[ Jocelyn Ireson-Paine's Home Page | Publications | Dobbs Code Talk Index | Dobbs Blog Version ]

In Silicon No One Can Hear You Scream

In his book of TV reviews The Crystal Bucket, Clive James writes of David Attenborough:

It is a lucky break that the presenter looks normal, because some of the life-forms he is presenting look as abnormal as the mind can stand. To Attenborough all that lives is beautiful: he possesses, to a high degree, the quality that Einstein called Einfühlung — the intellectual love for the objects of experience. Few who saw it will forget Attenborough's smile of ecstasy as he stood, some years ago, knee-deep in a conical mound of Borneo bat-poo. Miles underground, with cockroaches swarming all over him and millions of squeaking bats crapping on his head, he was as radiant as Her Majesty at the races.

How long will it be before such emotions can be inspired by the evolved objects of digital experience? I've been looking at Thomas Miconi's page Evolving Virtual creatures. Miconi is a postdoctoral researcher working on neural models of visual scene perception, a topic illustrated on his home page by a low-resolution image of Charles Darwin. Continuing the theme of evolution, the home page also links to Miconi's thesis, The Road to Everywhere: Evolution, Complexity and Progress in Natural and Artificial Systems, subtitled Everything you vaguely thought you'd like to know about evolution, but couldn't really be bothered to ask. But one thing I already know about evolution is that it works. And in case there are any creationists out there who still doubt this, look at the virtual-worlds videos linked from Evolving Virtual creatures.

I took this posting's title, "In Silicon No One Can Hear You Scream", from one of Miconi's papers. It's appropriate, because most of his videos depict creatures evolved to fight. For example, Evolving fighting creatures 7 shows one small creature which constantly tries to aim its swinging tail at its larger opponent, which hits back using what Miconi calls a "flail" appendage. And in Evolving box-grabbing creatures 1, a V-shaped creature resembling a caterpillar sliced down the middle and rejoined at the tail fights for possession of a box with a smaller and less nimble animal that moves by rolling. Miconi believes his evolution of fighting is a first: there had been previous attempts, but they were simplistic.

Miconi's work reimplements Karl Sims's Evolved Virtual Creatures project, which I summarised here, for the old Dobbs AI Newsletters. Sims's animations showed a simulated beachside world — complete with a physics engine which implemented both gravity and friction — in which he had placed an initial population of several hundred simple blocklike creatures with randomly-generated bodies and nervous systems. Each creature was tested for its ability to perform a pre-specified task. Those that were most successful survived, and their genes were copied, combined, and mutated to make offspring for a new population. The new creatures were once again tested, and so evolution continued. One of the tasks set for the creatures was to move as far as possible in a certain direction. Members of the first generation might only twitch or stand like sticks, but as simulated generations were born and died, impressive results emerged, such as a helical snake which had evolved to gain purchase on the ground by friction, moving forward by spiralling along its axis.

Miconi says that his work is the first successful reimplementation of Karl Sims' system. He has improved the genetic encoding and the way creatures develop after "birth". Moreover, whereas Sims's creatures had complicated neurons, into which he had preprogrammed a lot of behaviour, Miconi's have much simpler neurons that need considerably more help from evolution. There's more info, including a download, at Evolving Virtual creatures. With faster evolution and a more advanced physics engine, perhaps someone will be able to do fluid dynamics, wings, and flight. At least the bat-poo will be virtual.