Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search...
Transcript of Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search...
![Page 1: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/1.jpg)
Thresheld Convergencein PSO, DE, and Other Search Methods
Stephen Chen and James Montgomery
![Page 2: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/2.jpg)
Overview
Optimization in multi-modal search spaces
A study of attraction basins
An introduction to thresheld convergence
Applications of thresheld convergence
An introduction to leaders and followers
Conclusions about search in multi-modal spaces
Current and future work
PSODE
ES
MPS
SA
![Page 3: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/3.jpg)
Optimization in Multi-Modal Search Spaces
![Page 4: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/4.jpg)
Point Search – Escape from Local Optima
![Page 5: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/5.jpg)
Point Search – Escape from Local Optima
![Page 6: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/6.jpg)
Point Search – Escape from Local Optima
![Page 7: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/7.jpg)
Point Search – Escape from Local Optima
![Page 8: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/8.jpg)
Point Search – Escape from Local Optima
• Simulated annealing
• Cannot reject all worsening moves – no exploration
• To converge to an optimum, must consider escaping it to
get to a better one
• Escaping local optima is expensive and difficult
![Page 9: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/9.jpg)
Point Search – Constant Motion
![Page 10: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/10.jpg)
Point Search – Constant Motion
![Page 11: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/11.jpg)
Point Search – Constant Motion
![Page 12: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/12.jpg)
Point Search – Constant Motion
![Page 13: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/13.jpg)
Point Search – Constant Motion
![Page 14: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/14.jpg)
Point Search – Constant Motion
![Page 15: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/15.jpg)
Point Search – Constant Motion
• Tabu Search, restart methods
• Don’t try to converge to an optimum, try to find as many
(good ones) as possible
• Ideal for unstructured search spaces
![Page 16: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/16.jpg)
Population Search – Global Competition
![Page 17: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/17.jpg)
Population Search – Global Competition
![Page 18: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/18.jpg)
Population Search – Global Competition
![Page 19: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/19.jpg)
Population Search – Global Competition
![Page 20: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/20.jpg)
Population Search – Global Competition
• PSO and DE
• Converge around the best solutions in the population
![Page 21: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/21.jpg)
Population Search – Local Competition
![Page 22: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/22.jpg)
Population Search – Local Competition
![Page 23: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/23.jpg)
Population Search – Local Competition
![Page 24: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/24.jpg)
Population Search – Local Competition
![Page 25: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/25.jpg)
Population Search – Local Competition
• Techniques (e.g. PSO and DE) with niching
• Produce many local optima
• Computationally expensive (exhaustive search)
![Page 26: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/26.jpg)
Rethinking Multi-Modal Optimization
• Stop thinking about local optima
− Escaping from local optima…
− Finding multiple local optima…
• Start thinking about attraction basins
− Exploration search among different attraction basins
− Exploitation search within the same attraction basin
• Distinct two-phase process
− Find the fittest attraction basin
− Find its local optimum
![Page 27: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/27.jpg)
Population Search
– Find Fittest Attraction Basin
![Page 28: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/28.jpg)
Population Search
– Find Fittest Attraction Basin
![Page 29: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/29.jpg)
Population Search
– Find Fittest Attraction Basin
![Page 30: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/30.jpg)
Population Search
– Find Fittest Attraction Basin
![Page 31: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/31.jpg)
Population Search
– Get Local Optimum in Fittest Attraction Basin
![Page 32: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/32.jpg)
Population Search
– Get Local Optimum in Fittest Attraction Basin
![Page 33: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/33.jpg)
Population Search – Attraction Basins
Phase 1 – Exploration
• Find the fittest attraction basin
• (Not necessarily the fittest search point)
Phase 2 – Exploitation
• Stay within the fittest attraction basin (no further exploration)
![Page 34: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/34.jpg)
Attraction Basins
![Page 35: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/35.jpg)
Definitions
• An attraction basin is all of the points around a local
optimum that lead to that optimum when greedy local
search is used
• The fitness of an attraction basin is the fitness of its local
optimum
![Page 36: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/36.jpg)
Attraction Basins
![Page 37: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/37.jpg)
Attraction Basins
![Page 38: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/38.jpg)
Attraction Basins
![Page 39: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/39.jpg)
A Two-Phase Search Process
The global optimum is the local optimum in the fittest
attraction basin
1. Find the fittest attraction basin (exploration)
2. Find the local optimum in this attraction basin
(exploitation)
![Page 40: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/40.jpg)
Finding the Fittest Attraction Basin
(Trivial Case)
![Page 41: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/41.jpg)
Finding the Fittest Attraction Basin
