Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Stable Matching-Based Selection in Evolutionary Multiobjective
Optimization
Sam Kwong, Li Ke, M Wu and Qingfu Zhang
Department of Computer Science
City University of Hong Kong
Contents◦ Introduction◦ Stable Matching-Based Selection◦ Stable Matching-Based Selection with Incomplete Lists◦ Adaptive Stable Matching-Based Selection◦ Conclusion
2
Contents◦ Introduction◦ Multiobjective Optimization Problem◦ Pareto Dominance◦ Selection in EMO◦ MOEA/D
◦ Stable Matching-Based Selection◦ Stable Matching-Based Selection with Incomplete Lists◦ Adaptive Stable Matching-Based Selection◦ Conclusion
3
Introduction ◦ Multiobjective optimization problem (MOP)
◦ 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚/𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝐹𝐹 𝑚𝑚 = (𝑓𝑓1 𝑚𝑚 ,𝑓𝑓2 𝑚𝑚 , … , 𝑓𝑓𝑚𝑚 𝑚𝑚 )𝑇𝑇
◦ 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑚𝑚𝑠𝑠𝑠𝑠 𝑠𝑠𝑡𝑡 𝑚𝑚 ∈ Ω
4
Real World MOPs◦ Multiobjective traveling salesman problem (TSP)
◦ The goal of the planning is to minimize the travel time and travel cost
5
A salesman need to travel from the origin city to a couple of cities to meet customers. Each city need to be visited exactly once. There are different ways to travel from each city to another, each of which takes different travelling times and different costs. The sequence to visit each city could be arbitrary.
Real World MOPs◦ Multiobjective traveling salesman problem (TSP)
◦ Different visiting sequences result in significance differences in traveling time and cost
◦ Even if the visiting sequence is fixed, different transportation may result in different objectives
6
Short traveling timeHigh cost
Low cost Long traveling time
Real World MOPs◦ Antenna design
7
There is growing interest for small antennas that concurrently have higher functionality and operability. Designers are interested in antennas with higher gain and larger bandwidth.
Real World MOPs◦ Antenna design◦ E.g., [1]
8
Pareto Dominance◦ Pareto Dominance◦ 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝐹𝐹 𝑚𝑚 = (𝑓𝑓1 𝑚𝑚 ,𝑓𝑓2 𝑚𝑚 )𝑇𝑇
9
1
2
Solution 1 dominates solution 2
Pareto Dominance◦ Pareto Dominance◦ 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝐹𝐹 𝑚𝑚 = (𝑓𝑓1 𝑚𝑚 ,𝑓𝑓2 𝑚𝑚 )𝑇𝑇
10
1
2
Solution 1 and solution 2 cannot be comparedThey are nondominated with each other
Pareto Dominance◦ Pareto Dominance◦ 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝐹𝐹 𝑚𝑚 = (𝑓𝑓1 𝑚𝑚 ,𝑓𝑓2 𝑚𝑚 )𝑇𝑇
11
Attenable Set
Pareto font (PF) consists of the objective vectors of all Pareto optimal solutions
Pareto optimal solution
Selection ◦ EA framework
12
Start
Population initialization
Terminate?
Offspring reproduction
Selection
End
A set of solutions are randomly sampled in the search space
Produce a set of offspring solutions using reproduction operators
Promising solutions are selected to survive in the next iteration
Selection in EMO◦ Two main requirements◦ Convergence◦ Obtained solutions are as close to the PF as possible
◦ Diversity ◦ Obtained solutions are widely distributed along the PF
13
Convergence
Selection in EMO◦ Two main requirements ◦ Convergence◦ Obtained solutions are as close to the PF as possible
◦ Diversity ◦ Obtained solutions are widely distributed along the PF
14
Diversity
MOEA/D◦ Multiobjective evolutionary algorithm based on decomposition
[2]◦ Decomposes an MOP into a number of single-objective optimization
problems and optimizes them simultaneously ◦ Weighted sum, Tchebycheff and penalty-based intersection approaches
◦ E.g. Tchebycheff approach
15
MOEA/D
◦ Selection in MOEA/D ◦ Select a solution for each subproblem to minimize the objective of the
subproblem◦ Subproblems are constructed using a set of evenly distributed weight
vectors
16
Convergence
Diversity
Problem:During the optimization, the currently elite solutions of some relatively easier subproblems might also be good candidates for the others. In this case, these elite solutions can easily take over all these subproblems. MOEA/D may fail to maintain population diversity.
