From Article: 12062 in comp.ai.philosophy Newsgroups: comp.ai.philosophy Message-ID: <31mnh9$lvo@percy.cs.bham.ac.uk> References: <1994Jul28.163026.22838@galileo.cc.rochester.edu> Summary: Some simulations may be physically impossible Keywords: alife, symbols systems Date: 3 Aug 1994 00:14:33 GMT Organization: School of Computer Science, University of Birmingham, UK Subject: Re: Alife, real life, symbol systems From: A.Sloman@cs.bham.ac.uk (Aaron Sloman) [I may have got some of the attributions wrong.] jeff@aiai.ed.ac.uk (Jeff Dalton) writes, in response to Andrzej Pindor: > Date: Mon, 1 Aug 1994 17:44:53 GMT > Organization: AIAI, University of Edinburgh, Scotland (Jeff Dalton) > >>Can a virtual forest fire be hot? > >> (Greg Stevens I think ) > >> A virtual forest fire in a virtual world, if modeled accurately, is > >> as hot to virtual creatures in that world as one of our forest fires > >> is to us. It is hot WITH RESPECT TO that world, even if not with > >> respect to ours. (Jeff Dalton) > >>I think you might agree that any virtual worlds we've developed so > >>far (or computer simulations of various sorts) have not managed to > >>create virtual creatures that can experience heat in the way that > >>we and at least some animals do. [...] > >> (Andrzej Pindor) > >Unless you claim that there is an absolute meaning to the word "hot", > >whether something is hot (or not) is defined by its effects on other > >things. Things "hot" in the virtual world can have effects on other things > >in such a world which could be analogous to effects of "hot" things in > >the real world on other things here. Would it be then improper to say that > >they are (virtually) hot? (Jeff) > Well, suppose I'm playing a simulation game that includes forest > fires. (I don't know if any of the SimX games do, but we can imagine > a similar one that does. If you prefer to have a real example, > substitute the rain that occurs in SimAnt. The point is to have an > example from current simulation game technology). > > Now is _this_ forest fire hot? Or is the simulation not good enough? > Or is there a third possibility? In many philosophical discussions about criteria for something to be an X there's a serious danger of cross-purposes due to different (usually implicit) definitions of terms, like a discussion between a mathematician and a non-mathematician about whether a circle is "really" an ellipse. John W. Smith once pointed out to me that there are some things that can't be simulated without producing another of the same type: e.g. a picture of a picture is a real picture. I would argue that a simulation of a calculation is guaranteed to be a calculation in some sense. A simulation of a succession of events will be a succession of events. And so on. What about a simulation of a forest (never mind the fire)? Can the simulation contain real bits of bark, twigs, falling leaves, sap rising, butterflies fluttering by, etc.? Greg and Andrzej both seem to say yes. If you are inclined to the view that objects are defined by their causal powers and relationships as Andrzej suggests (he says "whether something is hot (or not) is defined by its effects on other things." and Greg seems to have similar views. (I've slightly generalised this to include causes as well as effects.) The problem is: how many of the causal relationships have to be simulated for it to be a forest (hot or otherwise)? There are enormously complex, varied, multi-layered causal processes at work in a forest: some at the level of what we can observe, some at microscopic levels, some at molecular or subatomic levels. Moreover, the causal relations extend outward in space and time, and possibly without any bounds, though the effects may be increasingly negligible as distance (in space-time) increases. Thus, perhaps it is impossible to simulate ALL the causal relations without simulating the whole universe, or at least that part of if that's reachable causally. This may be logically possible, but does not seem to be physically possible. It may even be logically impossible in general: Suppose we tried to put the computer doing the (complete) simulation into the forest. Then the simulation would have to include a (complete) simulation of the computer doing the simulation, and therefore a (complete) simulation of the simulation of the computer doing the simulation, and therefore..... However, even without the computer's presence it seems unlikely that all the causal relations can be simulated in full detail, though no doubt ever increasing subsets of them will be, as computers grow in power and shrink in physical size. This line of thinking suggests to me that a COMPLETE simulation of any portion of the physical universe including ALL causes and effects of that portion of the universe, is impossible except by replicating