(Hard Case)
![Page 42: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/42.jpg)
Finding the Fittest Attraction Basin
(Typical Case – Early)
![Page 43: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/43.jpg)
Finding the Fittest Attraction Basin
(Typical Case – Late)
![Page 44: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/44.jpg)
A Detailed Study with Rastrigin
Attraction basins of known size, shape, and location
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 45: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/45.jpg)
A Detailed Study with Rastrigin
• Focus on exploration
− The ability to identify the fitter attraction basin based on the
fitness of random sample solutions
• Demonstrate the negative effects of concurrent
exploration and exploitation
− Comparing only random solutions is the Easy Case
− Comparing random solutions against a local optimum is the
Hard Case
![Page 46: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/46.jpg)
A Detailed Study with Rastrigin
Success of exploration goes down with exploitation
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 47: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/47.jpg)
A Detailed Study with Rastrigin
– Initial Random Solution
![Page 48: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/48.jpg)
A Detailed Study with Rastrigin
– Partially Optimized Solution
![Page 49: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/49.jpg)
A Detailed Study with Rastrigin
– Locally Optimal Solution
![Page 50: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/50.jpg)
A Detailed Study with Rastrigin
50% optimization makes exploration almost impossible
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 51: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/51.jpg)
Concurrent Exploration and Exploitation
– From Successful Exploration
![Page 52: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/52.jpg)
Concurrent Exploration and Exploitation
– To Unsuccessful Exploration
![Page 53: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/53.jpg)
A Detailed Study with Rastrigin
PSO and DE perform local optimization very early
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 54: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/54.jpg)
A Detailed Study with Rastrigin
PSO and DE stall quickly – no more exploration
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 55: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/55.jpg)
Thresheld Convergence
an introduction
![Page 56: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/56.jpg)
Optimization in Multi-Modal Search Spaces
Phase 1: Find the fittest attraction basin
• Concurrent exploration and exploitation interferes with exploration
• Eliminate exploitation during this phase
Thresheld convergence
• Convergence is “held” back by a threshold function
• Small, exploitative moves which are smaller than the threshold value are disallowed
• Create an initial phase of pure exploration
![Page 57: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/57.jpg)
Working Definitions
• A new search point in a new attraction basin is exploration
• A new search point in the same attraction basin is
exploitation
• The first search point in a new attraction basin is likely to
have a random/average fitness
• Exploitation within an existing attraction basin leads to
solutions of better-than-random/above-average fitness
![Page 58: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/58.jpg)
Normal Search – No Minimum Step
![Page 59: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/59.jpg)
Normal Search – Exploitation Possible
![Page 60: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/60.jpg)
Thresheld Convergence – Minimum Step
![Page 61: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/61.jpg)
Thresheld Convergence – No Exploitation
![Page 62: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/62.jpg)
Thresheld Convergence – Exploration Only
![Page 63: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/63.jpg)
A Simple Threshold Function
• Early applications used a polynomial decay function
• In effect, this is exhaustive search to find the size of
attraction basins
𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 = 𝛼 ⋅ 𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙 ⋅𝐹𝐸𝑠 − 𝑖
𝐹𝐸𝑠
𝛾
𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙
𝛼 ⋅ 𝑑𝑖𝑎𝑔𝑜𝑛𝑎𝑙
![Page 64: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/64.jpg)
Applications of Thresheld Convergence
in PSO, DE, SA, ES and Minimum Population Search
![Page 65: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/65.jpg)
PSO with Thresheld Convergence
• Original update
if f(x) < f(pbest)
pbest = x;
end if
• New update
if f(x) < f(pbest)
AND distance (x, pbest) > threshold
AND distance (x, lbest) > threshold
pbest = x;
end if
![Page 66: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/66.jpg)
PSO with Thresheld Convergence
pbestlbest
![Page 67: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/67.jpg)
PSO with Thresheld Convergence
pbestlbest
![Page 68: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/68.jpg)
PSO with Thresheld Convergence
• Eliminating search points within the threshold wastes FEs
• Despite inefficiency, performance improves in multi-modal
search spaces
• Successful proof of concept
![Page 69: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/69.jpg)
DE with Thresheld Convergence
• Modified update
if ||new - base|| < threshold
direction = Normalize(new – base)
new = base + threshold direction
end if
![Page 70: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/70.jpg)
DE with Thresheld Convergence
base
new
![Page 71: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/71.jpg)
DE with Thresheld Convergence
base
new'
new
![Page 72: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/72.jpg)
DE with Thresheld Convergence
• No wasted FEs
• Minimum step size both improves exploration (in multi-
modal search spaces) and prevents premature
convergence (in all search spaces)
• Improved performance across a broad range of
benchmark functions
![Page 73: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/73.jpg)
SA with Thresheld Convergence
Create “hollow” sampling distribution
Images from S. Chen, C. Xudiera, and J. Montgomery, “Simulated Annealing with Thresheld Convergence”, IEEE CEC, 2012.
