1. Introduction
Background
Complex
Networks
Newscast
protocol
Model Design Peer-to-Peer Evolutionary Computation
The Evolvable
Agent
Model A Study of Viability
Properties
Experimental
Analysis
Goals
Methodology Juan Luis Jim´nez Laredo
e
Analysis of
Results
Test-Case 1 Dpto. Arquitectura y Tecnolog´ de Computadores
ıa
Test-Case 2
Test-Case 3 Universidad de Granada
Conclusions
27 de Mayo 2010
1 / 44
2. Scope
Introduction
Background
Complex
Networks
Newscast
protocol
Model Design
The Evolvable
• Status: Peer-to-Peer Evolutionary Computation (P2P EC)
Agent
Model represents a parallel solution for hard problems
Properties
Experimental
optimization
Analysis
Goals
• Objective: Find empirical evidences showing the viability
Methodology
Analysis of of the P2P EC paradigm
Results
Test-Case 1
Test-Case 2
• Modelling: Fine grained parallel EA using a P2P protocol
Test-Case 3
as underlying population structure
Conclusions
2 / 44
3. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
3 / 44
4. Introduction
Introduction
Background EAs: Bio-inspired population based optimization methods
Complex
Networks
Newscast
protocol
Model Design
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
4 / 44
11. Introduction
Introduction
Background
Complex
Networks
Newscast
protocol P2P EC
Model Design • Virtualization:
The Evolvable
Agent
Model
Single view at
Properties
application level
Experimental
Analysis • Decentralization:
Goals
Methodology
Analysis of
No central
Results
Test-Case 1
management
Test-Case 2
Test-Case 3 • Massive Scalability:
Conclusions Up to thousands of
computers
7 / 44
12. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
8 / 44
13. Population Structure as a complex network
Introduction
Background
Complex
Networks
Newscast
Watts-Strogatz Model
protocol
• Easy model for constructing small-world networks
Model Design
The Evolvable • Begins with a ring
Agent
Model • Rewired edges at random with a probability p
Properties
Experimental
Analysis
Goals
Methodology
p = 0: Ring p = 0.2: Small-world p = 1: Random
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
9 / 44
14. Population Structure as a complex network
Introduction
Panmictic Small-world Regular lattice
Background
Complex
Networks
Newscast
protocol
Model Design
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
10 / 44
15. Population Structure as a complex network
Introduction
Panmictic Small-world Regular lattice
Background
Complex
Networks
Newscast
protocol
Model Design
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
10 / 44
16. Population Structure as a complex network
Introduction
Panmictic Small-world Regular lattice
Background
Complex
Networks
Newscast
protocol
Model Design
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
10 / 44
17. Population Structure as a complex network
Introduction
Panmictic Small-world Regular lattice
Background
Complex
Networks
Newscast
protocol
Model Design
The Evolvable
Agent
Model n(n−1)
Properties
2
log(n) n
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
10 / 44
18. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
11 / 44
19. Newscast
Introduction
Background Basic Working Principles
Complex
Networks • Decentralized P2P protocol
Newscast
protocol • Every node has a cache acting as a routing table
Model Design • Dynamical self-organized network
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
Jelasity,02
12 / 44
20. Newscast
Introduction
Background Basic Working Principles
Complex
Networks • Decentralized P2P protocol
Newscast
protocol • Every node has a cache acting as a routing table
Model Design • Dynamical self-organized network
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
Jelasity,02
12 / 44
21. Newscast: Bootstrapping and Convergence
Introduction
Background
Complex
Networks
Experiment
Newscast
protocol
• Different network initializations: Watts-Strogatz and Random
Model Design • Network characterization: Average path length, Clustering coefficient
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
Jelasity,02
13 / 44
22. Newscast: Robustness
Introduction
Background
Complex
Networks
Experiment
Newscast
protocol
• System degradation: Up to 100%. Newscast, Random graph
Model Design • Network characterization: Size of largest cluster, Number of partitions
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
Jelasity,02
14 / 44
23. Newscast: Scalability
Introduction
Background
Complex
Networks
Experiment
Newscast
protocol • System traffic: Sizes of networks 1000 and 10000
Model Design • Network characterization: Probability of requests to a node
The Evolvable
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
Jelasity,02
15 / 44
24. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
16 / 44
25. The Evolvable Agent Model
Introduction
Background Design principles
Complex
Networks • Agent based approach
Newscast
protocol • Fine grain parallelization
Model Design
The Evolvable
• Spatially structured EA
Agent
Model
• Local selection
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
17 / 44
26. The Evolvable Agent Model
Introduction
Background Design principles
Complex
Networks • Agent based approach
Newscast
protocol • Fine grain parallelization
Model Design
The Evolvable
• Spatially structured EA
Agent
Model
• Local selection
Properties
Experimental
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
17 / 44
27. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
18 / 44
28. Multi-threading performance on a local computer
Introduction
Background Experiment
Complex
Networks
Newscast
• Scalability of virtual nodes running in a SMP desktop computer
protocol
• Fitness evaluation cost Tf ∈ [0.01 . . . 1] seconds
Model Design
The Evolvable
• Test-bed: 1 processor machine and a dual-core processor machine
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology T hroughputEA = evaluations
time
Analysis of
Results
Test-Case 1
T hroughput
Test-Case 2 Speedup = T hroughput EvAg
Test-Case 3 sequential
Conclusions
T imesequential
Speedup = T imeEvAg
Linear speedup up to the number of processors
19 / 44
29. Parallel performance on a P2P infrastructure
Introduction
Background Experiment
Complex
Networks
Newscast
• Scalability for a evaluation cost of Lζ=[1.5,2,3]
protocol
• L ∈ [1 . . . 100]
Model Design
The Evolvable
• N = L2 then N ∈ [1 . . . 10000]
Agent
Model
Properties
Experimental
Analysis
Goals
Methodology
Analysis of N Tf
Results Speedup = Tp
Test-Case 1
Test-Case 2
Test-Case 3 Tp = Tf + Tcomm + Tlat
Conclusions
Gagn´, 03
e
Linear speedups for demanding evaluation functions
20 / 44
30. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
21 / 44
31. Goals and Test-Cases
Introduction
Background
Complex
Networks
Goals
Newscast
protocol 1 Scalability: Suitability for tackling large problem
Model Design instances.
