Algorithm evolves more efficient engine

By Kimberly Patch, Technology Research News

University of Wisconsin-Madison researchers are using small-population genetic algorithms to better the emissions and fuel efficiency of diesel engines.

Genetic algorithms use the principle of evolution to find solutions to problems that have many variables. Like biological evolution, genetic algorithms start with a certain number of individuals that have varied sets of traits, mix the traits to produce another generation, select the survivors to parent the next generation, add some random changes as mutations, and repeat the process for many generations.

With computer-based genetic algorithms, the survivors are those who come closest to a desired result -- in this case engine efficiency.

Research Associate Peter Senecal used small populations of five "parents" each to run through many combinations of six engine variables. (see chart 1) Each variable has trade-offs, said Senecal. For example, increasing the amount of exhaust gas recycled through the engine reduces nitric oxide emissions but increases soot emissions.

Fifty generations and 2 1/2 weeks of computer time later the models converged on a common combination of the variables that used 15 percent less fuel and emitted one-third the nitric oxide and half the soot of today's standard engine designs.

Senecal and University of Wisconsin graduate student Matthew Thiel reproduced the results in steel and oil by tuning a real diesel engine according to the simulation. "The agreement was very good between what was measured and what was predicted in terms of combustion and also emissions, said Senecal. (see chart 2)

Engine design has traditionally been done by either trial and error or by holding many parameters constant and then changing one at a time and seeing what happens, said Senecal. But changing one parameter at a time is imprecise because "all these parameters are interdependent, so to achieve a true optimum you can't just change one factor at a time, you have to change them all together," said Senecal.

Trouble is, the possibilities increase exponentially when you're examining multiple parameters that affect each other. "If you look at all possible combinations of the six parameters that you could simulate or run in an experiment, it's over one billion," Senecal said. This makes even a computer simulation time-consuming and difficult. Genetic algorithms, however, use the much quicker process of natural selection to find the optimal solution rather than testing each possibility. Also, the small population micro-genetic algorithms Senecal used are faster than more traditional large population genetic algorithms.

Senecal's algorithms started with five sets of five parents; each set produced 50 generations for a total of 250 function evaluations -- or possible individuals. Extrapolating from simpler tests he ran comparing the small population method with the more traditional 50-to-100-parent models, he estimated that a 50-parent engine model would have taken about 6 1/2 months to run.

Traditional genetic algorithm models use many parents in order to cover a lot of ground in problems that have a lot of solutions -- or mathematical peaks in a mountain range of possibilities. Smaller populations may not roam over the whole area and consequently converge on a solution that is a tall peak, but not the tallest -- or best solution -- in the range.

The somewhat unusual micro-genetic method Senecal applied avoids this problem by introducing a lot of random mutations into the mix, he said. Senecal ran his set of five designs for as many generations as it took for them to reach a common solution, and then began again with the best design plus four random designs. The many random changes allowed the algorithms to find the best solution to a problem that had many possible solutions, said Senecal.

Although researchers have been hesitant to use small population genetic algorithms on complicated problems due to the potential of missing the best solution, Senecal has proved it practical for this type of problem, according to David E. Goldberg, Professor of Genetic Engineering at the University of Illinois and Chairman of the International Society for Genetic and Evolutionary Computation.

"The question is whether the diversity that's there is sufficient for the problem at hand -- here the results are so good that apparently it was," said Goldberg, who added that this is the first time he has seen genetic algorithms used in practical motor applications.

Senecal is currently readying a set of micro-genetic algorithms designed to find more efficient engine shapes. "We're using this tool [by] breaking up the geometry of the combustion chamber into a number of geometric parameters that we're optimizing over," Senecal said. Traditionally, engine shapes haven't changed much, largely because machining experimental parts is time-consuming and expensive.

The design efficiencies Senecal's work has uncovered "may influence the thinking of designers right away," said Rolf Reitz, a mechanical engineering professor at the University of Wisconsin who oversaw Senecal's work. Eventually, his methods many be used in commercial engine design, a process that takes three to four years from concept to production, Reitz said.

Senecal's exhaust parameter research is slated for publication in an upcoming issue of the International Journal of Engine Research. His research is funded by The Department of Energy, The Army Research Office, Tank-automotive and Armaments Command and Caterpillar Inc.

Timeline:   <1 year; >2 years
Funding:   Government; Corporate
TRN Categories:  Applied Computing; Artificial Life and Evolutionary Computing
Story Type:  News
Related Elements:  Chart 1; Chart 2; Paper, International Journal of Engine Research


June 28/July 5, 2000

Page One

Cortex chip goes both ways

Sampling ability broadens quantum computing

Nano-scale plotter goes parallel

NASA grasps intricacies of human hand

Algorithm evolves more efficient engine


Research News Roundup
Research Watch blog

View from the High Ground Q&A
How It Works

RSS Feeds:
News  | Blog  | Books 

Ad links:
Buy an ad link


Ad links: Clear History

Buy an ad link

Home     Archive     Resources    Feeds     Offline Publications     Glossary
TRN Finder     Research Dir.    Events Dir.      Researchers     Bookshelf
   Contribute      Under Development     T-shirts etc.     Classifieds
Forum    Comments    Feedback     About TRN

© Copyright Technology Research News, LLC 2000-2006. All rights reserved.