Sunday, August 08, 2010

deception in science

Instapundit links to this article: "Artificial life forms evolve basic intelligence". Practically everything that the article says about the experiment is literally false. The statements may be figuratively true, but the figurative language is misleading because it is intended to be taken as factual or at least reflective of facts and it is not at all reflective of facts. I'm not going to call this a lie because I think the researchers and the author of the article are themselves confused; they are not deliberately deceiving anyone, but it is false and misleading.

Let's start with the title : "Artificial life forms evolve basic intelligence". First, the things that they are talking about are not "artificial life forms". They are not alive in any sense. They do not have any organic structure or any physical structure at all. They are nothing but tokens in a computer program, marks in digital memory much like the marks on the screen you are reading represent words.

Second, these marks did not "evolve". They did not reproduce at all. What happened is that a computer program created a set of marks and then created another set of marks based on a set of rules. Those rules were set up to reflect what happens in reproduction and evolution, but the process was not reproduction and evolution. Just as the marks merely represented life forms, the operations merely represent reproduction and evolution. The author is making the very same mistake that ancient magicians made --confusing the symbol for the thing symbolized.

Third, the end result of all of this did not display intelligence in any literal sense. If the marks had been real life forms rather than just abstract representations of life forms, and if the events had been real reproduction rather than just abstract representations of reproduction, then the events would have suggested what biologists call irritable behavior --behavior that is influenced by outside sources. In other words, the events represented irritable behavior but were not irritable behavior. The behavior symbolized is a long way from anything that would be called "intelligence" in any case.

So what did they actually do, in real, rather than symbolic language? These sorts of experiments work something like this: the marks are random 5-letter words like "kwdez" and "qsbjl" (don't bother trying to pronounce them...). Each of these marks represents one organism. You start with a collection of these marks and the computer program goes through the collection periodically and adds new marks. Each set of new marks is a generation.

The way that they make the generations represent reproduction is by having each mark in the new generation based on a mark in the previous generation. You use a rule such as "take the previous mark and get a new mark by randomly changing one letter". The new mark would be called a descendant of the previous mark. For example, descendants of "kwdez" might include "kadez" and "kwdqz".

Notice that this reproduction is entirely figurative. The marks are not doing anything. It is the computer program that is creating new marks based on the old marks. There is nothing wrong with symbolic language but in the field of artificial intelligence they have a long history of confusing representations with reality.

Here is an example:
In early memory experiments, Laura Grabowski, now at the University of Texas-Pan American, Edinburg, set up a food gradient in a computer environment made of a grid of cells. First-generation Avidians were placed at the low end of the gradient, in a cell that had minimal food. Straight ahead of them, however, lay a cell that had more.
Note the continuing confusion between symbol and thing symbolized. There was no food. The "food gradient" would have been something like this. When the program is going through the list of current marks to create a new generation, each mark is reproduced 0 or more times based on some rule. For example, you might make a rule like this: to decide how many descendants to produce for a word, you flip a coin 5 times plus 1 more for each time that the letter "q" appears in the word. In other words a word with no "q"s gets 5 chances to reproduce and a word with 3 "q"s gets 8 chances to reproduce. Since the number of "q"s effects how many descendants a mark gets, this symbolizes "how close they are to food".

Notice how arbitrary the symbolism is. Instead of the number of "q"s representing "how close they are to food", it could just as well represent anything else that leads to more reproduction. The marks could symbolize bull elephant seals and the number of "q"s could represent the number of cow seals they inseminate. Just as easily, the marks could represent companies, the descendants could represent copy-cat companies and the number of "q"s could represent the profits.

This particular experiment seems to have had a more complex generation rule with some features representing genetic abilities and other features representing physical movement, but that doesn't matter to the primary point. The point is that the symbolism is completely arbitrary. Any process that can be mapped into the same abstract structure could just as easily be the thing symbolized. Instead of writing an article about the evolution of intelligence, they could have taken the exact same experiment and written an article about economics.


Foxfier, formerly Sailorette said...

But...but... the title "computer program model of reproduction successfully models response to programed stimulation" is so lame! And long!

First thing that came to MY mind was how plants respond to light-- a lack of light makes a plant grow. (because if there's light to your left, and shade on the right, growing on the right will make you tilt into the light, and on the left into the dark)

Marcel said...

Hmm, so I guess "First robot able to develop and show emotions is unveiled" is similarly misleading.