Tuesday, May 08, 2007

Evolution vs. ID.

I ran a simple test case due to my annoyance with the prophets of evolution proclaiming that evolution was a superior paradigm to ID. The basic problem is of the form: a1^m+a2^m+...an^m -> minimum subject to the constraint a1+a2+...+an=1. This is a trivial optimization problem with the solution: a1=a2=...=an=1/n. We can start it off with a random number generator to make initial values for the a's. Using values of m ranging from 0.5 to 3.5, and values of n=30, a sequential quadratic programming (SQP) method nails the solution to 6 significant digits in about 10 to 20 iterations. The only parameter that I tweaked was the convergence tolerance.

Attempting the same solution with a genetic algorithm requires about 4 million function evaluation and gives me something barely accurate to 2 significant digits. In this case, I needed to tweak the number of parents, the number of children, the mutation rate, the mutation size, the cross-over rates, ... In the end, more ID was required for the genetic algorithm than the ID algorithm! A book I picked up recently, "Evolutionary Computation for Modeling and Optimization" has a chapter on optimization of real valued systems. What is sad is that they do not inform the student that GA methods are hopelessly pathetic when compared to state-of-the-art optimization algorithms. The book does allude to the inefficient of GA, but instead encourages the reader to try IDing the GA methods to emulate aspects of ID methods. Thus, we have a GA implementation of a "line search" algorithm which is a key ID method. In the end, this merely diverts even more ID resources into the effort to prove that ID isn't necessary! Anything to appease the ghost of Darwin.

One final thing to note on this problem is that 99% of the ID goes into the problem definition. Simply by writing down the mathematical description clearly, I have invoked millenia of ID concepts in real numbers, integers, algebra, geometry and calculus. By the time I am ready to solve the problem, countless optimization techniques would work. The fact that a genetic algorithm also works is a no-brainer. This isn't due to the power of GA at all, but rather due to my ID prowess at defining a tractable problem in the first place. The evolution paradigm is worthless in its entirety.

2 comments:

LoneRubberDragon said...

If the task were setup to solve 100 types of equations of different types, using the atoms of mathematics, in genetic combinations, and a bank of memory codes for fruitful equations to simulate species of fruitful analysis, with utility of application functions, one million applications of the genetic algorithm would have solved the 100 types of equations; finding taylor series, quadratic convergence algorithms, differential solutions, and a heirarchy of equational genetic codes, that would be the ID product of one million GA iterations. Using GA to solve one optimization problem, is a misapplication of GA in its most efficient incarnation. SQP only solves SQP problems, the GA could have solved them all.

I could say the ID paradigm is worthless in its entirety, but I would be utterly dishonest in my formulation, against intelligent design, the product of evolution and intelligence in feedback and recording. GA are best when they use the atoms of the right idiomatic logos type, or else it takes billions of years to reach human intelligence. Using GA for number variation is a poor test comparison.

The image of data, that evolution can even accomplish the development of primates in 700 million years from single cells, is amazing, and yet it is painful and slow. But painful is knowing that 60,000,000,000 humans have already passed away, and God wants more death until the age ends. In 1000 more years at 7,000,000,000 humans average at 50 years average, will denote another 140,000,000,000 human deaths. The image of the earth's truth shows 1 trillion higher life forms with average 10 year lifespan, with 60,000,000,000,000,000,000 lifeform deaths over 600 million deaths. It isn't an easy picture God has painted for an old earth Creationist, assuming things are as they have always been of old times. Nothing new under the sun. But that is the force that has made evolution, in some human opinions, when it is done blindly, and noone is there to guide it intelligently every day over the planet, to shorten the entire age of the universe abut earth. So the logos of molecular bonding, isn't an easy path to human intelligence, but using GA inappropriately, and calling it worthless, misses so much. Call a baby worthless because it will never grow, and it will never learn, and it will never amount to anything, because the baby only reaches 2 digits of precision.

It is to a small extent like handling machine bits and machine bitfield unions in BASIC, or trying to write specific math code with iterations in LISP, or handling long ints in PASCAL, or represting human thought in HEXADECIMAL only with no pictures, sounds, feelings, emotions. Language restricts the potential of any problem when misapplied, like trying to describe quantum physics math with a kindergartener's concordance, or defining chess to a dog. The Logos defines the realms you can reach, but to close one's eyes to one leg in preference to the other means you cannot walk. It took 4 billion years for parallel biochemistry and lifeforms to reach human sentience, and now 1 million improper applications of GA is too long, and makes all of that potential evolution worthless in its entirety? If God told you He uses evolution in any way, that is an affront to the mathematics of God, and an abomination, as a lying tongue:

Proverbs "22 Lying lips are abomination to the LORD: but they that deal truly are his delight. 23 A prudent man concealeth knowledge: but the heart of fools proclaimeth foolishness. 24 The hand of the diligent shall bear rule: but the slothful shall be under tribute. 25 Heaviness in the heart of man maketh it stoop: but a good word maketh it glad."

Looney said...

LRD, just a few remarks:

First, there is no single genetic algorithm, but a vast array of them. In the book, "Evolutionary Computation for Modeling and Optimization", by Ashlock, he actually advocates coding different GA for every problem, but exploiting the techniques of optimization theory to help them converge.

Second, 60 million iterations really isn't a big number in this realm.