Monday, June 8, 2009

VatLife update

The latest on VatLife: I've done very little coding on it, mainly owing to a lack of time. I began an attempt at a tic-tac-toe playing environment, I've thought about adapting both Tom's noble ape brain as a controller, and Scott Davis' Bubble Pond as an environment, and I rambled about it on "Visions of the Evogrid".

Which on the one hand, seems entirely inappropriate, given that the Evogrid (opposed to Biota Eve) is now about artificial chemistry, and VatLife abstracts all the details of self-replication and has the genome as a primitive. And this isn't something I apologize for. I don't think one is necessarily better than the other, they are simply different approaches.

Yeah, the idea of self-replication occuring within an artificial chemistry is more pure and I can see why it would be more appealing to biologists in particular. However, the abstracted approach has produced tangible products, such as circuits and satellite antennae. This is a tiring debate...I wish the two approaches could simply co-exist in peace without the need to rank them or dismiss one or the other.

Anyway, after the recording with Tom I wrote down what I wished I had said, and here it is:

What I attempted to achieve with VatLife was to create a framework that met two main goals: make it a little easier to create a genome-based artificial life simulation by providing a simple set of generic tools, and to put an abstraction layer between the phenotypes of a-lifes and how they act within their environment. My thought there was while A-lifes within a 2D or 3D simulation might evolve to look radically different, they might share the same kind of evolving logic. For example, both a 2D and a 3D environment could contain a-lifes that used something like the Noble Ape brain to interact with their host environments, while these same environments could simulataneously host a-lifes that used different kinds of evolvable controllers.

An a-life in VatLife therefore has two genomes, much as a cell has both nuclear DNA and mitochondrial DNA. The first genome is used by its host environment to produce the a-life's initial phenotype, and the second which specifies its initial control logic. It also contains an ID to the controller that interprets the second genome to produce an initial brain, and interpret it against its host environment. Just as our brains are a product of both biological evolution and development after birth in response to our environment, the idea is that even something like a neural net would benefit from an initial seeding from a genome that evolved through natural selection.

A nice benefit of this structure is that you can author environments and controllers separately, which is particularly useful when the behaviors are not entirely determined from the phenotype. So you can author a simulation and immediately have a-lifes with evolving brains attempting to exploit the constraints of your simulation.

It would also be possible for a VatLife a-life to cross over from one environment to the next, although since each environment would intepret the phenotype-controlling genome differently, it would probably not be likely to succeed.

This model allows for a very different kind of interactions between simulations, one that's less focused on creating a common vocabulary so simulations can interact on a high level, and more focused on leveraging existing code for faster development of evolutionary simulations. At the same time, there's no reason why a VatLife simulation couldn't communicate with other Biota Eve style simulations.

A demonstration of VatLife is hosted at www.ripplingloop.com, which currently has a simple environment hosting a-lifes that can use one of two controllers to interact with their environment. Links to source code and to my blog are also on that page.

No comments: