Hi Steve and Nick, Sorry to have dropped off. I tried to read the very vigorous thread, to the extent I could, as it went by. There is a lot there that seems to remain in the core of one thing that brings this crowd and several others together, and is conceptually far from finished business. I can’t aim for Nick’s precision, or Steve’s coverage, unless there is some particular thing to solve, so my way of doing these things tends to be more limited than the main thread was.
On Russ’s question, I tried to give a lecture in an informal summer school a couple of years ago, to propose what sequence of changes in physical architecture would justify bringing in each of a series of new concepts. I don’t have worked examples behind any of these cases (for a couple of them there are toy-model ideas), so this is the kind of work that is probably of little value and even less trustworthiness. I don’t remember exactly the layout, but I think the sequence contained something like this (ALLCAPS are meant to be informal descriptors for concept keywords): 1. Protected degrees of freedom are a precondition to even the possibility of MEMORY. If you are a mere physical degree of freedom, and you are always coupled to your environment, you are nothing different than an instant-by-instant reflection of the immediate local state of your environment. All of the later concepts in the list require various forms of internal state that have enough insulation to be protected from constant harassment. So where in the physical world are suitably decoupled degrees of freedom available to be found? (Much later, to be built, but not yet.) 2. Some kind of dynamical variables need to be capable of being couplers that can become DOORWAYS, so that the other DOF are sometimes coupled and sometimes not. A DOF that is always behind a wall (a chemical reaction behind such a high energy barrier that it is never achieved) can’t remember anything because, although it can certianly hold a state, it is never in contact with the environment that would imprint anything on that state. This doesn’t yet talk about how the open/close states of the doorway happen, which will determine when and what it allows the environment to imprint on the memory variable, and for how long that imprint can be held. Here one can be quite precese with examples without invoking biology. Organic chemistry at low energy in water is largely non-active. Metal centers, particular d-block elements, are the major doorways that govern the sectors of organic chemistry available to early ocean-rock worlds. Many enzymes still use them in something not too far from a mineral or soluble metal-ligand complex state, with a little tuning. In this case, the doorway works just through physical drift. Molecules free in solution are inert; those that bump into a metal can potentially become active; when they dissolve and drift on, they become inert again. This leads to a very different set of relations between thermal energy and information in reactions, than simple thermally-activated reactions among the same species. Probably one can invoke many other examples. 3. Some of the internal variables need to be capable of carrying on an AUTONOMOUS dynamics or internal process. I guess a memory variable can sit there passively and still, at some level, categorize the way a system (set of DOF) responds to an environmental event, but for most of the later levels, there needs to be actual internal dynamics. This in itself is not so hard; the world is far from equilibrium in any number of dimensions, and for something to be moving in a direction is not rare. 4. Internal dynamics can be autonomous, but it isn’t really “about” anything unless something about the configuration constitutes a MODEL in the sense of Conant and Ashby from old 1950s control theory. How the model is registered, and how reflexive or self-referential the internal dynamics needs to be for a meaningful model to be imprinted, probably ramify to many differenent questions. I would of course be happy to produce an interesting case of the emergence of any of them. 5. At some stage, a protected internal process of which the state of the model is part needs to act back on the doorway, if we are to be justified in saying the basic relation of a CONTROL SYSTEM has come into existence. Here again I intend a Conant and Ashby line of thought: that “Every good controller “contains? entails?” a model of the system controlled. There has to be some internal state that is capable of being in different relations to the state of the world, and then the internal dynamics has to take an input from a comparison of those two states. Only if the resulting action feeds back on the state, does the system start controlling its own interaction with the world (for instance, what gets remembered). 