that portion of the universe and all the rest of the universe that's causally accessible from it. I.e. you can't get a replication by a simulation: it's the other way round. Whether a computer-based simulation can always replicate causal powers is itself a tricky question (discussed in my review of Penrose in the AI Journal 1993). For example engineers will often implement a system on a collection of computers including some redundancy, rather than a single computer in order to improve reliability. Now it is true that a single computer could simulate the redundant network: but could it ever replicate the extra reliability? If I am right, not all causal powers are capable of being replicated in a (single-processor) computer-based simulation: though not because of any argument based on dualism or mysticism, or special powers of the brain. The general argument is not based on causal powers of the brain (unlike the arguments of Searle and Penrose). If brains can't be simulated on a single computer (e.g. Turing machine) it may be for exactly the same reason that a fail-safe computer network cannot be. Anyhow, if some elaboration of this line of thought leads to a proof that you can't have a full simulation of a forest fire in any physically realisable computer, then it may help to justify the scepticism that Jeff was expressing, when he wrote > What I'm wondering is what's required for something in a simulation > or virtual world to count as heat. Is it _trivial_ to get heat, at > least sufficiently trivial so that ordinary simulation games might > already have accomplished it? Or does it require something more > sophisticated? And likewise for whether the virtual heat is hot > to creatures in the virtual world. How sophisticated does our > simulation of the creatures have to be? Maybe a simulation that leaves out some of the causal powers can be called a "low grade" simulation. What if most of the interesting causal powers are left out, as in computer games? This would challenge the confidence of Greg and Andrzej in the possibility of re-implementing all aspects of mind. However, it is still possible that various kinds of mental states and processes are not resistant to simulation because the relevant causal powers are separable from the implementation details, for the same reason that a simulation of a calculation is a calculation, and a simulation of a time-sharing process is a time-sharing process. I am inclined to think that most of the important features of mental states and processes (including qualia) are replicable in computers, possibly networks of computers to support asynchrony and some kinds of reliability. Aaron PS One reason the arguments are very tricky is that the analysis of the concept of "cause" (and its relationship to counterfactual conditional truths, about what would have happened if...) is one of the hardest problems in philosophy. Some very interesting and (as far as I know) highly original suggestions are made in Taylor, C.N, 1992 {A Formal Logical Analysis of Causal Relations} DPhil Thesis, Sussex University. Available as Cognitive Science Research Paper No.257 from Sussex University, School of Cognitive and Computing Sciences. (I was his supervisor, but he did the work independently of me.) > > I suspect there are conflicting intuitions on such points. I suspect > that some people regard it as trivial to get heat. All that's required > is that we can plausibly identify a correspondence, which we can > certainly do for current simulation games. (And, for that matter, > for films.) Other people will think that it's not trivial, but > there's no bar to simulations being good enough. > > I happen to disagree with the people (if there are any) who think its > trivial, but I'm not sure what I could say to convince them to change > their minds. So I'm hoping I can get a more precise characterization > for someone who doesn't think it's trivial but still thinks it's > possible. > > The Searle / Harnad position is, of course, that it's not possible. > I don't think > > A virtual forest fire in a virtual world, if modeled accurately, is > as hot to virtual creatures in that world as one of our forest fires > is to us. It is hot WITH RESPECT TO that world, even if not with > respect to ours. > > is a good enough answer as it stands. What does "modeled accurately" > amount to, for instance? > > >Talking about sensations of virtual creatures you are pushing the discussion > >into other tracks - would we ever be able to know what these creatures > >"feel", if anything? Hopefully you will agree that we are opening here > >a different can of worms. > > No, I think "hot to virtual creatures" already opened that can. > Note too that it's creatures and not, say, virtual stones or logs. > > -- jeff -- Aaron Sloman, School of Computer Science, The