ALife Project
03/22/2005 - The limitations of the current analog computers definately put a damper on my project plans; however, I now think there may be an opportunity to explore learning as a tool to guide genetic algorithms. To do so would within the context of analog as it currently is at IU [TODO: Write up a brief summary of this.] would require, I think, a fairly well-defined problem and its solution. I can only explore a small subset of the possible sheet configurations; I have to choose wisely.
Imported Notes
TODO: Update this. Also, way to timestamp this, dumbass!
There seems to be a lot of success getting tailored systems to exhibit emergent behavior [sidebar: what is emergence?] -- Tierra, Polyword, Blocky Creatures, etc. What I haven't noticed, and Dr. Yaeger confirms this, is that there are not a lot of open-ended systems in the works.
I would like to design a pseudo-value neutral world, where seeded components are free to interact as they may. The hope would be to see interesting systems emerge; or, at least, an increase in complexity [sidebar: what is complexity?]. The obvious way to do this would be to work with low-level physics and chemistry packages, to try to model the "primordial soup" that gave rise to the first nuecliec acid reactions [sidebard: yes, I am aware I do not know what I am talking about].
Why reinvent the wheel? Nature already provides the physics and chemistry engine. This suggest two things: one that I am really researching the "flash point," where randomness gives rise to self-catalytic complexity, and two that I am proposing a physical experiment.
I can accept one. I implicitly accept that evolution [sidebar: what is], whatever it may be [sidebar: oh, okay], will drive the system toward more "lifelike" states over time. As a sort of corollary to this point, I am not concerned with intelligence, or the appearance of intelligence. Only the appearance of life.
I do not accept two. I am not (that) interested in recreating the flash point. I'm more interested in exploring other possible flash points [sidebar: flash points, is that a good label?]. To that end, I would like to discard physics and biology and come up with an entirely new simulation that attempts to demonstrate a flash point of some sort.
Artificial Flash
I need to read the Tieran paper, Ray discusses "open-ended" systems. Ditto Sims, and all the other reading.
At initial glance, Karl Sims seems to have the right idea. Something about the use of directed graphs exposes the glimmer of a solution. The "traditional" ALife model of genotypes, agents, and genetic algorithms is too close to the biological model. The apple falls close to the tree, so to speak.
Put another way, traditional ALife simulations use a bit string to encode genes. The genes are mapped to features of the phenotype [sidebar: right?], either directly or indirectly. They directly influence limb structure, ability, and any number of other variables. This yields fascinating results, but it preselects a point in the evolutionary space far from the flash point. It is hardly a value-neutral system. The very nature of the simulation is biased to encourage a specific result. That the experiments often succeed is a testament to the robustness of the processes at work, but I don't think it gets at the deeper question of such systems come to be [sidebar: another way of looking at my "flash point" research interest? Or another topic altogether].
Sims' digraphs may be useful as an illustrative tool to define a new biology. Rather than borrow terms and concepts from natural biology, digraphs allow a whole new structure. I imagine a transition table of sorts. Defining elementary structures [sidebar: even this supposes values, I don't know if it is possible to simulate a "big bang," thought it would be neat to try] and their interactions in this table allows us to perfectly create new laws of physics. It doesn't matter how the architecture implements the behaviors, it simply does. The transition table enumerates the "laws of nature," the processor makes them so, and that's all the more we need to know [sidebar: this lends itself to interesting philosophical/computational questions, but not now!].
The difficult problem would be disambiguate [?] the "agents" from the environment. Presumably, the building blocks are all the same. It's just certain combinations yield rocks, certain combinations yield "nucleic acids." This is just speculation: but perhaps this is a world where there is no fitness function. The "failed" organisms [sidebar: watch those biology terms, bud] are just as much a part of the environment as the "fit" members.
Chaos
At this point, my train of thought is derailed. I'm speculating wildly. Maybe a sufficiently well designed transition table, plus some seed, yields the big bang, which differentiates into "atoms," "matter," and eventually ALife. That seems to be increasing complexity. Ahh well, I have to fill out this exam so I can go to Indy tomorrow night.
Emergent Properties (Tacitity)
I wonder. Assuming a deterministic system... Okay, maybe it doesn't need to be strictly deterministic. Better assumption: assuming quantum/random effect are not a necessary component of the system; that is, that any random effects are sufficiently localized as not to manifest themselves on a grand scale. [wow, what a grammatical mess you're making]
So assuming that, perhaps the definition of the explicit ruleset defines rules, implicitly, at the higher level. Now I've been trying to digest this notion of tacit information since Hakken's ethnography course. On one hand, I think I'm barking up the wrong tree; but on the other, I see it all over. Bardzell's composite picture from Marty's class, Josh's presentation on external cognition, [http://design.informatics.indiana.edu/wiki DesignWiki]. All of these point to the existence of higher-order, well, order. There seems to be an opening for a "macro calculus" of knolwedge at high levels.
Shannon's model, and in a broader sense, physics, try to define the information of a system [I'm mincing theories right now] at its most base level. However, there are clearly external dimensions to the system. Shannon's model might tell us something about a stream of 1s and 0s, but it doesn't really tell you what it means [even if you're encoding first-order logic, ain't that weird?]. I'm sure I'm disambiguating things a bit too broadly - Alife has convinced me that organism and environment are intrinsically linked - but it is hard to ignore.
Take Conway's game of life. Simple ruleset, complex behavior. However, once we identify a glider, we know what it is going to do. Slightly more complex ruleset (2nd-dimensional?), slightly more complex behavior [here's a trick: define complexity].
Another example: chess. Relatively simple ruleset, mind-numblingly complex behavior. At any one point, there are dozens of possible moves. Very complex behavior emerges, through the course of a game. Big Blue smashes through this, but Kasparof [sp?] doesn't get his ass kicked. Why? I think he knows the nth-dimensional ruleset. Big Blue is, despite clever programming, a brute force attack. Kasparof is a grand master. He "knows", if implicitly, the higher order rules of chess. He can see the gliders, so to speak. Further, he can operate on those gliders. Forsythe talks about "tacit information," Bardzell shows us a 2D array of images that forms a macro image. The rules exist. We just don't know how to manipulate them algebraically.
There has to be a calculus or an algebra [differentiate these terms] that operates on higher orders than what we hvae now. Maybe?