6. The next one is hard for me to say, even at the very low standards of the previous five: I can be a control system with a model of my world, even if I have only modest machinery. A membrane-bound protein that lets in some molecules and ignores others, and which is preserved in a population through some kind of filtering, is a perfectly good control-theoretic model in the C&A sense. But it only implicitly models its environment. I have not yet added the assumption that there is some kind of REFLEXIVITY or REFLECTION (in the sense of Quines) so that the model includes representations of possible counterfactual states of the internal variables themselves. If there is a physical process that drives a system’s parts into a configuration where that happens, then one of the things an internal process _could_ do is use the modeled futures to internally select among many responses to a situation of which it is capable. Only at that stage would I feel compelled to introduce a concept of AGENCY, where for my practical purposes, I am happy to use the word as game theorists use it. An agent is a kind of thing that fills one of the slots that games have for “players”, which must be provided for the mechanics of the game to execute, and where the agents have some way to convert specification of the game into a sequence of moves that are not individually dictated by the game itself. I am sure there are lots of other notions of agency (ABM has a much more permissive notion, which can be as little as a dynamical Monte Carlo, or can be full-blown game-theoretic player), but for the purpose of trying to draw levels from the foregoing, this one seemed enough to me to propose a concrete problem. I am sure there are more, but I think I stopped there, and this was about as far as Russ was asking, too, I think. So the challenge (speaking only for myself, of course) is to find places in matter where the structure of the dynamics as one starts with it, drives the activity into regions of material architecture that take on first one, and then another, of the above new patterns. I assume they have to occur in more or less this order, because it is hard for me to see how to build the later concepts without having the earlier ones as building blocks. I like chemistry as a medium, because the state space itself supports a lot of complexity, and the temporal variability of reactions, plus the fact that catalytic relations exist, offer large separations of timescales that can be used to fill functional classes like memories. Whether it becomes hard to build much hierarchy in any system that doesn’t benefit from the intensive way chemistry makes complexity easy, is a question I find interesting. I don’t know how one answers it with better than musing. This is all kind of armchair statements of the obvious, and I don’t mean to make it out as more. I know there are people like Rosen who made long careers of trying to tease all this out at length, and have written a lot on it. Maybe they include all this obvious stuff and also much more. But branching, to Nick’s point about the extent to which “a system” “chooses” something about the relevant delimitations of itself. I think this becomes an operationalizable question in the spirit of Leslie Valiant’s PAC learnability framework. (Probably Approximately Correct). Valiant’s wish was to show that the learnable tasks, like the computable functions, make a formally definable class. I don’t think that discussion is anywhere near being closed one way or another, but the attempt to systematize what can be learned, how hard it is, and how much either of those depends on the embedding context, seems very helpful and clarifying to me. The connection would be this: Suppose you have some internal state, and some internal dynamics, and the state under the influence of the dynamics — or even intervals within the dynamics under the influence of their longer trajectory — can pass through many different patterns. Suppose that somewhere there is a reinforcement learner working on those patterns in some systematic way. It could be an environment selectively filtering many copies of you with slight variations, or it could be some other subroutine within your internal dynamics. The kind of thing I have in mind is: suppose there is a synthetic organic chemistry generating small molecules, lots of copies of some of them, fewer copies of others, and as a by-product of that molecular pool, something like polymers large enough to be capable of function, but happening to have functions only at random, are one of the things that can arise. Out of all this mess, focus on the PAC-learnability problem of evolving an enzyme. The things that should determine whether a given selective protocol can find and then fix something should be: 1. how frequently is that substrate even encountered? If not often, it is hard to maintain any memory about it. It is easy for farmers to remember to water the crops during dry spring winds, because that happens every year. It is harder for a culture to remember to run uphill when the tide goes way out for a very long time, since maybe that hasn happened where they live in the past 600 years. 