Contents◦ Introduction◦ Stable Matching-Based Selection◦ Stable Matching-Based Selection with Incomplete Lists◦ Adaptive Stable Matching-Based Selection◦ Conclusion
17
Stable Marriage Problem◦ Stable marriage problem (SMP) [2] is a problem of finding a stable
matching between two sets of matching agents, given that each matching agent has a preference list over all matching agents in the other set.
◦ It is an early work of Nobel Prize Winner Lloyd S. Shapley
18
Stable Marriage Problem
19
Not a stable matching.
𝑚𝑚2 prefers 𝑤𝑤3 to 𝑤𝑤1, meanwhile, 𝑤𝑤3 prefers 𝑚𝑚2 to 𝑚𝑚3. Therefore, 𝑚𝑚2 and 𝑤𝑤3 may divorce and marry to each other.
This unstable situation should be avoided in a stable matching solutions.
Stable Marriage Problem
20
A stable matching solution.
Stable Marriage Problem◦ Stable matching algorithm proposed in [3]
21
Note: this is a men-oriented algorithm, where the number of men should be no more than the number of women.
Stable Marriage Problem◦ SMP is a one-one matching problem
◦ Matching problem can be generalized into many different problems◦ College admission problem/hospitals-residents problem (many-to-one
matching)◦ Stable roommate problem (only one set of matching agents)
22
Stable Matching-Based Selection◦ Model the selection in MOEA/D as an SMP◦ Subproblems and solutions are treated as two sets of matching agents ◦ Each subproblem has a preference list over all solutions◦ Each solution has a preference list over all subproblems
◦ Preference value of 𝑝𝑝 on 𝐱𝐱◦ ∆p 𝑝𝑝, 𝐱𝐱 = 𝑔𝑔(𝐱𝐱|𝑤𝑤, 𝑚𝑚∗)
◦ Objective function of subproblem 𝑝𝑝 given solution 𝐱𝐱◦ Subproblem prefers solution that can minimize its objective
◦ Preference value of x on 𝑝𝑝
◦ ∆x 𝐱𝐱, 𝑝𝑝 = �𝐹𝐹 𝐱𝐱 − 𝑤𝑤𝑇𝑇 �𝐹𝐹 𝐱𝐱𝑤𝑤𝑇𝑇𝑤𝑤
𝑤𝑤◦ Distance between the objective vector of solution 𝐱𝐱 and weight vector of subproblem 𝑝𝑝
◦ Solution prefers subproblem closer in objective space
23
Convergence
Diversity
Stable Matching-Based Selection ◦ Model the selection in MOEA/D as an SMP◦ The selection is achieved by computing a stable matching between
subproblems and solutions using proposed algorithm in [3]
◦ Since the preferences are defined considering convergence and diversity◦ A stable matching to the modeled SMP is regarded as a balance between
convergence and diversity
24
Stable Matching-Based Selection ◦ MOEA/D-STM [4]◦ K. Li, Q. Zhang, S. Kwong, M. Li, and R. Wang, Stable matching-based
selection in evolutionary multiobjective optimization. IEEE Trans. Evol. Comput., vol. 18, no. 6, pp. 909-923, Dec. 2014.
25
Stable Matching-Based Selection◦ Problem◦ A stable matching solution cannot always ensure that a subproblem select a
solution close to itself◦ Experiments in [5] show that for MOP test instances, MOEA/D-STM cannot
maintain diverse solutions over the entire PF
26Final result of MOEA/D-STM on MOP1
Stable Matching-Based Selection◦ E.g.
