Mary Beth Kurz, PhD

1
Utilizing Condor to Support Genetic Algorithm
Design Research

• Mary Beth Kurz, PhD
• Assistant Professor
• Department of Industrial Engineering
• Clemson University

2
Genetic algorithms are metaheuristics for
optimization
evaluation
Feasible solutions
Infeasible solutions
Chromosome 1
My research focuses on what chromosomes should
look like and asks whether the solution
representation impacts the quality of a genetic
algorithm
decoding
Chromosome 2
Chromosome N
Chromosome space
3
Lets make this more concrete the TSP
Each feasible solutions total travel time is
that solutions objective
Chromosome 1
The Traveling Salesman Problem asks how to route
the salesman through his cities so he returns
home as quickly as possible
decoding
Chromosome 2
Chromosome N

• How do I represent the tours?
• Directly the city list?
• Indirectly the list of roads taken?

Chromosome space
4
My Hypothesis Solution representation affects GA
design significantly
No chromosomes map to these solutions
Feasible solutions
Optimal objective value
Infeasible solutions
This is the optimal solution!!!
Chromosome 1
Chromosome 2
Fix or forbid these chromosomes?
Chromosome space
5
Genetic algorithms are motivated by an analogy to
real genetics
A chromosome is composed of genes, generally
randomly selected initially
Genetic Operators Randomness comes here
Selection picks some chromosomes as potential
parents in crossover
Crossover creates new chromosomes by taking genes
from 2 parents
Mutation changes a small number of genes in the
entire population
6
This research is empirical and requires immense
computational time

• Genetic Algorithms are inherently random
• Is it possible that some representation
  consistently finds better solutions for a
  specific problem?
• Most GA research currently uses 50 replications
  on numerous data files
• 180 problem types, 10 files each, 3
  representations 5400 files
• Simplest representation 1800 files would take
  about 45 hours in my Lab (a few years ago)
• 50 replications of 5400 files ? at least 241 days
  of running time!
• This is simply not feasible

7
Grid computing is saving me
Spring 2007 325,000 hrs Summer Fall
2007 212,000 hrs Spring 2008 124,000 hrs
Total about 660,000 hrs
8
Since last spring, Ive had to relearn how to do
research!

• How do I compile all this data?
• VBA and Excel!
• What can I actually analyze?
• Not pictures like this
• Reduce the data to correlations
• What statistics do I need to use?
• Needed to learn non-parametric statistics
• Needed to use SPSS for the analysis
• Used VBA to create the input files
• Reran to get different output data
• Summer Fall 2007 212,000 hrs

9
I dont know about random numbers

• I started using rand() in C!
• I use up to 600,000,000 random numbers in each
  run
• I have 270,000 runs (5400 50)
• Trying to use Mersenne Twister
• Period is 219937 1 which is plenty big
• How do I make sure I have independent sets of
  random numbers?
• Use the same initial seed, then burn through
  (n-1) numbers until we get to the nth set
• Would take over 4000 days to burn through 269,999
  sets for the last run
• Again not feasible!
• Tried to initialize using run number
• Spring 2008 124,000 hrs

10
I still love Condor

• But I dont know about random numbers
• Thought about saving the random numbers in an
  input file of 600,000,000 numbers each
• Stopped generating the first file after it got to
  3 GB
• This would mean 3270,000 GB of random number
  files!
• Looking at dynamic streams from Mersenne Twister
• Just heard about SPRNG from Todd on Tuesday
• Gearing up for another set of runs all I need
  is this set of runs to get a paper out!