implementation report: concurrent genetic algorithm with

1/25

Genetic Algorithm revisited Genetic Algorithm Operators Concurrency Some Results Future Work

Implementation Report:Concurrent Genetic Algorithm

with Island Migration

Markus Solbach

Laboratory for Active and Attentive VisionDepartment of Computer Science and Engineering

York University, Toronto, Ontario, Canada

November 10, 2015

Markus Solbach York University

Implementation Report: Concurrent Genetic Algorithm with Island Migration

2/25


Overview

Genetic Algorithm revisited

Genetic Algorithm Operators

Concurrency

Some Results

Future Work Figure :Evolution ‖i.livescience.com (Oct. 5. 15)



3/25


Genetic Algorithm revisited



4/25


Genetic Algorithm revisitedGeneration 0



5/25


Genetic Algorithm revisitedGeneration n



6/25


VLSI Design Problem

Figure :VLSI Design Problem



7/25


VLSI Design Problem - Traveling Salesman Problem

Problem changed for good reasons:

VLSI very little implementation details

VLSI faced problems had nothing to do with concurrency

TSP well known problem

TSP rich implementation details in literature (sequential)

TSP able to concentrade more on concurrency



8/25


Traveling Salesman Problem

Find a route on a mapRequirements:

I Visit each City only once

I Find shortest path

Complexity (20 city):

I O(n!) (Brute Force, WC)

I 20! = 2.432902x1018

Figure :TSP Map



9/25


Genetic Algorithm Operators



10/25


Initialization

Figure :High Level Individual Representation



11/25


Mutation

Figure :High Level Mutation Representation



12/25


Crossover

Figure :High Level Crossover Representation



13/25


Fitness Function

I Travelled distance over all city (inverse fitness)

I Each City has a location (x , y)

I Euclidean distance

I Fitness1 =∑n

i=1(xi−1 − xi )2 + (yi−1 − yi )

2



14/25


Concurrency



15/25


Concurrent Random Number Generator

Genetic Algorithms rely heavily on random numbers

I Math.random() is not concurrent

I Multiple threads use similar or same seeds

I ThreadLocalRandom 1

I Generator with an internally generated seed

I java.util.concurrent.ThreadLocalRandom

1http://docs.oracle.com/javase/7/docs/api/java/util/

concurrent/ThreadLocalRandom.html



http://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ThreadLocalRandom.html

http://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ThreadLocalRandom.html

16/25


Threads and Islands

Each Thread represents one Island

I Genetic Algorithm (GA) Logic

I Implements Runnable

I Concurrent Execution

= GA’s Island Migration extension

Figure :Simplified UML-Class of CGaimIsland



17/25


Threads and Islands

Each Thread represents one Island

I Genetic Algorithm (GA) Logic

I Implements Runnable

I Concurrent Execution

= GA’s Island Migration extension

Figure :Island Migration overview



18/25


Threads and Barriers

Barriers2 for synchronization

I new CyclicBarrier(# Islands)

I Waits that all Islands are ready

I Evolution ↪→ random process

2http://docs.oracle.com/javase/7/docs/api/java/util/

concurrent/CyclicBarrier.html



http://docs.oracle.com/javase/7/docs/api/java/util/concurrent/CyclicBarrier.html

http://docs.oracle.com/javase/7/docs/api/java/util/concurrent/CyclicBarrier.html

19/25


Island Migration

Sequentially (as proposed).

I Dependens on epoch length

I Joins all Threads

I Avoids shared memory access

I Performs cyclic Migration

I Copies Individuals

I Java’s array.clone()[i ]



20/25


Island Migration

Sequentially (as proposed).

I Dependens on epoch length

I Joins all Threads

I Avoids shared memory access

I Performs cyclic Migration

I Copies Individuals

I Java’s array.clone()[i ]



21/25


Some Results



22/25


Some Results (Island Migration)

Figure :4 Islands - 150 Individuals - 2 Migrants - 70 Generations Epoch (≈ 5 Sec.)



23/25


Some Results (Island Migration vs. Sequential)

Figure :100 City - Sequential (150 Individuals) vs. 4 Islands (as before)



24/25


Future Work



25/25


Future WorkI More tests / Parameter improvements

I intel i7 (8 cores x 4GHz) using 8 IslandsI 5 times faster than sequential GA on same machine

I More debuggingI bugs in MigrationI lack of .clone()I ...

I Execution time differences based on number of Islands



implementation report: concurrent genetic algorithm with

Documents