27
Subproblems’ preference lists Solutions’ preference lists
Stable matching-based selection result
Reason:Some part of PF can be located easily, while other part is not. All subproblems prefer solutions closer to the PF.
Outcome:Stable matching cannot ensure that the selected solutions are matched to their closest subproblems. Population diversity may be destroyed.
Contents◦ Introduction◦ Stable Matching-Based Selection◦ Stable Matching-Based Selection with Incomplete Lists◦ Stable Matching with Incomplete Preference Lists◦ Two-level one-one stable matching-based selection◦ Many-one stable matching-based selection
◦ Adaptive Stable Matching-Based Selection◦ Conclusion
28
Stable Matching with Incomplete Lists◦ Special case for a matching problem
◦ Stable matching with incomplete lists (STM-IC) [6]◦ A matching agent does not need to have a complete preference list over all
matching agents in the other sets◦ A matching agent can only be matched with one of the matching agents
present on its preference list
29
Some men or women refuse to be matched to some candidates. They just remove these candidates in their preference lists. A matching will never to made to these rejected candidates.
Stable Matching with Incomplete Lists◦ If we only keep a few most preferred subproblems on a solution’s
preference list, then◦ A solution is only allowed to be matched with its one of most preferred
subproblems ◦ We can ensure that the selected solutions are matched to its closest
subproblems after stable matching
30
Stable Matching with Incomplete Lists◦ E.g.
31
Subproblems’ preference lists Solutions’ preference lists
Stable matching with incomplete preference lists
The length of solution’s preference list is reduced to be 2. Only the two favorite subproblems remained.
Population diverse is ensured.
Stable Matching with Incomplete Lists
32Note: an SMP with incomplete lists cannot guarantee to match a stable solution for all subproblems.
Stable Matching with Incomplete Lists◦ A stable matching to an SMP with incomplete lists cannot
guarantee to match a stable solution for all subproblems◦ Two versions of stable matching-based selection mechanisms
with incomplete preference lists are developed in [9]◦ Two-level one-one matching-based selection ◦ Many-one matching-based selection
33
Two-Level One-One Matching-Based Selection
◦ Frist level (SMP-IC)◦ The preference list of a solution is set
to a limited length 𝑟𝑟◦ Only the 𝑟𝑟 most preferred subproblems
remain in solution’s preference list
◦ Second level (original SMP)◦ Since the stable matching of an SMP-IC
might not match a stable solution for all subproblems
◦ The remaining subproblems and solutions are matched using complete preference lists
34
Subproblems with complete preference list
Solutions with incomplete
preference list
First-level stable matching
Matched subproblems and solutions
Unmatched subproblems and solutions with complete preference list
Second-level stable matching
Two-Level One-One Matching-Based Selection
35
Two-Level One-One Matching-Based Selection
◦ MOEA/D-STM2L [7]◦ r is set to be 8 for UF test instances and
4 for MOP test instances◦ Achieves competitive performance on
MOP test instances◦ Inherits the main ability of MOEA/D-
STM on UF test instances
36
Many-One Matching-Based Selection◦ The selection process is modeled as a many-one matching
problem with a common quota [8] (college admission problem)◦ Each solution can only match with one subproblem◦ Each subproblem can match with more than one solutions without a limit◦ The common quota is set to be equal to 𝑁𝑁 (the number of subproblems)◦ At most 𝑁𝑁 solutions can be matched in a stable matching
◦ In such a way, the unmatched subproblems caused by the incomplete preference lists give the matching opportunities to other subproblems that have already been matched with a solution but still have openings.