![Page 74: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/74.jpg)
Original Simulated Annealing
![Page 75: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/75.jpg)
SA with Thresheld Convergence
![Page 76: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/76.jpg)
SA with Thresheld Convergence
• Local search/optimization makes escaping from the
current attraction basin more difficult
• With threshold, search becomes almost random
− Concurrent exploration and exploitation is required for
useful convergence in a memory-less search system
− E.g. simulated annealing like a real physical system,
genetic algorithms like a real biological system
• SA with thresheld convergence requires a memory
![Page 77: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/77.jpg)
ES with Thresheld Convergence
• Adaptive σ optimizes rate of convergence
• In a multi-modal search space, σ will shrink into a single
attraction basin
• Need to balance both threshold size and σ
![Page 78: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/78.jpg)
ES with Thresheld Convergence
Maintain step size/prevent collapsing σ
Images from A. Piad-Morffis, S. Estevez-Velarde, A. Bolufe-Rohler, J. Montgomery, and S. Chen, “Evolution Strategies with Thresheld Convergence”, IEEE CEC, 2015
![Page 79: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/79.jpg)
ES with Thresheld Convergence
• Adaptive σ optimizes rate of convergence (in a unimodal
search space)
• Threshold size can be matched to size of attraction basins
• Leads to large performance improvements in multi-modal
search spaces
![Page 80: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/80.jpg)
Minimum Population Search
• New metaheuristic is designed to use the minimum
population size of two (in two dimensions)
− A population size of one is indistinguishable from point
search
• Small population size should improve scaling to higher
dimensional search spaces
• With two populations members,
− Can use gradient/difference vector for intensification
− Use thresheld convergence for diversification
![Page 81: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/81.jpg)
Minimum Population Search
In-line step of intensification is balanced with orthogonal
step for diversification. Threshold specifies overall step size.
Image from A. Bolufe-Rohler and S. Chen, “Minimum Population Search – Lessons from building a heuristic technique with two population members”, IEEE CEC, 2013
![Page 82: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/82.jpg)
Minimum Population Search
• MPS performs well in multi-modal search spaces
• Small population size leads to very efficient performance
• Performance advantage increases when FEs are limited
![Page 83: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/83.jpg)
Early Applications of Thresheld Convergence
• Lots of good results in the targeted multi-modal search
spaces
• Results limited by the simple, preliminary threshold
function
![Page 84: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/84.jpg)
Improvements to Thresheld Convergence
• Threshold size should match the size of attraction basins
• Decay function is exhaustive search for the size of
attraction basins
• Non-zero threshold size during exploitation phase is
inefficient
![Page 85: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/85.jpg)
Step-based Threshold Functions
• More time at ideal threshold size
• More efficient exploitation
![Page 86: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/86.jpg)
Current and Future Work
• Step-based threshold functions need explicit
measurement for the size of the attraction basins
• Size of attraction basins in Rastrigin varies from 1 to √n
depending on orientation of the line segment
• Experimenting with multiple options
1
√n
1
√n
![Page 87: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/87.jpg)
Limitations of Thresheld Convergence
• Ideal case assumes same relative fitness for each sample
solution
• Random variations naturally occur in the relative fitness of
sample solutions
• Random variations lead to random search
− Relative fitness improves over time
− Even when threshold size is larger than attraction basins
− Even when no exploitation is possible
• Exploration with thresheld convergence still stalls
![Page 88: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/88.jpg)
Ideal Case
![Page 89: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/89.jpg)
Typical Case
![Page 90: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/90.jpg)
Typical Case
![Page 91: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/91.jpg)
Limitations of Thresheld Convergence
Random search leads to improved relative fitness
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 92: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/92.jpg)
Limitations of Thresheld Convergence
Better relative fitness means less effective exploration
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 93: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/93.jpg)
Leaders and Followers
Towards a New Metaheuristic
![Page 94: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/94.jpg)
Towards a New Metaheuristic
• Exploration is only effective for a very few number of
iterations
• Need rapid restarts
− Each restart renews opportunity for effective exploration
• Goal of optimization is to find better solutions
− Better solutions reduce effectiveness of exploration
• Solution: two populations
− Leaders: best solutions which guide the search
− Followers: new solutions which can explore effectively
![Page 95: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/95.jpg)
Leaders and Followers
• A new metaheuristic explicitly developed for multi-modal
search spaces
• Population control scheme designed to promote accurate
comparison of attraction basins
![Page 96: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/96.jpg)
Leaders
The best solutions found to date
Guide the search
• Like pbests in PSO which guide/attract the current solutions
Do not evaluate the search
• Unlike PSO, new solutions are not compared to leaders/pbests
![Page 97: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/97.jpg)
Followers
The current solutions looking for better attraction basins
Cannot be compared with leaders
• New search points are likely to have poor relative fitness
• Direct comparison with leaders/solutions with high relative fitness will lead to inaccurate comparisons
Compare with other followers
• After each restart, followers can have similar relative fitness
![Page 98: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/98.jpg)
Leaders and Followers
• Leaders lead
• Followers follow
• How can search progress unless followers can become
leaders?
