Genetic Algorithm Helps Identify Criminals 84
Ponca City, We love you writes to tell us that a new software approach to police sketch artists is finding surprising success in a trial run of 15 police departments in the UK and a few other sites. The software borrows principles from evolution with an interactive genetic algorithm that progressively changes as witnesses try to remember specific details. Current field trials are reporting an increase in successful identification by as much as double conventional methods. A short video with a few working shots of the new "EFIT-V" system is also available on YouTube. "[Researcher Christopher Solomon]'s software generates its own faces that progressively evolve to match the witness' memories. The witness starts with a general description such as 'I remember a young white male with dark hair.' Nine different computer-generated faces that roughly fit the description are generated, and the witness identifies the best and worst matches. The software uses the best fit as a template to automatically generate nine new faces with slightly tweaked features, based on what it learned from the rejected faces. 'Over a number of generations, the computer can learn what face you're looking for,' says Solomon. The mathematics underlying the software is borrowed from Solomon's experience using optics to image turbulence in the atmosphere in the 1990s."
How valid does it turn out to be? (Score:3, Insightful)
How do you measure success? (Score:3, Insightful)
How do they know if this thing actually works? If they're using the computer generated sketch to finger a suspect, and then presenting that sketch as evidence to a jury who convicts, and then using that conviction as evidence of the algorithms accuracy that's just circular reasoning.
The memory is not an immutable thing. It's quite possible that in the process of generating the sketch you are leading the witness on, even implanting memories. So what happens if you generate a sketch that doesn't look like the actual criminal, and present that to a jury and get a conviction. Is that going to be counted as a success?
This is actually very cool... (Score:3, Insightful)
This technology, at its core, is a little bit like PicBreeder [slashdot.org]. It doesn't include the complexification, but the principle is the same.
There is an argument about 'leading the witness' being bandied about as if that makes this thing useless. If you read the articles, they talk about that, and they show that it is no worse than any existing techniques, gets good results, and works for people that can't work with sketch artists.
The reality is, this technology has applications beyond what it is being used for.
Personally, I would *LOVE* to be able to tinker with technology like this.
You got that right. (Score:4, Insightful)
The test case:
Get a group of test subjects - college students are always great for this. Have your "assailant" run up to the subject and Yell, "Hi!" and then hand the "victim" a flower and then run off. Right then and there, the "victim" goes a "files a police report" with the researchers following typical police procedure.
After about a thousand tests on different subjects with statistically significant positive results, then and only then, will I start to believe this "technology" and maybe with more tests will I think it should be allowed as evidence in a court of law.
Other than that it just a gimmick - we're talking about taking people's freedom here or sentencing them to death.
Re:How do you measure success? (Score:3, Insightful)
They won't present the sketch as evidence to the jury. They will call the witness and ask him to identify the suspect. They will be able to do other things like take fingerprints and DNA samples from the scene and match them to the suspect.
Re:GA vs. Hillclimbing (Score:3, Insightful)
Instead of a sketch artist listening to a description and modifying based on feedback, the system will be "prompting" the witness.
Prompting has been shown to cause false memories of details, so I imagine it will be even worse when you consider the "the computer generated this, it must be right" phenomenon.
Re:It's about time for GP (Score:2, Insightful)
I think GAs have definitely had a time when they were popular at least as an idea, mostly sometime in the early 90's or so, and there was quite a bit of research into applying them to various problems. They haven't always turned out to perform very well, though. Quite a few attempts have been made towards using GAs as a heuristic to traditional NP-hard combinatorial problems, for example, and while there has been some success, quite often other heuristics have beaten GAs.
My impression of the beauty of GAs in general isn't quite as positive as yours. The idea certainly is aesthetically pleasing, and you can, in theory, try to apply a GA to pretty much any optimization problem, but how well GAs work really depends a lot on the problem: the very nature of the problem (does it fulfill the building block hypothesis, or whatever magic is that makes GAs work for some problems?), what kind of a landscape the search space provides, what kinds of cases of the problem are more likely in your application, etc. That's not including all the nontrivial problem-specific tweaking that will be needed in a practical application of a GA, such as how to encode or represent the solutions (has a big effect on how much good genetic crossover does).
I'd rather say that GAs have worked well for some specific problems, and some new specific applications will probably still emerge, but I'm not sure they will ever become very generally applicable. They had a chance, but it turned out that they mostly work just for some particular problems, not others, and nobody seems to really know very well why.