37
Many-One Matching-Based Selection◦ Stable condition◦ A matching to a many-one matching problem with a common quota is called
stable if there does not exist any pair of subproblem 𝑝𝑝 and solution 𝐱𝐱, where:◦ 𝑝𝑝 and 𝐱𝐱 are acceptable to each other but not matched together;◦ 𝐱𝐱 is unmatched or prefers 𝑝𝑝 to its assigned subproblem;◦ the common quota is not met or 𝑝𝑝 prefers 𝐱𝐱 to at least one of its assigned
solutions.
38
Many-One Matching-Based Selection◦ Proposed many-one matching procedure◦ Main loop◦ Step 1:
◦ Randomly select an unmatched solution x; (line 7)◦ Match x with its current favorite subproblem 𝑝𝑝 according to its incomplete preference
list; (line 9-11)◦ Remove 𝑝𝑝 from 𝐱𝐱’s preference list; (line 10)
◦ Step 2:◦ If the number of current matching pairs 𝑀𝑀 > 𝑁𝑁, find a substitute subproblem 𝑝𝑝′ and
adjust its matching pairs by releasing the matching relationship with its least preferred solution 𝐱𝐱′. (line 12-18)
39
Many-One Matching-Based Selection◦ Proposed many-one matching procedure◦ The substitute subproblem 𝑝𝑝′ is selected according to the following criteria
1. Choose the subproblems that have the largest number of matched solutions to form �𝑃𝑃. Its underlying motivation is to reduce the chance for overly exploiting a particular subproblem; (line 13)
2. If �𝑃𝑃 > 1, the subproblems in �𝑃𝑃, whose least preferred solution holds the worst rank on that subproblem’s preference list, are used to reconstruct �𝑃𝑃; (line 14)
3. A substitute subproblem 𝑝𝑝′ is randomly chosen from �𝑃𝑃. (line 15)
40
Many-One Matching-Based Selection
41
Stable Matching-Based Selection with Incomplete Preference lists
◦ E.g., 𝑟𝑟 = 2
42Note: selection results of two-level one-one matching and many-one matching might not be the same.
Many-one stable matching-based selection resultTwo-level one-one stable matching-based selection result
Matched in the second level
Stable Matching-Based Selection with Incomplete Preference lists
◦ Steps◦ 𝐱𝐱1 proposes a matching request to 𝑝𝑝1 (|M|=1)◦ 𝐱𝐱2 proposes a matching request to 𝑝𝑝1 (|M|=2)◦ 𝐱𝐱3 proposes a matching request to 𝑝𝑝1 (|M|=3)◦ 𝐱𝐱4 proposes a matching request to 𝑝𝑝2 (|M|=4)◦ 𝐱𝐱5 proposes a matching request to 𝑝𝑝2 (|M|=5)
43
Stable Matching-Based Selection with Incomplete Preference lists
◦ Steps◦ 𝐱𝐱6 proposes a matching request to 𝑝𝑝2 (|M|=6)◦ |M|>5, find a substitute problem 𝑝𝑝′◦ Find subproblems with most matched solutions �𝑃𝑃 = 𝑝𝑝1, 𝑝𝑝2
◦ Find subproblems in �𝑃𝑃, whose least preferred solution holds the worst rank on that subproblem's preference list, to reconstruct �𝑃𝑃 = 𝑝𝑝2
◦ Randomly select a subproblem from �𝑃𝑃 to be 𝑝𝑝′ = 𝑝𝑝2
◦ Release the matching relationship of 𝑝𝑝2 with its least preferred solution 𝐱𝐱6(|M|=5)
44
Stable Matching-Based Selection with Incomplete Preference lists
◦ Steps◦ 𝐱𝐱7 proposes a matching request to 𝑝𝑝3 (|M|=6)◦ Release the matching relationship of 𝑝𝑝1 with its least preferred solution 𝐱𝐱3
◦ 𝐱𝐱8 proposes a matching request to 𝑝𝑝1 (|M|=6)◦ Release the matching relationship of 𝑝𝑝1 with its least preferred solution 𝐱𝐱8