University of Birmingham, B15 2TT, England EMAIL A.Sloman@cs.bham.ac.uk OR A.Sloman@bham.ac.uk Phone: +44-(0)21-414-4775 Fax: +44-(0)21-414-4281 From Article: 12087 in comp.ai.philosophy Newsgroups: comp.ai.philosophy Message-ID: <31s30i$k7d@percy.cs.bham.ac.uk> References: <1994Jul28.163026.22838@galileo.cc.rochester.edu> <31mnh9$lvo@percy.cs.bham.ac.uk> <31nkft$qom@infa.central.susx.ac.uk> Date: 5 Aug 1994 01:01:06 GMT Organization: School of Computer Science, University of Birmingham, UK Subject: Re: Alife, real life, symbol systems From: A.Sloman@cs.bham.ac.uk (Aaron Sloman) inmanh@cogs.susx.ac.uk (Inman Harvey) writes: > Date: 3 Aug 1994 08:28:45 GMT > Organization: University of Sussex > > Aaron Sloman (A.Sloman@cs.bham.ac.uk) wrote: > : John W. Smith once pointed out to me that there are some things that > : can't be simulated without producing another of the same type: e.g. > : a picture of a picture is a real picture. I would argue that a > : simulation of a calculation is guaranteed to be a calculation in > : some sense. A simulation of a succession of events will be a > : succession of events. And so on. > > BUT... suppose I was directing a team of people on a bridge-building > project, where I knew that the calculation of materials needed is a 1 > day operation for 1 person, and the task of getting bureaucratic > permissions is a succession of formfilling and rubber-stamping which > requires 2 people for 1 week.... > > ....THEN in this context, and for my purposes of planning who does what job in > what order, I can simulate the calculation by a red square on a > wall-chart, the succession of bureaucratic operations by a blue line on > the wall-chart - and these squares/lines are neither calculations nor > successions. Your general warning about being careful with generalisations is well taken. And you have produced a nice try at a counter example to what I wrote. Unfortunately I think this becomes one of those many points that turn on definitions (like is a circle and ellipse?). I would be happy to say your red square was a representation of the calculation but I would be disinclined to say that it was a simulation of it: it's too atomic and my understanding of "simulate" and "simulation" includes a notion of structural decomposition. (Similarly, I would be uncomfortable calling a dot a picture of a man far away: "being a picture of" also seems to require some structural decomposition and mapping.) However, just as a picture need not be fully decomposed, as in a rouch sketch, or a stick figure of a person, so could you also produce a modified version of your counter-example in which a complex calculation was simulated only in the sense that the timing of its main sub-processes was simulated. This might be a useful exercise in doing some sort of estimation of how long a job was going to take. Similarly, a picture of a picture could leave out a lot of the detail of the original picture, though it would remain a picture. In that sense a simulation of a calculation could leave out so much detail that I no longer would be inclined to call it a calculation: looking at it one could not see a justification for the answer, which is a characteristic feature of a calculation. Having granted you all that, I think the point intended still stands, namely that there are some things that are inherently abstract structures (proofs, patterns, calculations) and if you produce a representation of such a thing in such a way as to duplicate all the structural details then you also replicate, not merely simulate. The rest of my note was claiming that when you are talking not about an abstract structure but about a collection of interacting causal processes then in some cases a computer based simulation may be incapable of replicating the thing because of the richness and diversity of the causal relationships (including counterfactuals). > > Some people would argue (myself included) that there is a regrettable > tendency for people to propose scenarios, eg scenarios of simulation, in > what they think is a context-free manner, forgetting that simulations > are things done by somebody for a purpose. This leads them into all the > confusion associated with objectivism; the unstated assumptions that are > 'obvious' to one protagonist differ from those that are 'obvious' to > another. I agree that that can be a mistake. I think my mistake wasn't leaving out the context, but simply not noticing that a simulation of a calculation might not simulate sufficient detail to be able to fulfil the functions of a calculation (namely, not merely ending with the right answer but also providing a demonstration that it is the right answer). So I should have defined more precisely what sort of simulation I was talking about. > .... (Jeff Dalton) > : > What I'm wondering is what's