toward a natural genetic / evolutionary algorithm for multiobjective optimization
DESCRIPTION
New Genetic or Evolutionary Algorithm for multiobjective optimization, that attempts to find tradeoff solutions and scales easily with increase in parameter space as well as objective space. Does not use complex niche calculation that is used in existing multiobjective genetic algorithms.TRANSCRIPT
Toward a NaturalGenetic/Evolutionary Algorithm for
Multiobjective OptimizationHariharane Ramasamy
Evolutionary Systems Inc., Cupertino, CA.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 1/58
Outline
1. Problem Description
2. Practical Examples
3. Classical Algorithms
4. Genetic Algorithms
5. Multiobjective Genetic Algorithms
6. Extended Genetic Algorithms
7. Results
8. Protein Folding—Lattice Model
9. Future Research
10. Conclusion
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 2/58
Problem Description
1. In most practical optimization problems, one deals withsimultaneous optimization of multiple objectives that mayconflict with one another.
2. We seek efficient ways to optimize systems specified by atuple of parameters (not necessarily numerical). Eachtuple determines a system with fitnesses for the individualobjectives, this maps the parameter space into the fitnessspace.
3. The best compromise solutions are called the tradeoff ornondominated solutions and form the Pareto front. Theparameter set associated with the Pareto front is calledthe Pareto set
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 3/58
An Example
Figure 1: Multiobjective problem example with two objectives
The plot is an example of a biobjective problem (f1, f2) that depends on two parameters (x1,x2). Assuming minimization, the thick dark line represents the best tradeoff solutions andconstitute the Pareto front. In the above figure, point O is dominated by all points in theinterior of the quadrant OAB. The lines ~OA, ~OB, ~OC, and ~OD define the boundary pointswhere f1 or f2 gets better or worse by moving either horizontally or vertically with respect topoint O. However, the points inside the region OAC and ODB are neither dominated nornondominated by the point O. Each one of them has either f1 or f2 optimal, but not both withrespect to point O.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 4/58
Local and Global Pareto Front
A local Pareto front dominates every point in its neighborhood and a global Pareto frontdominates every point in the objective space. In the left plot, A and B are two disjointobjective regions. Assuming minimization of objectives, B contains the global Pareto frontand A the local Pareto front. In the right plot, the curve AOBCD contains the entire objectivespace; OB is A local Pareto front, and the global Pareto front is the union of AO and CD.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 5/58
Multiobjective problem
Thus a multiobjective problem can be stated as:
Minimize F (x) = (f1(x), f2(x), ..., fm(x))
such that x ∈ S
x = (x1, x2, ..., xn)
S is the feasible parameter space
m is the number of objectives
we seek feasible configurations, tuples (x1, x2,...,xn) that maps into points(f1, f2,...,fm) on objective space, which can’t be improved w.r.t anyobjective without worsening them w.r.t. some other objective.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 6/58
Practical Examples
1. Protein folding problem
2. Multiple knapsack problem
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 7/58
Protein Folding Problem
Protein Fold
The primary sequence of a protein (left), based on its composition of amino acids, folds intoa unique three dimensional structure (right) under certain physiological conditions. Thefolding is driven by multiple physio-chemical properties like hydrogen bonds, protectinghydrophobes from water- a multiobjective problem.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 8/58
Protein Folding—Lattice Model
Allowed Moves Buried Hydrophobe Rule
in in
.
The amino acids in the primary sequence of a protein is connected to each other by theformation of peptide bond that resembles a rectangular lattice. Using a three dimensionalrectangular lattice, protein folding is simulated with move rules defined by the left picture.The right picture defines the condition under which a side chain in the rectangular lattice isburied. The mutiobjective optimization is to find the three dimensional conformation that
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 9/58
Multiple Knapscack Problem
Problems such as bin packing, cutting stock, and financial managementcan all be modeled as multiple knapsack problems, in which;
1. we are given a set of items with specific weights and profits andmultiple knapsacks with fixed capacity.
2. We have to fill up each knapsack, maximizing its profits and spaceutilization.
We present results for two variations of multiple knapsack problems.