The Evolvable
Agent
Model 2 Fault-tolerance: Suitability for tolerating the system
Properties
Experimental
degradation.
Analysis
Goals
Methodology
Analysis of
Test-Cases
Results
Test-Case 1 • Test-Case 1: Scalability against canonical approaches in
Test-Case 2
Test-Case 3 failure-free environments
Conclusions
• Test-Case 2: Scalability against other population
structures in failure-free environments
• Test-Case 3: Fault-tolerance of the model under churn
22 / 44
32. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
23 / 44
33. Generalised l-trap function
Introduction
Background
Complex
Networks
Newscast
protocol
Model Design • l-trap function (Ackley,
The Evolvable
Agent 1987):
Model
Properties • 2-trap: not-deceptive
Experimental • 3-trap: partially
Analysis
Goals deceptive
Methodology
Analysis of
• 4-trap: fully deceptive
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
• L = 12 . . . 60
24 / 44
34. Experimental settings
Introduction
Background
Complex
Networks
Newscast
protocol
Model Design
• Population size
The Evolvable
Agent
• Estimated by bisection
Model
Properties
• Selectorecombinative
Experimental GA (Mutation less)
Analysis • Minimum population
Goals
Methodology size able to reach 0.98
Analysis of
Results of SR
Test-Case 1
Test-Case 2
Test-Case 3
• Uniform Crossover
Conclusions • Binary Tournament
25 / 44
35. Outline
Introduction
1 Introduction
Background
Complex
Networks
2 Background
Newscast
protocol Complex Networks
Model Design Newscast protocol
The Evolvable
Agent
Model 3 Model Design
Properties
Experimental
The Evolvable Agent
Analysis Model Properties
Goals
Methodology
Analysis of
4 Experimental Analysis
Results
Test-Case 1 Goals
Test-Case 2
Test-Case 3 Methodology
Conclusions Analysis of Results
Test-Case 1
Test-Case 2
Test-Case 3
5 Conclusions
26 / 44
53. Test-Case 3: Fault-tolerance
Introduction
Background Settings
Complex
Networks
Newscast Equally parameterized approaches with and without churn
protocol
Model Design
Problem instance: L=36 4-trap
The Evolvable
Agent
Pop. Size: 600
Model
Properties
Max. Eval: 393000
1
Experimental Mutation: Bit-flip Pm = L
Analysis
Goals
Methodology
Analysis of
Results
Test-Case 1
Test-Case 2
Test-Case 3
Conclusions
42 / 44
54. Conclusions
Introduction
Background Selected publications
Complex Peer reviewed journal papers :
Networks
Newscast
protocol 1 J.L.J. Laredo, A.E. Eiben, M. van Steen, and J.J. Merelo. Evag: A scalable
peer-to-peer evolutionary algorithm. GPEM, 2010.
Model Design http://dx.doi.org/10.1007/s10710-009-9096-z.
The Evolvable
Agent 2 J.L.J. Laredo, P.A. Castillo, A.M. Mora, J.J. Merelo, and C. Fernandes.
Model Resilience to churn of a peer-to-peer evolutionary algorithm. IJHPSA,
Properties 1(4):260-268, 2009.
Experimental 3 J.L.J. Laredo, P.A. Castillo, A.M. Mora, and J.J. Merelo. Evolvable agents, a
Analysis fine grained approach for distributed evolutionary computing. Soft
Goals Computing, 12(12):1145-1156, 2008.
Methodology
Peer reviewed conference papers and book chapters :
Analysis of
Results
Test-Case 1 1 J.L.J. Laredo, P.A. Castillo, A.M. Mora, J.J. Merelo, A.C. Rosa, and C.
Test-Case 2 Fernandes. Evolvable agents in static and dynamic optimization problems. In
Test-Case 3 PPSN X, pages 488-497. Springer, 2008
Conclusions 2 J.L.J. Laredo, A.E. Eiben, M. van Steen, P.A. Castillo, A.M. Mora, and J.J.
Merelo. P2P evolutionary algorithms: A suitable approach for tackling large
instances in hard optimization problems. In Euro-Par’ 08, pages 622-631.
Springer, 2008.
3 J.L.J. Laredo, P.A. Castillo, A.M. Mora, and J.J. Merelo. Exploring
population structures for locally concurrent and massively parallel
evolutionary algorithms. In IEEE WCCI2008 Proceedings, pages 2610-2617.
IEEE Press, Hong Kong, June 2008.
43 / 44