45
Stable Matching-Based Selection with Incomplete Preference lists
◦ Steps◦ 𝐱𝐱9 proposes a matching request to 𝑝𝑝1 (|M|=6)◦ Release the matching relationship of 𝑝𝑝1 with its least preferred solution 𝐱𝐱9
◦ 𝐱𝐱10 proposes a matching request to 𝑝𝑝5 (|M|=6)◦ Release the matching relationship of 𝑝𝑝2 with its least preferred solution 𝐱𝐱5
46
Stable Matching-Based Selection with Incomplete Preference lists
◦ Steps◦ 𝐱𝐱3 proposes a matching request to 𝑝𝑝2 (|M|=6)◦ Release the matching relationship of 𝑝𝑝2 with its least preferred solution 𝐱𝐱4
◦ 𝐱𝐱4 proposes a matching request to 𝑝𝑝1 (|M|=6)◦ Release the matching relationship of 𝑝𝑝1 with its least preferred solution 𝐱𝐱4
◦ 𝐱𝐱5 proposes a matching request to 𝑝𝑝1 (|M|=6)◦ Release the matching relationship of 𝑝𝑝1 with its least preferred solution 𝐱𝐱5
47
Stable Matching-Based Selection with Incomplete Preference lists
◦ Steps◦ 𝐱𝐱6 proposes a matching request to 𝑝𝑝3 (|M|=6)◦ Release the matching relationship of 𝑝𝑝6 with its least preferred solution 𝐱𝐱7
◦ 𝐱𝐱7 proposes a matching request to 𝑝𝑝2 (|M|=6)◦ Release the matching relationship of 𝑝𝑝2 with its least preferred solution 𝐱𝐱7
◦ 𝐱𝐱8 proposes a matching request to 𝑝𝑝2 (|M|=6)◦ Release the matching relationship of 𝑝𝑝2 with its least preferred solution 𝐱𝐱8
◦ 𝐱𝐱9 proposes a matching request to 𝑝𝑝2 (|M|=6)◦ Release the matching relationship of 𝑝𝑝2 with its least preferred solution 𝐱𝐱9
48
Contents◦ Introduction◦ Stable Matching-Based selection◦ Stable Matching-Based Selection with Incomplete Lists◦ Adaptive Stable Matching-Based Selection◦ Conclusion
49
Adaptive Stable Matching-Based Selection
◦ Parameter sensitivity studies of 𝑟𝑟◦ For some test instances, performance is very sensitive to the setting of r◦ If 𝑟𝑟 is too large, the advantage of incomplete lists degenerates◦ If 𝑟𝑟 is too small, diversity is strengthened while the convergence is scarified
◦ 𝑟𝑟 is a problem dependent parameter
50
IGD performance of MOEA/D-STM2L on MOP test suite with different r settings [6]
Adaptive Stable Matching-Based Selection
◦ E.g.
51
Different settings of 𝑟𝑟 affect the selection result.Larger 𝑟𝑟, better convergence;Smaller 𝑟𝑟, better diversity.
Adaptive Stable Matching-Based Selection
◦ Example◦ 𝑝𝑝3 and 𝑝𝑝4 select solutions far away because 𝐱𝐱2 and 𝐱𝐱4 contributes smaller
objectives than solutions around them, e.g., 𝑝𝑝4 and 𝐱𝐱9 will be unstable as long as 𝐱𝐱2 and 𝐱𝐱4 are available
◦ We can assume that it is the better/competitive solutions near the PF that have a side effect on the diversity
◦ A straightforward way to overcome it is to reduce the preference list of a competitive solution to avoid it being matched with a subproblem far away
Subproblems’ preference lists Solutions’ preference lists Stable matching based selection result52
Adaptive Stable Matching-Based Selection
◦ Concept of the local competitiveness◦ At first, all solutions are associated with their closest subproblems. For each
subproblem, choose the one which offers the best objective value, as its representative solution.