1. In one problem, when an item is selected, it is included in allknapsacks.
2. In the second, the selected item is included in only one knapsack.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 10/58
Existing Methods
1. Traditional methods
2. Genetic algorithm methods
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 11/58
Summary of Traditional Methods
Name Advantages Disadvantages
WeightedMethods
Simple, works well when Paretofront is simple.
Computationally expensive, failswith complicated Pareto fronts.
ǫ-constraintMethod
Works with complicated regions. Success largely depends on theinitial solution, which might beselected in an infeasible region.
LexicographicApproach
Simple, optimizes objectives se-quentially by predefine priority.
A limited number of Pareto opti-mal points are found.
Normal Bound-ary Intersec-tion (NBI)
Finds a well-spread Paretopoints.
Fails with complicated land-scapes in objective space.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 12/58
Genetic Algorithm
Genetic algorithm, motivated from natural evolution, repeatedly apply genetic operators on apool of solutions called populations.
generate initial population P randomlyset new population Pn
while desired convergence is achieved in the population doperform crossover with probability pc
if the fitness of offspring are better then parents thenadd the offspring to Pn
elsedecide with very low probability to include offspring in Pn
end ifperform mutation with probability pm
perform reproduction with probability pr
if size of(Pn) ≥ N thenreplace P with N members from Pn
reduce size of Pn by Nend if
end while
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 13/58
Genetic Algorithm—Metrics
1. Diversity—Genetic Algorithm often converge to a few dominant solutions andlose diversity. Any further optimization is ineffective since repeatedapplication of operators yield the few converged dominant solutions. Diversityis essential for multiobjective problems.
2. Elitism—selecting individuals with a bias to create better individuals is calledelitism; selecting the best parents to replace the less fit members drives thepopulation to converge to a few best parents.
3. Scalability—Performance of the algorithm should not deteriorate with anincrease in the number of objectives and parameters.
4. Exploration—Ability of an algorithm to find new solutions or reproduce lostsolutions is called exploration.
5. Exploitation—Ability to retain the current best solutions in the population iscalled exploitation.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 14/58
Niching
1. Goldberg and Richardson proposed the initial method to promote thediversity within the population.
2. The method used a distance function, called sharing function, whichcalculates similarities within the population members using aparameter called niche radius in parameter or objective space.
3. The members in a crowded neighborhood get their fitness degraded,thus preventing their selection for next generation.
4. The sharing function helps to keep the solutions diverse.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 15/58
Niching
1. Niche radius is difficult to determine. Too large a radius will ignore good tradeoffsolutions; too small can consider points that are already dominated in the set.
2. When there are more than two objectives, the number of solutions on the Pareto frontincreases, along with the complexity of the niche calculation.
3. In practical problems with more than two objectives, the sharing function has met littlesuccess.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 16/58
Multiobjective Genetic Algorithms
Name Advantages Disadvantages
ParetoArchivedEvolutionStrategy
Gives a clear description forclass of problems that genetic al-gorithm should excel.
Does not scale well with problemdimension.
Vector Evalu-ated GeneticAlgorithm
Executes by population reshuf-fling using each objective func-tion.
Loses good solutions and doesnot have a good spread of Paretooptimal set.
Weight-BasedGenetic Algo-rithm
Similar to weighted methods,evolves the weight along withthe optimization of the objectivefunction.
Fails when the objective functionis nonconvex.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 17/58
Multiobjective Genetic Algorithms
Name Advantages Disadvantages
Niched ParetoGenetic Algo-rithm
Builds a small nondominated setalong with the algorithm.
Success largely depends on thesmall nondominated subset.
NondominatedSorting Ge-netic Algorithm
Niching is performed in the deci-sion space.
Niching function is too complex,and scales poorly as number ofobjectives increases.
StrengthPareto Evo-lutionaryAlgorithm
Finds well-spread Pareto pointsby maintaining an external non-dominated set of fixed size.
clustering is performed to main-tain the size of the externalPareto set. This clustering pro-cess scales poorly with the size.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 18/58
Extended Genetic Algorithms
Will be published in paper, only results are presented here
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 19/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. find more that one distinct points in the Pareto set that map to the same pointin the Pareto front?
8. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 20/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. find more that one distinct points in the Pareto set that map to the same pointin the Pareto front?
8. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 21/58
Schaeffer-1
0
5
10
15
20
25
0 5 10 15 20 25
f2
f1
Table 1: Summary of Results—Schaeffer 1Name Pop. Size Total - Pareto front Points Duration
NSGA 100 ≤ 30
sec
SPEA 100 100 ≤ 30 sec
MOEAD 100 100 ≤ 30 sec
NSGA 1000 1000 2 min
SPEA 1000 1000 8 min
MOEAD 1000 300 11 sec
MOEAD 10000 300 11 sec
SEGA 100 2541 ≤ 30 sec
SEGA 100 18500+ 3 min
PEGA 100 608 ≤ 30 sec
PEGA 100 900+ 3 min
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 22/58
Schaeffer-2
0
5
10
15
20
25
30
35
-1 0 1 2 3 4
f2
f1
Table 2: Summary of Results—Schaeffer 2Name Pop. Size Total - Pareto front Points Duration
NSGA 100 100 ≤ 30
sec
SPEA 100 100 ≤ 30
sec
MOEAD 100 100 ≤ 30
sec
SEGA 100 3767 ≤ 30
sec
PEGA 100 x 2 409 ≤ 30
sec
SEGA 100 114000+ 3 min
PEGA 100 x 2 2300+ 3 min
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 23/58
Viennet
0
10
20
30
40
50
60
0 1 2 3 4 5 6 7 8 9
f2
f1
0
0.05
0.1
0.15
0.2
0 1 2 3 4 5 6 7 8 9
f3
Table 3: Viennet—Summary of ResultsName Unique PF points Unique PS points
NSGA 100 100
SPEA 1 1
MOEAD 40 65
SEGA 2394 2405
PEGA 3522 3605
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 24/58
Kalyanamoy
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
f3
DTLZ4 - SEGA Population Plot
SEGAPEGA
f1
f2
f3
Table 4: Kalayanomoy Scalable Multiobjective ProblemDTLZ4—Seven Objectives—Summary of Results
Name Unique Pareto front points Unique Pareto set points
NSGA2 999 999
SEGA 744 773
PEGA 3728 4815
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 25/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. find more that one distinct points in the Pareto set that map to the same pointin the Pareto front?
8. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 26/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 500 Generations
1000
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 1000 Generations
1000
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 1500 Generations
1500
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 2000 Generations
2000
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 2500 Generations
2500
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 3000 Generations
3000
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front—Progress
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
Knapsack 750-3 - F1 Versus F2 after 3300 Generations
3300
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 27/58
Pareto front
20000
25000
30000
35000
20000 25000 30000 35000
f 3 P
rofit
s
f1 Profits
Knapsack 750-3 - PF F1 Versus F2 after 3300 Generations
500100015002000250030003300
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 28/58
Pareto front
20000
25000
30000
35000
20000 25000 30000 35000
f 3 P
rofit
s
f1 Profits
Knapsack 750-3 - PF F1 Versus F3 after 3300 Generations
500100015002000250030003300
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 28/58
Pareto front
20000
25000
30000
35000
20000 25000 30000 35000
f 3 P
rofit
s
f1 Profits
Knapsack 750-3 - PF F2 Versus F3 after 3300 Generations
500100015002000250030003300
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 28/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. find more that one distinct points in the Pareto set that map to the same pointin the Pareto front?
8. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 29/58
Exploration And Exploitation
20000
25000
30000
35000
20000 25000 30000 35000
f 2 P
rofit
s
f1 Profits
PEGA - F1 Vs F2 Population
F1F2F3
F1-F2F1-F3F2-F3
F1-F2-F3
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 30/58
Exploration And Exploitation
20000
25000
30000
35000
20000 25000 30000 35000
f 3 P
rofit
s
f2 Profits
PEGA - F2 Vs F3 Population
F1F2F3
F1-F2F1-F3F2-F3
F1-F2-F3
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 30/58
Exploration And Exploitation
20000
25000
30000
35000
20000 25000 30000 35000
f 3 P
rofit
s
f1 Profits
PEGA - F1 Vs F3 Population
F1F2F3
F1-F2F1-F3F2-F3
F1-F2-F3
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 30/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 31/58
Scaling with Objectives
Kalyanamoy Scalable Objective Test function - Five
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 32/58
Scaling with Objectives
DTLZ4—f1 vs f2—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 2
f1
DTLZ4 - 5 Objectives - PF f1 vs f2 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 2
f1
DTLZ4 - 5 Objectives - PF f1 vs f2 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 2
f1
DTLZ4 - 5 Objectives - PF f1 vs f2 - PEGA
PEGA
Figure 2: Results of DTLZ4 with five objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 33/58
Scaling with Objectives
DTLZ4—f3 vs f4—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 4
f3
DTLZ4 - 5 Objectives - PF f3 vs f4 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 4
f3
DTLZ4 - 5 Objectives - PF f3 vs f4 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 4
f3
DTLZ4 - 5 Objectives - PF f3 vs f4 - PEGA
PEGA
Figure 3: Results of DTLZ4 with five objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 34/58
Scaling with Objectives
DTLZ4—f1 vs f5—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 5
f1
DTLZ4 - 5 Objectives - PF f1 vs f5 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 5
f1
DTLZ4 - 5 Objectives - PF f1 vs f5 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 5
f1
DTLZ4 - 5 Objectives - PF f1 vs f5 - PEGA
PEGA
Figure 4: Results of DTLZ4 with five objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 35/58
Scaling with objectives
Kalyanamoy Scalable Objective Test function - Seven
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 36/58
Scaling with Objectives
DTLZ4—f1 vs f2—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 2
f1
DTLZ4 - 7 Objectives - PF f1 vs f2 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 2
f1
DTLZ4 - 7 Objectives - PF f1 vs f2 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 2
f1
DTLZ4 - 7 Objectives - PF f1 vs f2 - PEGA
PEGA
Figure 5: Results of DTLZ4 with seven objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 37/58
Scaling with Objectives
DTLZ4—f3 vs f4—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 4
f3
DTLZ4 - 7 Objectives - PF f3 vs f4 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 4
f3
DTLZ4 - 7 Objectives - PF f3 vs f4 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 4
f3
DTLZ4 - 7 Objectives - PF f3 vs f4 - PEGA
PEGA
Figure 6: Results of DTLZ4 with seven objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 38/58
Scaling with Objectives
DTLZ4—f5 vs f6—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 6
f5
DTLZ4 - 7 Objectives - PF f5 vs f6 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 6
f5
DTLZ4 - 7 Objectives - PF f5 vs f6 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 6
f5
DTLZ4 - 7 Objectives - PF f5 vs f6 - PEGA
PEGA
Figure 7: Results of DTLZ4 with seven objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 39/58
Scaling with Objectives
DTLZ4—f1 vs f2—NSGA (left), SEGA (middle), PEGA(right)
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 7
f1
DTLZ4 - 7 Objectives - PF f1 vs f7 - NSGA2
NSGA2
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 7
f1
DTLZ4 - 7 Objectives - PF f1 vs f7 - SEGA
SEGA
0
0.5
1
1.5
2
0 0.5 1 1.5 2
f 7
f1
DTLZ4 - 7 Objectives - PF f1 vs f7 - PEGA
PEGA
Figure 8: Results of DTLZ4 with seven objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 40/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 41/58
Scaling with Parameter Space
Kalyanamoy Scalable Objective Test function - Five
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 42/58
Scaling with Parameter Space
DTLZ4—Pareto set ( x1, x2, x3)—NSGA (left), SEGA (middle), PEGA(right)
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x3
DTLZ4 - 5 Objectives - PS (x1,x2,x3) - NSGA2
NSGA2
x1
x2
x3
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x3
DTLZ4 - 5 Objectives - PS (x1,x2,x3) - SEGA
SEGA
x1
x2
x3
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x3
DTLZ4 - 5 Objectives - PS (x1,x2,x3) - PEGA
PEGA
x1
x2
x3
Figure 9: Results of DTLZ4 with five objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 43/58
Scaling with Parameter Space
Kalyanamoy Scalable Objective Test function - Seven
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 44/58
Scaling with Parameter Space
DTLZ4—Pareto set ( x1, x2, x3)—NSGA (left), SEGA (middle), PEGA(right)
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x3
DTLZ4 - 7 Objectives - PS (x1, x2, x3) - NSGA2
NSGA2
x1
x2
x3
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x3
DTLZ4 - 7 Objectives - PS (x1, x2, x3) - SEGA
SEGA
x1
x2
x3
0 0.2
0.4 0.6
0.8 1 0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
x3
DTLZ4 - 7 Objectives - PS (x1, x2, x3) - PEGA
PEGA
x1
x2
x3
Figure 10: Results of DTLZ4 with seven objectives
Among the tests we did, only the NSGA results can be compared qualitatively with SEGAand PEGA. The graph clearly show the superior performance of EGAs.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 45/58
Multiple Knapsack Problem(Exclusive)
We present results for 750 items and 3 knapsacks
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 46/58
Three Knapsacks, 750 Items—Profits
SEGA f1 vs f3 PEGA f1 vs f3
0
5000
10000
15000
20000
25000
30000
35000
0 5000 10000 15000 20000 25000 30000 35000
f 3
f1
Knapsack 750-3 - f1 vs f3 Profits - SEGA
0
5000
10000
15000
20000
25000
30000
35000
0 5000 10000 15000 20000 25000 30000 35000
f 3
f1
Knapsack 750-3 - f1 vs f3 Profits - SEGA
Figure 11: Profits KP 750-3. The left and right columns show the profits-to-weightsratio obtained by the sequential and parallel extended genetic algorithms for f1 vsf3. The conflicting nature of the objective functions can be easily seen in the plots
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 47/58
Three Knapsacks, 750 Items—Items
SEGA f1 vs f3 PEGA f1 vs f3
0
250
500
750
0 250 500 750
Ite
ms in
KS
3
Items in KS1
Knapsack 750-3 - f1 vs f3 Items - SEGA
0
250
500
750
0 250 500 750
Ite
ms in
KS
3
Items in KS1
Knapsack 750-3 - f1 vs f3 Items - PEGA
Figure 12: Profits KP 750-3. The plots show the items obtained by the sequentialand parallel extended genetic algorithms for f1 vs f3. The conflicting nature of theobjective functions can be easily seen in the plots
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 48/58
Parameter Space—Exploration
Solutions that Map to Similar Sum
600
650
700
750
800
0 25 50 75 100
Tota
l Item
s Inc
luded
in a
ll Kna
psac
ks(S
EGA+
PEGA
)
Number of Occurrences
Knapsack 750-3 - Number of unique solutions
Figure 13: Item KP 750-3. The plots show the number of solutions assigned to thethree knapsacks that map to the same sum
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 49/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 50/58
Scattered Pareto Set
Non Dominated Points/Non Dominated Set
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
f 2
f1
F1 versus F2 Corresponding x1, x2, x3 - Set
0 0.2 0.4 0.6 0.8 1x1 -1
-0.5 0
0.5 1
x2-1
-0.5
0
0.5
1
x3
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
f 2
f1
F1 versus F2 Corresponding x1, x2, x3 - Set
0 0.2 0.4 0.6 0.8 1x1 -1
-0.5 0
0.5 1
x2-1
-0.5
0
0.5
1
x3
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 51/58
Results
We’ll see that our versions outperformed a large number of existing algorithms, ona battery of tests commonly used in the literature to evaluate optimizationmethods. Specifically, do our proposed methods :
1. find more Pareto front points with ease?
2. effectively move towards Pareto front?
3. explore and exploit the search space effectively?
4. scale with increase in number of objectives and parameters?
5. find different Pareto set points that map to one point in objective space?
6. scale well with scattered Pareto set?
7. have a framework for flexible extension?
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 52/58
Flexibility
The results presented so far were ran with the same algorithmwithout altering any parameters.
Niching methods require suitable parameter selection to reachPareto front which is totally absent here.
Population size is not increased to obtain a larger Pareto set. This isa great advantage as we cannot provide infinite source to hold thepopulation.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 53/58
Protein Folding - Interesting Results
Helix Beta Sheets
Figure 14: Two interesting results obtained by extended genetic algorithms. The red and bluerepresents hydrophobic and hydrophilic amino acids. The two green lattices, connecting thetwo helices, represent the glycine, which has four additional lattice movements. Glycine isoften found near sharp turns in the protein due to its small size. In the right result, every othermember in the sequence is hydrophobic, and hence the algorithm produced a beta sheet thatfolds against itself to bury the hydrophobes
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 54/58
Conclusion
Pseudo-parallelism is a feature in the genetic algorithm in which parameter space isexplored simultaneously by different members in the population. Extended geneticalgorithms successfully extended pseudo-parallelism not only in parameter space butalso in objective space.
