analysis of two algorithms for multi-objective min-max optimization
DESCRIPTION
Analysis of two algorithms for multi-objective min-max optimization. Simone Alicino Prof. Massimiliano Vasile Department of Mechanical and Aerospace Engineering University of Strathclyde , Glasgow, UK BIOMA 2014 13 th September 2014, Ljubljana, Slovenia. Design under uncertainty. bba. - PowerPoint PPT PresentationTRANSCRIPT
ANALYSIS OF TWO ALGORITHMS FOR MULTI-OBJECTIVE MIN-MAX OPTIMIZATIONSimone Alicino
Prof. Massimiliano Vasile
Department of Mechanical and Aerospace Engineering
University of Strathclyde, Glasgow, UK
BIOMA 2014
13th September 2014, Ljubljana, Slovenia
2
Model of the System
d, u f(d,u)
f
u
Bel/Pl
f
Design under uncertainty
u
aleatory
m1
m2
m3
u
bbaepistemic
3
Belief and Plausibility that f(u) < ?
Bel(f < ) = m(q1) + m(q2) = 0.4
Pl(f < ) = m(q1) + m(q2) + m(q3) = 0.8
q1 q2 q4
m(q1) = 0.3 m(q3) = 0.4 m(q4) = 0.2
q3
m(q2) = 0.1
Evidence theory
4
MACS2• Population-based search• Pareto ranking + Tchebycheff scalarization• Exploitation: sampling of neighborhood• Exploration: differential evolution• ArchiveIDEA• Population-based search• Hybrid DE + MBH• Improves local convergence• Avoids stagnation
CROSS-CHECKS• To rank and check and • Increase prob. of finding global maxima
Methodology
ν𝑚𝑎𝑥=mind∈𝐷 [maxu∈𝑈
𝑓 1 (d ,u ) ,maxu∈𝑈
𝑓 2 (d , u ) ,…,maxu∈𝑈
𝑓 𝑞 (d ,u ) ]MACS2ν𝑚𝑎𝑥=min
d∈𝐷[ 𝑓 𝑚𝑎𝑥
1 (d , u ) ,…, 𝑓 𝑚𝑎𝑥𝑞 (d ,u ) ]
d 𝑖
IDEA
5
Individualistic Actions
Subproblem selection
Initialization
Cross-Check
Cross-Check
Archive resize
Validation
Social Actions
All other individuals
Social individuals
Min-Max Selection
Cross-Check
Min-Max Selection
SOCIAL ACTIONS Child generated by interaction (DE) of agents with neighbours or global archive.
SUBPROBLEM SELECTIONupdate of the composition of
the social population and their associated scalar
subproblems.
GLOBAL ARCHIVEan external repository in
which non-dominated solutions are stored. The
archive is kept below a maximum size.
INDIVIDUALISTIC ACTIONS Child generated by random moves (pattern search) of each agent.
INITIALIZATIONInitial population is randomly
generated (LHS) in the search domain D.
MACS
6
MACS: Cross-check
f2
f1
𝑓 (d ,u )
𝑓 𝑎(d 𝑎 ,u 𝑎)
𝑓 12(d ,u 𝑎)
𝑓 21(d 𝑎 , u)
If agent in the population dominates or is dominated by the archive
7
MACS: Min-max selection
f
Uunewu
f
Dd dnew
otherwise
8
Run global optimization over U
until
MACS: Validation
f2
f1
𝑓 𝑙= 𝑓 𝑙
𝑓 2
9
SO GLOBAL MAXIMIZATION Performed by IDEA on u space, for each di solution of MO global minimization
ARCHIVE MAXIMAStore in U-archive solution of
IDEA only if it is better than solution of MACS2
(maximization might fail to find global optimum)
FINAL CROSS-CHECKLocal search to refine accuracy of U-archive
MO GLOBAL MINIMIZATION Performed by MACS2 on d space, and uses u’s stored in U-archive (internal cross-check).
INITIALIZATIONInitial population is randomly
generated (LHS) andU-archive is initialized.
