SoylentNews is people
SoylentNews
AI researchers from 34 institutions have compiled 27 examples of ways in which digital evolution produces surprising and creative solutions. They also make the case that such surprise is the rule rather than the exception. These 27 are just a small subset of what are for now amusing anecdotes and for every story they received or heard, there are likely to be many others that have been already forgotten as researchers retire.
The process of evolution is an algorithmic process that transcends the substrate in which it occurs. Indeed, many researchers in the field of digital evolution can provide examples of how their evolving algorithms and organisms have creatively subverted their expectations or intentions, exposed unrecognized bugs in their code, produced unexpectedly adaptations, or engaged in behaviors and outcomes uncannily convergent with ones found in nature. Such stories routinely reveal surprise and creativity by evolution in these digital worlds, but they rarely fit into the standard scientific narrative.
[...] One obstacle to their dissemination is that such unexpected results often result from evolution thwarting a researcher's intentions: by exploiting a bug in the code, by optimizing an uninteresting feature, or by failing to answer the intended research question. That is, such behavior is often viewed as a frustrating distraction, rather than a phenomenon of scientific interest. Additionally, surprise is subjective and thus fits poorly with the objective language and narrative expected in scientific publications. As a result, most anecdotes have been spread only through word of mouth, providing laughs and discussion in research groups, at conferences, and as comic relief during talks. But such communications fail to inform the field as a whole in a lasting and stable way.
Limited to the lab, the examples given are currently either humorous or intriguing or both. Those outside of AI work and, maybe some inside it, forget the completely alien nature of the algorithms and their ability to deliver exactly what was asked of them. These examples help illustrate that nature.
From Arxiv.org : The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities
(Score: 2) by fyngyrz on Saturday March 17 2018, @10:35AM (2 children)
At one point, we (my company) were selling a product that used evolutionary mechanisms to alter digital critters, or "crits." There was an environment that contained food and hazards; the critters could mate, and you could tune the mating habits to the fitness of the crits, basically causing them to prefer to mate with higher performance individuals.
The crits "DNA" was a series of mid-level behaviors such as move forward, eat, mate, turn left if there's a hazard, that sort of thing. Basically a set of serially executed opcodes that defined the behavior of each crit. Crits could starve and die, or eat, mate and thrive using various strategies. Offspring had DNA that was a mix of the parent's DNA.
You could randomly load the crits' DNA with these opcodes at any time (typically when you started), and you'd dependably get a very low-performing crit population. However, let them breed for a few generations against a set of reasonable metrics, and the performance of the population would rise, often very quickly. The population's fitness is available as a graph at all times.
Here's the funny thing: I – and I'm the one that conceived of and wrote this application – tried over and over again to manually contrive a set of ops using my knowledge of the system that would start out with a high-performance crit population. I never succeeded. I could let the crits evolve such a set of behaviors and then can those for use, but... I could not come up with one myself, and examining the behaviors that the crits themselves developed did not result in "aha" moments. The ended-up-as-high-performance evolved strategies were... weird.
If you've got, or manage to get, an Amiga emulation such as UAE running you can play with this yourself. I've released the product for free here. [datapipe-blackbeltsystems.com] It has built-in documentation.
You can examine and/or change the crits' DNA at any point, and there's considerable control over the environment and stresses. It's designed so that you can easily wrap your head around everything that's happening in most ways. But like any system with many variables, there are an enormous range of potential behaviors and outcomes, and in this case, that manifests directly as non-obvious (to me, anyway) DNA sequences.
(Score: 0) by Anonymous Coward on Saturday March 17 2018, @12:04PM
Oh hey you were part of Digital Soup. I remember this! Kudos!
(Score: 0) by Anonymous Coward on Saturday March 17 2018, @08:03PM
I never had an amiga, but I remember hearing about a few of these as examples of why they were so much better than PCs (along with Video Toaster and some of the other well known software.) Cool to know we have such interesting characters amongst our membership, even if we don't always get along.