Video results… empty calories, but taste great

Some results simulating agents in a 2D plane, with completely distributed controllers (though there is the assumption of total connectivity.  Circles around the black agents represent force fields, which tend to keep agents at a particular distance from one another.  A new, larger force field is created when an agent thinks it is correctly paired.  In addition, the coordinate system continually zooms outward to enclose all the agents, though this isn’t obvious from the blank background.

Video with force fields drawn

Video without fields, showing raw agent motion

Waiting for update…

As per usual, updating Eclipse proves to be a PITA.  Thought I had everything planned out this time, ready for the drop, but a quick update on top of the update brought the program to its (gigabyte sized) knees.  While I’m re-updating, figure I may as well post my recent exploits… such as they are.

After my meeting with Inman this week, in which I was able to show my toy “unlimited” self-assembly system, we’ve decided the best way to go forward is to get complicated single-scale assembly going with arbitrary parts.  That should, in theory, get me to assembling “complicated” structures of any size.  I’ve been looking for a drop-in system for assembling complicated swarm formations, and may have finally found one from Elkaim and Kelbey, but only for identical agents.  A consequence of multi-scale assembly is that you get multi-scale parts at the end of the day / week / sim, and ideally we want to assemble things out of both little parts and big parts.  I’d call this process “heterogeneous self-assembly,” but google suggests that either nobody knows anything about it or other people found a better term for it.  I’ll keep looking.

I’ve also had some interesting (to me) thoughts about defining structures of multiple agents as informational entities.   There was an interesting paper about using information theory to automatically locate sensors from raw sensor data by Polani’s people?, if you considered the sensor data as large information vectors.  You could do the same for coordinates of agents moving through an environment, and dynamically place the agents in the same “information metric” as the sensors.  Agents in structures would have their information correlated, is my hypothesis, and so would tend to group in this information space.  More interestingly, if this was the case, multi-scale assembly might generate hierarchical information grouping.  If this was a requirement for multi-scale assembly, that would give some really pretty bounds on what kind of systems can exhibit this behavior in information terms.  Of course, I don’t really know anything about information theory, so getting to the testable point of these hypothesis is gonna be a steep slope involving lots, and lots, and lots of statistical practice.  Hmm.

A Manditory “Hello World”

This is my first post to the research blog, and as such I’ve been considering the general format of these entries. I’ll try for once-a-week updates, assuming there is anything to update about, and once I get some other webspace @ Sussex I’ll be posting current research results, links to papers and related work / blogs, as well as my thoughts as to my current research directions.

On that note, coming to Sussex my main stated research focus was in understanding scale transitions in evolutionary terms. “Living” systems, over the past 4 billion years they existed, have slowly developed into a heiarchy of scaled systems. Though Maynard Smith’s “Transitions” probably explains it more comprehensively than any other source, the main idea is that life has slowly organized from “simple” autocatalytic chemical systems to self-maintaining distributed ecosystems of organisms made themselves of distributed cells. How is it possible for evolution to drive toward distributed organization, so that, for instance, the units of selection change from cells to organisms? The answer may be (and probably is) highly complex, having to do with particular organisms and environments throughout history, but there must be some set of minimal requirements for evolution to act on distributed systems instead of single replicators.

In order to address this (highly theoretical) topic, I’ve currently decided to take a rather meandering approach. Instead of tackling evolution of distributed systems head-on, with no preconcieved notions of what these systems might be, I plan to assume that all of the distributed systems evolution sees in the natural world are self-organized. After all, for evolution to work on any system, that system must self-replicate, and the first of these replicating systems must logically have emerged in a self-organized way. Through self-assembly then, and self-replication, we can put borders on the potentially unbounded set of organization types evolution uses when scale transitions occur.

As a bonus to this approach, self-assembling / self-organizing systems (these terms basically view the same process from different perspectives) of multiple scales are rare in engineering, and of huge potential use.  A system which has predictable (and controllable) behavior from the very large to very small would essentially “decouple” the assembly problem from the scale at which the assembly occurs.  I.e., building a robot the size of a man becomes as computationally “easy” as building a robot the size of a city. I assume that there are fundamental limitations as to how far this decoupling could be applied in the real world, and hope to discover what these are.

I now (after debugging feverishly last night) have a simulation of swarm robots exhibiting this kind of scalable building behavior, which assumes only that each robot is able to move at a constant speed, know the distance to every other robot, and each timestep communicate three numbers with one another on a large number of virtual “channels.” Hopefully I’ll be able to post the video online soon. This particular simulation seems a good testbed to “verify” (to myself) assumptions and mathematical models of multi-scale assembly, as the behavior of the agents is really very simple (gravitationally inspired, actually), but the coordination entirely scale free.