MACSminmax
SO maximizationsIDEA
Archive min solutions
Archive max solutions
MO minimizationMACS2
Initialization
Cross-Check
Cross-Check
Dominance
10
MACSminmax: restoration
Archived maximum
Candidate minimum in d
Solution
Selected minimum in d
11
MACSminmaxMACSν
Individualistic Actions
Subproblem selection
Initialization
Cross-Check
Cross-Check
Archive resize
Validation
Social Actions
Min-Max Selection
Cross-Check
Min-Max Selection
SO maximizationsIDEA
Archive min solutions
Archive max solutions
MO minimizationMACS2
Initialization
Cross-Check
Cross-Check
Dominance
Comparison
Global vs. local search, same purpose: make sure that each d is associate to a global
maximum u
Local search vs. cross-checkfor every agent of the minimization
Both implement similar mechanisms to
increase probability of archiving global
maxima
12
Performance metrics
• Convergence
• Spreading
• Success rate
𝑀 𝑠𝑝𝑟=1𝑀𝑝
∑𝑖=1
𝑀𝑝
min𝑗∈𝑁 𝑝
100‖ 𝑓 𝑗−𝑔𝑖
𝑔𝑖‖
𝑀 𝑐𝑜𝑛𝑣=1𝑁𝑝
∑𝑖=1
𝑁 𝑝
min𝑗∈𝑀 𝑝
100‖𝑔 𝑗− 𝑓 𝑖𝑔 𝑗
‖
13
Test cases
Settings
MACS2• 200n function evaluations• 10 agents• 5 (1/2) social agents• F = 1• CR = 0.1
IDEA• 200n function evaluations• 5 agents• F = 1• CR = 0.1
14
Test case 1𝑓 1=∑
𝑖=1
𝑛=2
𝑑𝑖𝑢𝑖2
𝑓 2=∑𝑖=1
𝑛=2
(5−𝑑𝑖 ) (1+cos𝑢𝑖 )+(𝑑𝑖−1 ) (1+sin𝑢𝑖 )
Max f1 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 100% 100% 0.2 1.7 100 / 0.5 79 / 2
MACSminmax 100% 100% 0.2 1.3 100 / 0.5 100 / 2
15
Test case 2𝑓 1=∑
𝑖=1
𝑛=8
(𝑑𝑖−𝑢𝑖 )2
𝑓 2=∑𝑖=1
𝑛=8
(2𝜋 −𝑢𝑖 )cos (𝑢𝑖−𝑑𝑖 )
Max f1 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 100% 65% 0.5 16.1 100 / 1 0 / 2
MACSminmax 100% 60% 0.6 2.0 100 / 1 64 / 2
16
Test case 3𝑓 1=∑
𝑖=1
𝑛=8
(𝑑𝑖−𝑢𝑖 )2
𝑓 2=∑𝑖=1
𝑛=8
(𝑢𝑖−3𝑑𝑖 ) sin𝑢𝑖+(𝑑𝑖−2 )2
Max f1 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 100% 100% 0.6 7.5 46 / 0.5 3 / 2
MACSminmax 100% 100% 0.1 0.3 100 / 0.5 100 / 2
17
Test case 4𝑓 1=∑
𝑖=1
𝑛=2
(2𝜋−𝑢𝑖 )cos (𝑢𝑖−𝑑𝑖 )
𝑓 2=∑𝑖=1
𝑛=2
(𝑑𝑖−𝑢𝑖 ) cos (3𝑑𝑖−5𝑢𝑖 )
Max f1 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 100% 91.3% 0.3 0.9 83 / 0.5 97 / 2
MACSminmax 100% 85.7% 0.4 1.0 77 / 0.5 91 / 2
18
Test case 5𝑓 1=∑
𝑖=1
𝑛=4
(2𝜋−𝑢𝑖 )cos (𝑢𝑖−𝑑𝑖 )
𝑓 2=∑𝑖=1
𝑛=4
(𝑢𝑖−3𝑑𝑖 ) sin𝑢𝑖+(𝑑𝑖−2 )2
Max f1 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 98.6% 54.1% 1.2 5.8 48 / 1 60 / 6
MACSminmax 92.8% 87.6% 2.7 8.0 24 / 1 42 / 6
19
Test case 6𝑓 1=∑
𝑖=1
𝑛=1
(𝑑𝑖+𝑢𝑖 )cos [3𝑑𝑖−𝑢𝑖 (5|𝑑|+5 ) ]
𝑓 2=∑𝑖=1
𝑛=1
(𝑑𝑖−𝑢𝑖 )cos (3𝑑𝑖−5𝑢𝑖 )
Max f1 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 100% 100% 0.3 1.2 95 / 0.5 97 / 2
MACSminmax 100% 100% 0.3 2.0 91 / 0.5 63 / 2
20
Test case 7
Max f1 Max f2 Max f2 Mconv Mspr pconv / tconv pspr / tspr
MACS 100% 100% 95.3% 5.0 9.3 50 / 5 8 / 5
MACSminmax 100% 100% 98.3 4.6 2.1 66 / 5 100 / 5
𝑓 1=∑𝑖=1
𝑛=8
(𝑑𝑖−𝑢𝑖 )2
𝑓 2=∑𝑖=1
𝑛=8
(2𝜋 −𝑢𝑖 )cos (𝑢𝑖−𝑑𝑖 )
𝑓 2=∑𝑖=1
𝑛=8
(𝑢𝑖−3𝑑𝑖 ) sin𝑢𝑖+(𝑑𝑖−2 )2
21
Conclusions
• Worst-case design– Evidence Theory to model epistemic uncertainty– Maximization of Belief function: worst-case scenario design
• Two multi-objective algorithms– Cross-checks to increase probability to find global maximum– MACS: bi-level algorithm, modification of MACS2– MACSminmax: restoration methodology, works with any MO/SO algorithm
• Test cases– 6 bi- and 1 three- objective cases, with different dimensions and complexity– Global fronts identified, with good to excellent accuracy– Comparable performance between MACS and MACSminmax
• Limitations– Limited number of cases, objectives, and dimensions– Test suite: neither fronts, nor global maxima analytically known (difficult to
assess performance)
23
Belief and Plausibility that f(u) < ?
Evidence theory
24
1. Worst-case solution (Bel = 1) (best d that gives the minimum of the maxima of f over u)• Above this point the design
is certainly feasible given the current information.
1
2
Bel 3
Pl
Computational approach
2. Best possible solution (Pl = 0) • Below this point the design
is certainly not possible
3. Belief and Plausibility of every intermediate solution between best and worst• Trade-off curve
25
Results
26
Fronts