The new algorithms leave much room for extension and improvements. A few keyenhancements are mentioned in Future research.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 55/58
Conclusion
Pseudo-parallelism is a feature in the genetic algorithm in which parameter space isexplored simultaneously by different members in the population. Extended geneticalgorithms successfully extended pseudo-parallelism not only in parameter space butalso in objective space.
Parallel and sequential extended genetic algorithms represent two different extensionsof the genetic algorithm. Both the algorithms were applied for the first time to theprotein folding problem and were presented in several conferences in the year 1996.
The new algorithms leave much room for extension and improvements. A few keyenhancements are mentioned in Future research.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 55/58
Conclusion
Pseudo-parallelism is a feature in the genetic algorithm in which parameter space isexplored simultaneously by different members in the population. Extended geneticalgorithms successfully extended pseudo-parallelism not only in parameter space butalso in objective space.
Parallel and sequential extended genetic algorithms represent two different extensionsof the genetic algorithm. Both the algorithms were applied for the first time to theprotein folding problem and were presented in several conferences in the year 1996.
Extended genetic algorithms have better performance in exploring objective andparameter space than do existing multiobjecitve genetic algorithms.
The new algorithms leave much room for extension and improvements. A few keyenhancements are mentioned in Future research.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 55/58
Conclusion
Pseudo-parallelism is a feature in the genetic algorithm in which parameter space isexplored simultaneously by different members in the population. Extended geneticalgorithms successfully extended pseudo-parallelism not only in parameter space butalso in objective space.
Parallel and sequential extended genetic algorithms represent two different extensionsof the genetic algorithm. Both the algorithms were applied for the first time to theprotein folding problem and were presented in several conferences in the year 1996.
Extended genetic algorithms have better performance in exploring objective andparameter space than do existing multiobjecitve genetic algorithms.
Duplication and transposon operators were introduced for the first time in the geneticalgorithm.
The new algorithms leave much room for extension and improvements. A few keyenhancements are mentioned in Future research.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 55/58
Conclusion
Pseudo-parallelism is a feature in the genetic algorithm in which parameter space isexplored simultaneously by different members in the population. Extended geneticalgorithms successfully extended pseudo-parallelism not only in parameter space butalso in objective space.
Parallel and sequential extended genetic algorithms represent two different extensionsof the genetic algorithm. Both the algorithms were applied for the first time to theprotein folding problem and were presented in several conferences in the year 1996.
Extended genetic algorithms have better performance in exploring objective andparameter space than do existing multiobjecitve genetic algorithms.
Duplication and transposon operators were introduced for the first time in the geneticalgorithm.
Extended genetic algorithms’ performed better with increase in the number ofobjectives and variables than were the multiobjective genetic algorithms.
The new algorithms leave much room for extension and improvements. A few keyenhancements are mentioned in Future research.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 55/58
Future Research
Natural Genetic Algorithm—described in the next slide represents the combination ofparallel and sequential extended genetic algorithm.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 56/58
Future Research
Natural Genetic Algorithm—described in the next slide represents the combination ofparallel and sequential extended genetic algorithm.
Coding—binary coding is used in most of the problems. The selection of coding dependson the type of problem and the parameters associated with the problem.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 56/58
Future Research
Natural Genetic Algorithm—described in the next slide represents the combination ofparallel and sequential extended genetic algorithm.
Coding—binary coding is used in most of the problems. The selection of coding dependson the type of problem and the parameters associated with the problem.
Operators—duplication and transposon operators used an arbitrary method. There aredifferent ways the operators can be performed.
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 56/58
Future Research
Toward a Natural Genetic / Ramasamy Algorithm
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 57/58
Questions
please email sundar_hariharane @ yahoo.com
Toward a Natural Genetic/Evolutionary Algorithm for Multiobjective Optimization – p. 58/58