• When can leaders and followers be compared fairly?
• Ideally, when both populations have similar relative fitness
• Practically, when both populations have the same median
fitness
![Page 99: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/99.jpg)
Leaders and Followers
Initialize Leaders (to random solutions)
Loop allowed iterations/function evaluations
Initialize Followers to random solutions
Loop until Followers have similar relative fitness
Each follower moves towards a leader
Each follower is best of current or previous position
End Loop
Select new Leaders
End Loop
Return best leader
![Page 100: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/100.jpg)
Leaders and Followers
Relative fitness of pbests in PSO improves rapidly
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 101: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/101.jpg)
Leaders and Followers
Relative fitness of leaders also improves rapidly
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 102: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/102.jpg)
Leaders and Followers
Search stalls in PSO, does not stall in Leaders and
Followers – followers are not compared against leaders
Image from Y. Gonzalez-Fernandez and S. Chen, “Leaders and Followers – A New Metaheuristic to Avoid the Bias of Accumulated Information”, IEEE CEC, 2015.
![Page 103: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/103.jpg)
Leaders and Followers
• Followers reset every time their median fitness reaches
the median fitness of the Leaders
• Each reset allows unbiased comparison of attraction
basins
• Improved comparisons lead to improved performance in
multi-modal search space
• Rapid restarts are implicit thresheld convergence
− No threshold function used to encourage diversity
![Page 104: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/104.jpg)
Future Work for Leaders and Followers
• Add explicit thresheld convergence
• Better exploration in Followers
− Don’t create followers in the same attraction basins
• Better diversity in Leaders
− Each leader in its own attraction basin
![Page 105: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/105.jpg)
Thresheld Convergence
and what we have learned from its development
![Page 106: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/106.jpg)
Optimization in multi-modal spaces
should be two distinct phases
Thresheld convergence focuses on attraction basins in multi-modal search spaces
Optimization should be a two-phase process:
• Phase 1: Exploration find the fittest attraction basins
• Phase 2: Exploitation find its local optimum
Accuracy of comparing attraction basins is affected by the relative fitness of their sample solutions
Goal of thresheld convergence is to maintain similar relative fitness of all search solutions
![Page 107: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/107.jpg)
Leaders and followers is a new metaheuristic designed explicitly for multi-modal search spaces
Key focus is on how solutions are compared
• Avoid comparison of solutions with dissimilar relative fitness
Most metaheuristics focus on how solutions are created
• PSO attraction vectors
• DE difference vectors
• UMDA distributions
Improvements to how solutions are created should lead to large performance gains in leaders and followers
Metaheuristic design should focus on how
solutions are compared
![Page 108: Thresheld Convergence - IEEE CEC 2015 Tutorial · Overview Optimization in multi-modal search spaces A study of attraction basins An introduction to thresheld convergence Applications](https://reader035.fdocuments.us/reader035/viewer/2022071218/6051357746981012b64dec42/html5/thumbnails/108.jpg)
Summary paper
A Review of Thresheld Convergence
Stephen Chen, James Montgomery, Antonio Bolufé-Röhler,
Yasser Gonzalez-Fernandez
GECONTEC: Revista Internacional de Gestión del
Conocimiento y la Tecnología, Vol. 3(1). 2015
http://www.upo.es/revistas/index.php/gecontec/article/view/1410