A NSGAII, Web Enabled, Parallel Optimization Framework for NLP and MINLP

1
(No Transcript)
2
A NSGA-II, Web Enabled, Parallel Optimization
Framework for NLP and MINLP
IT Administrator
Designer/End User

• David J. Powell
• Joel Hollingsworth
• Elon University

Application Programmer
Optimization Expert
3
Overview

• Problem Making Optimization a common work place
  tool for today and tomorrow
• Pragmatic Solution Component based framework
  using industry standards
• Growing Usage In Industry
• Bench Marks / Results
• Project Status

4
Problems with Optimization

• Modeling
• Cost
• Optimizer Selection and Tuning
• Performance
• No Standards Based Development.

5
Modeling Solution

• Use AMPL for Optimization Formulation.
• Single or multiple objectives min or max
• Continuous or mixed integer design variables
• Bounds or constraints can be one sided or two
  sided.
• Mathematical language to express in intuitive
  fashion
• Extendable with user supplied C functions
• Interpretive (iterative exploration)

6
Low Cost Solution

• Use widespread free tools AMPL, NSGA-II, Apache
  Axis (Tomcat), SGE, Java.
• Supports industry standards Web Services, DRMAA
• Large newsgroups for support
• Documentation is wide spread
• Works on heterogeneous platforms

7
Optimizer Selection and Tuning Solution

• Single algorithm/implementation NSGA-II
• Supports multiple objectives (min and max)
• Supports mixed integer design variables.
• No landscape assumptions.
• Minimal tuning parameters.
• Easily supports parallel evaluation.
• Other world class algorithms for plug and play

8
Performance Solution

• Moores Law
• Continue through 2011 - consistent with future.
• Cost of a single design evaluation in 1990 would
  in 2007 equal 10 generations of population size
  50.
• SGE
• Coarse grained, heterogeneous, parallel
  evaluation
• Supports fine grained parallel evaluation
• DRMAA 1.0
• C and java support
• Is cost of GA really much greater than cost of
  finite differences and branch and bound methods?

9
Standards Based, Mainstream Solution

• AMPL coupling with NSGA-II with sequential
  evaluation (C)
• AMPL coupling with NSGA-II with parallel
  evaluation (DRMAA - SGE)
• Java Class Wrapper for AMPL
• Web Service for AMPL - SOA

10
Single Machine Sequential View
Command promptgt ampl model file
11
Multiple Machines Parallel Evaluation
12
Classic Example
Page 534 Engineering Optimization Theory and
Practice
13
NSGA-II AMPL MO-MI-NLP specification
14
NSGA-II AMPL Controllable Options

• Population size
• Number of generations
• Real variable crossover and mutation rates
• Real variable distribution indices for crossover
  and mutation
• Integer variable crossover and mutation rates
• Random number seed
• Seed of initial population with start point
• Delta for equality constraints

15
Wrapping Ampl Inside Java Class
16
Using Wrapper Inside Web Service
17
Benchmark Results of Parallel Speedup
18
Engineering Survey
19
Engineering Survey
20
Benchmark Results with Sandren
21
Project Status

• Framework proof of concept completed for inside
  firewall.
• Key Items to industrialize
• End User Administration/Security
• Asynchronous callbacks.
• Uploading design codes.
• Globus grid engine

22
Support

• Work funded by North Carolina Office of the
  President for a Consortium to Promote
  Scientific Computation and High Performance
  Computing

23
Questions?

• dpowell2_at_elon.edu
• jhollingsworth_at_elon.edu

24
Survey of GA Usage in Japan, Korea and Singapore

• Primary Application Automotive.
• Popularity due to parallel processing and global
  search.
• Clusters are usually Linux/Unix with LSF. Size
  varies from 4 to 128 CPUs
• Design point evaluations range from a few seconds
  on a single CPU to 12 hours on a 8 CPU machine.
• GA run varies from 1 hour to 10 days.
• 30 population size for 30 generations
• 60 to 100 population size for 100 generations