Multi strategy intelligent optimization algorithm for computationally expensive cae

NAFEMS World Congress 2015 | 21-24 June | San Diego | California | USA

Multi-Strategy Intelligent Optimization

Algorithm For Computationally

Expensive CAE Simulation

S. Costanzo, Z. Xue, M. Engel,

S. Parashar, C. Chuang


Goal

• Reduce number of evaluations

• Solve a complex constrained MDO problem

• Handle computationally expensive CAE

simulations

• Case study:

–MDO of Ford Taurus 2001


MDO Ford Taurus 2001

Our target was to improve the baseline design of

the 2001 Ford Taurus model based on the

National Crash Analysis Center (NCAC) criteria.

Disciplines considered:

• safety (subdivided into Full Frontal and 40%

offset impact)

• NVH (noise, vibration & harshness)


MDO Problem Description


For a vehicle model with over a million elements a

single design evaluation takes about 5 hours on

32-CPUs HPC clusters.

The challenge


Software Platform

is an integration platform for multi-objective and multi-

disciplinary optimization. It provides a seamless coupling with

third party engineering tools, enables the automation of the design

simulation process, and facilitates analytic decision making.


Optimization Workflow

Problem: identify most appropriate algorithm


Algorithm suite

Taking into account a subset of available

optimization algorithm categories:

• Gradient-Based

• Heuristic

• Multi-Strategy


Algorithm suite



• Gradient-Based

• Heuristic

• Multi-Strategy

Main focus: few number of evaluations.


Algorithm suite



• Gradient-Based

• Heuristic

• Multi-Strategy

Main focus: few number of evaluation.


Heuristics: Genetic Approach

• Well known algorithms, recognized in

literature

• Allow for parallel computing

• High robustness and design space

exploration capabilities

• Elitism allows the GA to focus on the best

solutions and explore the most interesting

regions


Heuristic: Genetic Approach

MOGA-II

Multi-objective Genetic Algorithm II is an

improved version of MOGA developed by C.

Poloni, that uses a smart multi-search

elitism for robustness and a directional

crossover for fast convergence.


Heuristic: Genetic Approach

NSGA-II

Non-dominated Sorting Genetic Algorithm is

a well-known multi-objective optimization

algorithm developed by K. Deb,

implementing a fast and clever elitism and

non-dominated sorting.


Multi-Strategy Approach

• Combine heuristics with other techniques:

– Local search for refinement

– response surfaces to speed up convergence

• Suitable for MDO problems where

correlations between disciplines may

require different optimization techniques

to achieve the best results



FAST

Automatic iterative algorithm focused on the

exploration, exploitation and validation cycle.

Fast optimizer uses different internal

adaptive Response Surface Metamodels to

speed up the optimization process.



HYBRID

Combines a genetic algorithm and a

gradient-based SQP local search algorithm

within a steady-state evolution scheme,

which can keep the computational resources

saturated with concurrent design

evaluations.



pilOPT

Exploits the advantages of local and global

search algorithms while automatically

adjusting the ratio between different

optimization strategies based on their

performance. It also uses Response

Surfaces to speed up the optimization.


Preliminary Benchmark

• Comparison based on a limited number of

evaluations

• Mathematical test functions from literature

– Michalewicz test library

– Zitzler benchmark library

• Target: Find the most appropriate strategy for the Ford

Taurus model optimization


Preliminary Benchmark: T01

Objective Function:

Constraints: Bounds:


Preliminary Benchmark: T01


Preliminary Benchmark: ZDT2

Objective Functions:

Bounds:




Inverted Generational Distance

•




Preliminary Benchmark

• Multi-strategy algorithms outperform GA

on short runs

• All candidate algorithms have shown

remarkable results on a long run

• Accordingly, we decided that we could

afford three short optimizations


Application Test

• Also in this case only default algorithm

parameter settings were used

• Maximum number of evaluations for each

algorithm was set to 400

• In spite of the use of significant parallel

computing resources, one whole run took more

than three days


HYBRID


HYBRID


FAST


pilOPT


Conclusions

• Good validation for multi-strategy algorithms on

a complex MDO problems

• The baseline design weight has been

successfully reduced with all algorithms

• The best result has been obtained with pilOPT,

with 13.72% weight reduction

• The remarkable performance indicates great

potential for intelligent multi-strategy algorithms


Future Work

• Single-parameter multi-strategy algorithm

• Improve automatic controls in pilOPT

• Increase number of available internal algorithms

• Find other complex MDO cases where intelligent

algorithms could be effectively applied


Thank you for your attention

Multi strategy intelligent optimization algorithm for computationally expensive cae

Engineering