required for something in a simulation > : > or virtual world to count as heat. Is it _trivial_ to get heat, at > : > least sufficiently trivial so that ordinary simulation games might > : > already have accomplished it? Or does it require something more > : > sophisticated? And likewise for whether the virtual heat is hot > : > to creatures in the virtual world. How sophisticated does our > : > simulation of the creatures have to be? (Inman) > ...to count as heat for whom, in what context? > ...to count as 'hot to creatures in the virtual world' for whom, in what > context? I am not as much of a relativist as you are. If something is hot then it's hot because it has all those atoms/molecules with high kinetic energy, not because there's some observer around ``for'' whom it is hot, whatever that might mean. But `simulation' is a functional concept and what fulfils the function may depend on who needs it. So you are right about some cases. > > Personally I blame it all on the use of the passive tense making people forget > that you always require a subject to do simulating, computing, representing, > etc, and you must be prepared to identify the subject. I would make use of the > passive tense in this fashion in philosophical discussion a criminal offence. > (:^' 1/2 a smiley, I am only half joking) Actually a far worse cause of muddle and confusion is the use of nouns and noun phrases, e.g. consciousness intelligence vision emotion, free-will (ugh!) etc. People should stick to adjectives and adverbs when trying to talk about these things. (I think it was the philosopher J.L,Austin who pointed this out.) Aaron -- Aaron Sloman, School of Computer Science, The University of Birmingham, B15 2TT, England EMAIL A.Sloman@cs.bham.ac.uk OR A.Sloman@bham.ac.uk Phone: +44-(0)21-414-4775 Fax: +44-(0)21-414-4281 From Article: 12189 in comp.ai.philosophy Newsgroups: comp.ai.philosophy Message-ID: <32igsd$j11@percy.cs.bham.ac.uk> References: <326esf$jio@tadpole.fc.hp.com> Date: 13 Aug 1994 13:12:45 GMT Organization: School of Computer Science, University of Birmingham, UK Subject: Re: Aaronn, Is a quale an object or an abstraction? From: A.Sloman@cs.bham.ac.uk (Aaron Sloman) [edited later, replacing "qualium" with "quale" as singular of "qualia"] Apologies for delayed response. I have been away. (Also, sending me an email copy is advisable if a response is required, as I do not read everything in comp.ai.philosohy -- too busy.) allsop@fc.hp.com (Brent Allsop) writes: > Date: 8 Aug 1994 23:25:03 GMT > Organization: Hewlett-Packard Fort Collins Site > > In the "Alife, real life, symbol systems" thread Aaronn Sloman said: > > I am inclined to think that most of the important features of mental > > states and processes (including qualia) are replicable in computers, I guess including that comment was asking for trouble! (BRent asked) > Do you define neural based computational devices that use ************ === > qualia instead of transistor states to represent things as computers? =================================== +++++++++ I have great difficulty responding to this, because I am not sure how Brent interprets the phrases I have underlined using "*", "+" and "=". The only neural based computational devices I am aware of are animal brains (which I see as computers of a sort), though I suppose that if someone ever makes artificial neurons then things created using them might also be described as artifical "neural based" computational devices. (Of course, some people might define "computational" so narrowly as to exclude both natural and artificial neural based devices.) I have no reason to believe, as Brent, apparently does, that neural based computational devices (natural and artifical?) use qualia "instead of" tansistor states: I am not even sure what that could mean. Natural neural devices certainly use states of brain components, e.g. states of nerve cells, including, I presume chemical states, electrical states, and probably all sorts of things I don't understand. Whether and how qualia are involved in brains is another matter. This would require considerable discussion including unpacking the notion of a "quale" in some detail, but whatever their role is I don't see how qualia can replace transistor states. In computer-based intelligent agents qualia (in the sense in which I understand the term) would be objects in a high level virtual machine *implemented* in the machine made of transisters, but not *replacing* transistors. Similarly, if human minds contain qualia then I assume they are *implemented* in neural states and processes, but could not possibly *replace* them. This comment presupposes that the concept "quale" can be defined in an implementation- independent way, just as the concepts subtraction, data-structure, list-structure, 3-D array, etc. can be defined