO for Finding Nadir PointEMO for Finding Nadir PointEmphasize only extreme Emphasize only extreme pointspointsM≤3 find complete M≤3 find complete front, else use front, else use extremized crowded extremized crowded NSGA-IINSGA-II
Dagstuhl Seminar, 5-10 February, 2006
51
EMO for Finding Nadir Point (cont.)EMO for Finding Nadir Point (cont.)DTLZ problems extended up to 20 objectivesDTLZ problems extended up to 20 objectives
Dagstuhl Seminar, 5-10 February, 2006
52
Multi-Modal EMOsMulti-Modal EMOsDifferent solutions having identical objective values Different solutions having identical objective values Multi-modal Pareto-optimal solutions: Design, Multi-modal Pareto-optimal solutions: Design, Bioinformatics Bioinformatics
Dagstuhl Seminar, 5-10 February, 2006
53
Multiple Gene Subsets for Multiple Gene Subsets for Leukemia SamplesLeukemia Samples
Deb and Reddy Deb and Reddy (BioSystems, 2003)(BioSystems, 2003)
Multiple (26) four-Multiple (26) four-gene combinations gene combinations for 100% for 100% classificationclassification
Discovery of some Discovery of some common genes common genes
Dagstuhl Seminar, 5-10 February, 2006
54
Parameter Versus Objective-space NichingDistribution depends on the space niching is Distribution depends on the space niching is performedperformed
Dagstuhl Seminar, 5-10 February, 2006
55
Redefining Elites To aid in better diversity
Controlled Elitist EMO (Deb and Goel, 2001)
Dagstuhl Seminar, 5-10 February, 2006
56
Controlled Elitism Keep solutions
from dominated fronts in GP
Dagstuhl Seminar, 5-10 February, 2006
57
Controlled Elitism (cont.)ZDT4 has many
local P-O frontsg()=1 is globalControlled elitism
can come closer to global P-O front
Dagstuhl Seminar, 5-10 February, 2006
58
Omni-Optimizer:Omni-Optimizer:Motivation from ComputationMotivation from Computation
Multiple is a generic case, single is specific Multiple is a generic case, single is specific Single objective as a degenerate case multi-Single objective as a degenerate case multi-objective case objective case One algorithm for single and multi-objective One algorithm for single and multi-objective problem solving problem solving ((Deb and Tiwari, 2005Deb and Tiwari, 2005))Accommodating NFL theorem, not violating itAccommodating NFL theorem, not violating it
Single-objective, uni-optimum problemsSingle-objective, uni-optimum problemsSingle-objective, multi-optima problemsSingle-objective, multi-optima problemsMulti-objective, uni-optimal front problemsMulti-objective, uni-optimal front problemsMulti-objective, multi-optimal front problemsMulti-objective, multi-optimal front problems
Dagstuhl Seminar, 5-10 February, 2006
59
Structure of Omni-optimizer Very much like NSGA-II
Epsilon-dominance Variable-space and
objective space niching Use maximum of both crowding distances
Dagstuhl Seminar, 5-10 February, 2006
60
Single-Objective, Uni-Optimum Dominance reduced to simple ‘<‘ Epsilon-dominance to fa< fb-ε
Allows multiple solutions within ε to exist Elite-preservation is similar to CHC and
(μ+λ)-ES
Dagstuhl Seminar, 5-10 February, 2006
61
Shinn et al.’s 12 Problems12 problems
Dagstuhl Seminar, 5-10 February, 2006
62
Single-Objective, Multi-Optima Variable-space niching help find multiple
solutions Weierstrass function
16 minima with f=0
Dagstuhl Seminar, 5-10 February, 2006
63
101044SinSin22(x): 20 Minima(x): 20 Minima
Dagstuhl Seminar, 5-10 February, 2006
64
Himmelblau’s Function: 4 MinimaHimmelblau’s Function: 4 Minima
Dagstuhl Seminar, 5-10 February, 2006
65
Multi-Objective, Uni-Pareto frontMulti-Objective, Uni-Pareto front Constrained and
unconstrained test problems
Dagstuhl Seminar, 5-10 February, 2006
66
More Results Comparable
performance to existing EMO methods
Dagstuhl Seminar, 5-10 February, 2006
67
Multi-Objective, Multi-Optima Multi-Objective, Multi-Optima
Nine regions leading to the same Pareto-optimal front
Multiple solutions cause a single Pareto-optimal point
Dagstuhl Seminar, 5-10 February, 2006
68
Nine Optimal RegionsOm
ni-o
ptim
izer
Omni
-opt
imize
rNSGA-IINSGA-II
Dagstuhl Seminar, 5-10 February, 2006
69
Nine Optimal Fronts Nine Optimal Fronts
Dagstuhl Seminar, 5-10 February, 2006
70
ConclusionsConclusionsFunctional decomposition of NSGA-IIFunctional decomposition of NSGA-II
Non-domination for convergenceNon-domination for convergenceNiching for diverse set of solutionsNiching for diverse set of solutionsElite-preservation for reliable convergenceElite-preservation for reliable convergence
For a new problem-solving, find the For a new problem-solving, find the suitable place to changesuitable place to changeMany different problem-solving tasks Many different problem-solving tasks achieved with NSGA-IIachieved with NSGA-IIOmni-optimizer provides a holistic Omni-optimizer provides a holistic approach for optimizationapproach for optimization
Dagstuhl Seminar, 5-10 February, 2006
71
Thank You for Your AttentionThank You for Your Attention
For further information:For further information:http://www.iitk.ac.in/kangalEmail: [email protected]: [email protected]
Acknowledgement:Acknowledgement: KanGAL students, staff and collaboratorsKanGAL students, staff and collaborators