◦ Given a neighborhood (defined as 𝐱𝐱’s ℓ closest subproblems), ◦ 𝐱𝐱 is defined as a locally competitive solution if there exists a representative
solution 𝐱𝐱(p) in this neighborhood dominated by 𝐱𝐱◦ ∃𝑝𝑝∈Neighbor(𝐱𝐱): 𝐱𝐱(p)≺𝐱𝐱
◦ Otherwise, solution x is denoted as a locally noncompetitive solution
53
Adaptive Stable Matching-Based Selection
◦ Adaptive Mechanism ◦ If solution 𝐱𝐱 is locally competitive and better than a representative solution𝐱𝐱(𝑝𝑝) in 𝐱𝐱’s neighborhood, then 𝐱𝐱 will be preferred by the subproblem 𝑝𝑝. When 𝐱𝐱 locates far from 𝑝𝑝, the diversity may be affected.
◦ To avoid 𝑝𝑝 appearing on 𝐱𝐱’s preference list ◦ The length of preference list 𝑟𝑟 of solution 𝐱𝐱 is set to be the maximum
neighborhood size ℓ that keeps 𝐱𝐱 locally noncompetitive◦ In such way, solution 𝐱𝐱 will not be matched with a subproblem whose
reprehensive solution is worse than 𝐱𝐱
54
Adaptive Stable Matching-Based Selection
◦ Example of determining 𝑟𝑟◦ Step 1: determine representative solutions◦ Solutions are associated to their closest subproblems ◦ Each subproblem select the best solution from its associated solutions◦ Representative solutions:
◦ 𝑝𝑝1: 𝐱𝐱1
◦ 𝑝𝑝2: 𝐱𝐱4
◦ 𝑝𝑝3: 𝐱𝐱7
◦ 𝑝𝑝4 does not have a representative solutions◦ 𝑝𝑝5: 𝐱𝐱10
55
◦ Example of determining 𝑟𝑟◦ Step 2: determine 𝑟𝑟 for each solution as the maximum neighborhood size
that keeps the solution locally noncompetitive◦ 𝐱𝐱1:
◦ When ℓ = 1, the neighborhood is 𝑝𝑝1, the included representative solution is 𝐱𝐱1. 𝐱𝐱1does not dominate 𝐱𝐱1, so 𝐱𝐱1 is locally noncompetitive.
◦ When ℓ = 2, the neighborhood is {𝑝𝑝1,𝑝𝑝2}, the included representative solution are {𝐱𝐱1, 𝐱𝐱4}. 𝐱𝐱1 dominates 𝐱𝐱4, so 𝐱𝐱1 becomes locally competitive.
◦ Set 𝑟𝑟1 = 1
◦ Similarly, set 𝑟𝑟𝟐𝟐 = 𝑟𝑟𝟑𝟑 = 1
Adaptive Stable Matching-Based Selection
56
neighborhood
◦ Example of determining 𝑟𝑟◦ 𝐱𝐱4:
◦ When ℓ = 1, the neighborhood is 𝑝𝑝2, the included representative solution is 𝐱𝐱4. 𝐱𝐱4 does not dominate 𝐱𝐱4, so 𝐱𝐱4 is locally noncompetitive.
◦ When ℓ = 2, the neighborhood is {𝑝𝑝2,𝑝𝑝1}, the included representative solution are {𝐱𝐱4, 𝐱𝐱1}. 𝐱𝐱4 does not dominate any solution in {𝐱𝐱4, 𝐱𝐱1}, so 𝐱𝐱4 is locally noncompetitive.
◦ When ℓ = 3, the neighborhood is {𝑝𝑝2,𝑝𝑝1, 𝑝𝑝3}, the included representative solution are {𝐱𝐱1, 𝐱𝐱4, 𝐱𝐱7}. 𝐱𝐱4 dominates 𝐱𝐱7, so 𝐱𝐱4 becomes locally competitive.