in implementation-independent ways. I suspect that my understanding of the concept "qualium" is very different from Brent's: since in my understanding there is no more sense in qualia replacing transistors than in subtraction or lists replacing transistors. I presume therefore that Brent assumes that qualia are states of neurons, which I don't, even though I believe that human qualia are *implemented* in states of neurons (and maybe other things in the brain, e.g. various chemical soups and tissues surrounding neurons). (brent) > Your usage here seems to indicate you are treating qualia as > an abstract notion Yes, in the sense in which most concepts referring to objects and processes in virtual machines are abstract notions. > ...like a number. No. There are differences between the data-structures used by the editor into which I am typing this message and numbers. The data-structures are abstract in a sense, in that they are not physical objects, they have no weight, size, electrical resistivity, temperature, etc. They are abstract in the sense in which most of the states and processes that concern software engineers are abstract: i.e. they exist in virtual machines (e.g. the lisp virtual machine, the C++ virtual machine) which can be implemented in all sorts of physically different concrete machines. Nevertheless the particular instances of abstract data-structures, e.g. the vector of strings making up my current editor buffer, are things that, *unlike numbers* can play a role in causal processes. When I press a key, that event causes one of the strings to be changed. When one of the strings is changed, or the "current cursor" location changes in the editor's virtual machine, then that causes what is displayed on my screen to change, and so on. (HOW events in virtual machines can enter into causal relations is a long and complex story. It is often assumed that they cannot, because of the mistaken assumption that only physical objects and events can have causal relationships). Thus, when I talk about qualia I am talking about things that are abstract insofar as they exist in virtual machines that could have different physical implementations, but, like my editor's data-structures, they are also causally efficatious, and certainly real. But they are different from numbers. (brent) > ..Numbers are abstract, and not real. (Many mathematicians would agree that they are abstract, but disagree with your claim that they are not real. I am not sure what it means to argue about the reality of numbers, as compared with arguing about the reality of the Loch Ness monster.) (brent goes on about numbers) > They must be represented by some object. That object can be a set of > beads on a string, the state of a set of transistors, or an idea in > the mind; or anything that can assume a state. Numbers are different from data-structures in a virtual machine. The number 93.275 is abstract in a different way from the vector of strings I am currently "typing into". The number exists in no particular place or time, and never changes (unlike the vector, whose components change). It does not enter into causal relationships. (Something that cannot change cannot be affected by causes.) Of course, there may be *representations* of numbers in virtual machines, and we often loosely refer to them as numbers (e.g. a 5 by 5 array might contain 15 numbers). Strictly we are talking about representations of numbers: abstract *numerals*. Those are data-structures, and are capable of entering into causal proceses. They are also implemented in lower level virtual machines and still lower level physical machines. E.g. a number in a lisp virtual machine might be represented as a certain bit-pattern on a SPARC or HP-PA machine and a different bit pattern on a DEC Alpha machine. Those bit patterns, in turn, will be implemented in states of transistors in physical machines (which will be different next year, no doubt). When I wrote > > I am inclined to think that most of the important features of mental > > states and processes (including qualia) are replicable in computers, I was expressing the view that mental states and processes occur in virtual machines that are capable of having different implementations, in something like the way in which states and processes in my editor occur in a virtual machine that can have different implementations. Changing the implementations may affect some properties of instances of those virtual machines, e.g. speed of operation, error conditions, robustness. In some cases, changing the implementation of a mind may introduce new limitations on some of the possible states (e.g. moving from one implementation to another with far more physical memory might allow far higher resolution visual processing, and therefore, perhaps, richer and more detailed visual