2. how consequential is the particular molecule. If very consequential, selection can be more severe, and leave a stronger signal, which maybe can be remembered a little longer. There is probably lots of other fine structure to learnability, such as whether the environment is effectively serving as a “teacher” with respect to that particular problem (Valiant’s term, used to illustrate concrete cases), but I won’t ramble more than this. How does that relate to Nick’s point; one more indirection on the way to getting there: Steve mentioned (in some thread, a few weeks ago) the concept of Order Parameter, which is a kind of predictive statistic that suddenly starts to have a lot more predictive content, and to be more stable, when a system goes into an ordered phase. If you are going to try to use reinforcement learning to select higher-order structure on some low-order patterns that you are already producing, the order parameters are the things that take the most regular values, and they most robustly support induction, which is what all reinfocement learning is. (A finite system cannot help but induct: in a world of potentially unlimited variability, it has only finitely many possible states, so perforce it will make infinite equivalence classes of environmental states, by responding to many situations that in detail are different, with the same response. That doesn’t mean that “the problem of induction” “has” any solution. It only means that every finite system is perforce a commitment to some inductive hypothesis.) So I would argue that, with respect to the accumulation of hierarchy, there is a natural sense of a system’s own delimitation, to the extent that the parts that are sufficiently stable and sufficiently consequential to build something on top of by reinforcement become the foundation that holds other parts together. I agree with the purpose underlying Steve G’s point: that this can depend in part on what kind of environment there is, since this is part of the learning protocol. But we also all recognize that — at least insofar as the statistical concepts found useful in equilibrium thermo and fundamental processes continue to be useful in more elaborate dynamical realms — the Order Parameter as a Minimum Sufficient Statistic for distributions over future states is an informationally special quantity. Sorry for long harangue, and I don’t know whether this has anything new in it that the list hasn’t revisited many times. All best, Eric > On May 29, 2017, at 8:29 PM, Stephen Guerin <stephen.gue...@simtable.com> > wrote: > > > > _______________________________________________________________________ > stephen.gue...@simtable.com > CEO, Simtable http://www.simtable.com > 1600 Lena St #D1, Santa Fe, NM 87505 > office: (505)995-0206 mobile: (505)577-5828 > twitter: @simtable > > On Mon, May 29, 2017 at 12:10 PM, Nick Thompson <nickthomp...@earthlink.net> > wrote: > SG, > > > > There are now THREE issues lurking here between us. > > > > IS THE CRITERION FOR A SYSTEM ARBITRARY: You say yes; I say no. We’ve > already covered that ground. > > > In my post, I said it is not arbitrary. It's a function of what the > researcher is trying to use it for or explain. > > > > > IS A HURRICANE A SYSTEM: For me, that is the question of whether the > collection of thunderstorms we call a hurricane interact with one another > more than they interact with their collective surroundings. Another way to > put this question is in terms of redundancy. If we were to go about > describing the movements of the thunderstorms of a hurricane, would we get a > simpler, less redundant description if we referred their movements to the > center of the hurricane. I think the answer to this question is clearly YES. > > > Yes you could model the movement in a simpler way by modeling the movement of > the center point. And that was my second model of a hurricane as a random > walker biased by a global wind vector and Coriolis curve term. And I said > that was not a complex system. > > > > > IS A HURRICANE COMPLEX? For me, complexity means “multi-layered” . So, a > complex system is one composed of other systems. A hurricane is a system of > thunderstorms which themselves are a system of thermals (handwaving, here). > Thus a hurricane is at least a three-level system. So, yes. It is complex. > > > I agree about complex systems as having multiple layers - a macro scale and a > micro scale. I would say there's one system. If I was trying to model a > hurricane in my first example of an emergent vortex dissipating temperature > and pressure gradients, I would model the air with a combination of air > particles and patches of air - at LANL they would describe these as particle > in a box models or hybrid lagrangian and eulerian models. I would not > introduce thunderstorms at the micro level. But there's many ways to skin a > hurricane :-) > > Some would say the micro level air particles and air cell components which I > would model as finite state machines (agents with a lower case "a") are > systems in their own right and have boundaries. I don't see the benefit of > calling them systems as their aren't multiple interacting components within > them. But don't feel like arguing too hard here. > > > Eric Smith? > > > > > Yes, where are you Eric Smith? ============================================================ FRIAM Applied Complexity Group listserv Meets Fridays 9a-11:30 at cafe at St. John's College to unsubscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ by Dr. Strangelove