◦ Set 𝑟𝑟4 = 2
◦ Similarly, set 𝑟𝑟5 = 2
Adaptive Stable Matching-Based Selection
57
neighborhood
◦ Example of determining 𝑟𝑟◦ 𝐱𝐱6:
◦ When ℓ = 1, the neighborhood is 𝑝𝑝2, the included representative solution is 𝐱𝐱4. 𝐱𝐱6 does not dominate 𝐱𝐱4, so 𝐱𝐱6 is locally noncompetitive.
◦ When ℓ = 2, the neighborhood is {𝑝𝑝2,𝑝𝑝3}, the included representative solution are {𝐱𝐱4, 𝐱𝐱7}. 𝐱𝐱6 dominates 𝐱𝐱7, so 𝐱𝐱6 becomes locally competitive.
◦ Set 𝑟𝑟6 = 1
Adaptive Stable Matching-Based Selection
58
neighborhood
◦ Example of determining 𝑟𝑟◦ 𝐱𝐱7:
◦ When ℓ = 1, the neighborhood is 𝑝𝑝3, the included representative solution is 𝐱𝐱7. 𝐱𝐱7 does not dominate 𝐱𝐱7, so 𝐱𝐱7 is locally noncompetitive.
◦ When ℓ = 2, the neighborhood is {𝑝𝑝3,𝑝𝑝2}, the included representative solution are {𝐱𝐱7, 𝐱𝐱4}. 𝐱𝐱7 does not dominate any solution in {𝐱𝐱7, 𝐱𝐱4}, so 𝐱𝐱7 is locally noncompetitive.
◦ When ℓ = 3, the neighborhood is {𝑝𝑝3,𝑝𝑝2,𝑝𝑝4}, the included representative solution are {𝐱𝐱7, 𝐱𝐱4}. 𝐱𝐱7 does not dominate any solution in {𝐱𝐱7, 𝐱𝐱4}, so 𝐱𝐱7 is locally noncompetitive.
Adaptive Stable Matching-Based Selection
59
◦ When ℓ = 4, the neighborhood is {𝑝𝑝3,𝑝𝑝2,𝑝𝑝4,𝑝𝑝1}, the included representative solution are {𝐱𝐱7, 𝐱𝐱4, 𝐱𝐱1}. 𝐱𝐱7does not dominate any solution in {𝐱𝐱7, 𝐱𝐱4, 𝐱𝐱1}, so 𝐱𝐱7is locally noncompetitive.
◦ When ℓ = 5, the neighborhood is {𝑝𝑝3,𝑝𝑝2,𝑝𝑝4,𝑝𝑝1, 𝑝𝑝5}, the included representative solution are {𝐱𝐱7, 𝐱𝐱4, 𝐱𝐱1, 𝐱𝐱10}. 𝐱𝐱7 does not dominate any solution in {𝐱𝐱7, 𝐱𝐱4, 𝐱𝐱1, 𝐱𝐱10}, so 𝐱𝐱7 is locally noncompetitive.
◦ Set 𝑟𝑟7 = 5
◦ Similarly, set 𝑟𝑟𝟖𝟖 = 𝑟𝑟𝟗𝟗 = 𝑟𝑟𝟏𝟏𝟏𝟏 = 5
neighborhood
Adaptive Stable Matching-Based Selection
◦ Adaptive Mechanism ◦ Step 1:◦ Determine representative solutions;
(line 1-8)
◦ Step 2:◦ Gradually determine the maximum
noncompetitive neighborhood size of each solution 𝐱𝐱 and assign it to 𝑟𝑟; (line 9-16)
◦ Note that the search of 𝑟𝑟 is conducted within [𝑚𝑚, ℓ𝑚𝑚𝑚𝑚𝑚𝑚]
60
Adaptive Stable Matching-Based Selection
61
Adaptive Stable Matching-Based Selection
62
The tragectaries of adaptively determined 𝑟𝑟 values of the solutions selected by subproblems 𝑝𝑝1, 𝑝𝑝34, 𝑝𝑝67 and 𝑝𝑝100 during the run on MOP1.
