moea/d: a multiobjective evolutionary algorithm based on decomposition reporter: steven date:...
Post on 21-Dec-2015
248 views
TRANSCRIPT
![Page 1: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/1.jpg)
MOEA/D: A Multiobjective Evolutionary Algorithm
Based on Decomposition
Reporter: Steven
Date: 2011/5/4
![Page 2: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/2.jpg)
Why decompose the MOP ? Most MOPs may have many or even infinite Pareto
optimal vectors. It is very time-consuming to obtain the complete PF.
Decision maker may not be interested in having an unduly large number of Pareto optimal vectors to deal with due to overflow of information.
Many MO algorithms are to find a manageable number of Pareto optimal vectors which are evenly distributed along the PF, and thus good representatives of the entire PF.
![Page 3: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/3.jpg)
DECOMPOSITION OF MULTIOBJECTIVE OPTIMIZATION A. Weighted Sum Approach
: be a weight vector
: Object solution
m
i i
Tm
1
1
1
),...,(
Tm xfxfxF ))(),...,(()( 1
m
iii
ws xfxgMaximize1
)()|(
![Page 4: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/4.jpg)
DECOMPOSITION OF MULTIOBJECTIVE OPTIMIZATION B. Tchebycheff Approach
: is the reference point
|})(|{max*),|( *
1iii
mi
te zxfzxgMinimize
Tmzzz *)*,...,(* 1
1f
2f
0
*Z
|)(| *ii zxf
![Page 5: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/5.jpg)
DECOMPOSITION OF MULTIOBJECTIVE OPTIMIZATION C. Boundary Intersection (BI) Approach
xdxFztosubject
dzxgMinimize bi
)(*
*),|(
![Page 6: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/6.jpg)
DECOMPOSITION OF MULTIOBJECTIVE OPTIMIZATION C. Penalty-based Boundary Intersection (BI)
Approach
||)*()(||||||
||)(*||
,*),|(
121
21
dzxFdandxFz
dwhere
xtosubjectddzxgMinimizeT
bi
![Page 7: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/7.jpg)
THE FRAMEWORK OF MOEA/D
At each generation , MOEA/D with the Tchebycheff approach maintains:
![Page 8: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/8.jpg)
THE FRAMEWORK OF MOEA/D
The algorithm works as follows:
![Page 9: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/9.jpg)
Step 1) Initialization:
![Page 10: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/10.jpg)
Step 2) Update:
![Page 11: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/11.jpg)
1f
2f
0
*Z
)(
)(,...)( 111
lB
xfxf T
)(
)(,...)( 212
kB
xfxf T
y
y‘
![Page 12: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/12.jpg)
Step 3) Stopping Criteria: If stopping criteria is satisfied,then stop and
output EP. Otherwise, go to Step 2.
![Page 13: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/13.jpg)
Discussions of MOEA/D 1) Why a Finite Number of Subproblems are
Considered in MOEA/D:
MOEA/D spends about the same amount of effort on each of the N aggregation functions, while MOGLS randomly generates a weight vector at each iteration, aiming at optimizing all the possible aggregation functions.
Since the computational resource is always limited, optimizing all the possible aggregation functions would not be very practical, and thus may waste some computational effort.
![Page 14: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/14.jpg)
Discussions of MOEA/D 2) How Diversity is Maintained in MOEA/D:
NSGA-II and SPEA-II → crowding distancesMOEA/D → The “diversity” among these subproblems will naturally lead to diversity in the population.
3) Mating Restriction and the Role of in MOEA/D:T is too small :the solution could be very close to their parents, the algorithm lacks the ability to explore new areas in the search space.T is too large :the exploitation ability of the algorithm is weakened.
![Page 15: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/15.jpg)
Multiobjective 0–1 Knapsack Problem Given a set of n items and a set of m
knapsacks, the multiobjective 0–1 knapsack problem (MOKP) can be stated as:
is the profit of item j in knapsack i
is the weight of item j in knapsack i
is the capacity of knapsack i
item i is selected and put in all the knapsacks.ic
1ix
0ijp
0ijw
![Page 16: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/16.jpg)
![Page 17: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/17.jpg)
Experimental Results- CPU time
![Page 18: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/18.jpg)
Experimental Results- C metric C(A,B) is defined as the percentage of the
solutions in B that are dominated by at least one solution in A
![Page 19: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/19.jpg)
Experimental Results- D metric Distance from Representatives in the PF ( D-
metric):
![Page 20: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/20.jpg)
![Page 21: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/21.jpg)
Fig. 4. Plots of the non-dominated solutions with the lowest D-metric in 30 runs of MOEA/D and MOGLS with the weighted sum approach for all the 2-objective MOKP test instances.
![Page 22: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/22.jpg)
Fig. 5. Plots of the nondominated solutions with the lowest D-metric in 30 runsof MOEA/D and MOGLS with the Tchebycheff approach for all the 2-objectiveMOKP test instances.
![Page 23: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/23.jpg)
We use five widely used bi-objective ZDT test instances and two 3-objective instances in comparing MOEA/D with NSGA-II
![Page 24: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/24.jpg)
A Bit More Effort on MOEA/D: Can MOEA/D with other advanced decomposition methods such as the PBI approach find more evenly distributed solutions for 3-objective test instances like DTLZ1 and DTLZ 2
PBINSGA-11MOEA/D Te
![Page 25: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/25.jpg)
A Bit More Effort on MOEA/D: Can MOEA/D with objective normalization perform better in the case of disparately scaled objectives as in ZDT3
normalization
f2→10f2
Without normalization
![Page 26: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/26.jpg)
Sensitivity of in MOEA/D
MOEA/D is not very sensitive to the setting of , at least for MOPs that are somehow similar to these test instances
![Page 27: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/27.jpg)
MOEA/D Using Small Population
![Page 28: MOEA/D: A Multiobjective Evolutionary Algorithm Based on Decomposition Reporter: Steven Date: 2011/5/4](https://reader033.vdocuments.mx/reader033/viewer/2022061607/56649d615503460f94a43985/html5/thumbnails/28.jpg)
Scalability