Adaptive Stable Matching-Based Selection
63
Adaptive Stable Matching-Based Selection
64
Adaptive Stable Matching-Based Selection
65
Adaptive Stable Matching-Based Selection
66
Adaptive Stable Matching-Based Selection
67
Adaptive Stable Matching-Based Selection
68
Adaptive Stable Matching-Based Selection
69
Adaptive Stable Matching-Based Selection
70
Adaptive Stable Matching-Based Selection
71
Adaptive Stable Matching-Based Selection
72
Adaptive Stable Matching-Based Selection
73
Contents◦ Introduction◦ Stable Matching-Based Selection◦ Stable Matching-Based Selection with Incomplete Lists◦ Adaptive Stable Matching-Based Selection◦ Conclusion
74
Conclusion ◦ Conclusions ◦ The stable matching-based selection mechanism paves a new avenue to
address the balance between convergence and diversity from the perspective of achieving the equilibrium between the preferences of subproblems and solutions.
◦ With the help partial preference information, a two-level one-one stable matching-based selection mechanism and a many-one stable matching-based selection mechanism are proposed to address the diversity issue of original one-one stable matching.
◦ An adaptive mechanism is proposed to dynamically control the length of the preference list of each solution according to the local competitiveness information.
75
Conclusion ◦ Future work◦ Allow elite solutions to be matched with more than one subproblems ◦ Improve convergence
◦ Stable matching may be used for mating selection ◦ Preferences between solutions need to be well defined
76
References ◦ [1] S. Koulouridis, D. Psychoudakis, and J. L. Volakis, Multiobjective Optimal Antenna Design
Based on Volumetric Material Optimization. IEEE Trans. on Antennas and Propagation, vol. 55, no. 3, pp. 594-603, March 2007.
◦ [2] Q. Zhang, and H. Li, MOEA/D, A Multiobjective evolutionary algorithm based on decomposition. IEEE Trans. Evol. Comput., vol. 11, no. 6, pp. 712-731, Dec. 2007.
◦ [3] D. Gale, and L. S. Shapley, “College admissions and the stability of marriage,” Amer. Math. Mon., vol. 69, no. 1, pp. 9–15, 1962.
◦ [4] K. Li, Q. Zhang, S. Kwong, M. Li, and R.Wang, Stable matching-based selection in evolutionary multiobjective optimization. IEEE Trans. Evol. Comput., vol. 18, no. 6, pp. 909-923, Dec. 2014.
◦ [5] H. Liu, F. Gu, and Q. Zhang, Decomposition of a multiobjective optimization problem into a number of simple multiobjective subproblems. IEEE Trans. Evol. Comput., vol. 18, no. 3, pp. 450-455, Jun. 2014.
◦ [6] K. Iwama, D. Manlove, S. Miyazaki, and Y. Morita, Stable marriage with incomplete lists and ties. ICALP, vol. 99, pp. 443-452, July, 1999.
◦ [7] M. Wu, S. Kwong, Q. Zhang, K. Li, R. Wang and B. Liu, "Two-Level Stable Matching-Based Selection in MOEA/D," Systems, Man, and Cybernetics, 2015 IEEE International Conference on, pp. 1720-1725, Oct. 2015.
77
References ◦ [8] P. Bir¡äo, T. Fleiner, R. W. Irving, and D. Manlove, The college admissions problem
with lower and common quotas, Theor. Comput. Sci., vol. 411, no. 34-36, pp. 3136-3153, 2010.
◦ [9] M. Wu, K. Li, S. Kwong, Y. Zhou and Q. Zhang, Matching-Based Selection With Incomplete Lists for Decomposition Multiobjective Optimization, IEEE Trans. on Evol. Comput., vol. 21, no. 4, pp. 554-568, Aug. 2017.
78
79
Thank you