qualia.) None of this actually tells you what I think qualia are, except to hint that they involve states of perceptual systems. In fact I think the concept (as used by philosophers) is a mixture of some good phenomenology (i.e. the concept refers to aspects of real human experience) and some bad philosophical theorising (i.e. the concept often carries extraneous baggage derived from a bad philosophical theory about what minds are.) I believe that the "good" part of the concept of a qualium can be given a role within a theory of the type of architecture required for a system to constitute an autonomous human-like agent whose perceptual systems involve several different layers of interpretation of sensory input where the architecture also provides a mechanism that is to some extent capable of inspecting internal system states. The full account of such an architecture and how qualia would fit in would have to be be long and complicated, and I don't think I have a finished version of it yet, which is why I wrote "I am inclined to think..." I.e. it's a conjecture about the outcome of future research in this area rather than any kind of dogmatic claim. Future research would show that I was wrong about qualia (as I understand the term) if robots could be built that were comparable in capabilities to humans, and as strongly inclined to report the existence of qualia in THEIR experience, but without any of the types of intermediate perceptual databases nor the types of internal self-monitoring of those databases, which I suspect (together) constitute the existence of qualia. (Don't ask me for a more precise or complete account of all this: I don't have one.) I think Brent was essentially disagreeing with the idea that qualia are somehow components of a virtual machine that could be implemented differently, just as (some would say) water is not a component of a virtual machine which could be implemented differently. He seems to think that an experience of painful heat or of freezing cold should somehow be thought of like a bucket of water: its a kind of stuff that exists in its own right, and not an emergent feature of a virtual machine. Anyhow that's how I interpret this part of his message: (Brent:) > Any cause and effect system can be modeled (or 'replicable') > by a different cause and effect system. Various abstract notions can > be represented by anything that can assume a state but this doesn't ^^^^^^^^^^^^^^^^ > change the fact that there are different fundamental things (earth, ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > fire, water, air, and the different ideas that represent them...???) ^^^^^^^^^^^^^^^^ I don't see why he is bringing in the ideas that represent them here, and I can't see the point of the next bit. I have tried to make sense of all this an can do so only if in some contexts I take Brent to use "idea of X" as equivalent to something like "experience of X", as some philosophers used to do (Hume, perhaps?). (Brent writes) > It seems to me that "What are ideas?" is the same question as: "What > is earth?". What follows seems to fit my interpretation of his views, provided I read "idea" as "experience", rather than as "concept". (Brent) > .We know that one idea (painful heat) is different than > another idea (freezing cold). Just as we know that earth is different > than water, and a transistor that is on is different than a transistor > that is off. What is the difference between earth and water or a > transistor that is on and off? Although we know much about what each > of these are "like" or the differences between them we still don't > know fundamentally why the two are and why they are able to be > different both causely and stately. Someday we may know as much about > hot and cold ideas as we know about the hot and cold forest fires they > can represent but knowing fundamentally why they are and why they are > different is a much more difficult question. I am not sure I can make any sense of the question why things are, e.g. why electrons are, why clouds are. What they are is another matter: that's a topic for scientific investigation. Is that what you are talking about? (i.e. why use "why" rather than "what"?) (Brent) > I think it is an unproductive error to consider qualia as mere > abstractions. Who said anything about "mere" abstractions? The word "mere" suggests some kind of low grade existence. Virtual machines are NOT low grade. They can be very powerful (e.g. the operating system you are currently using.) The data-structures in my editor are abstractions in the sense I mentioned above (the same editor can run on different kinds of machines), but they are not "mere abstractions" in any sense of that phrase that I can understand: the data-structures play an important role in the operations of the editor, and that plays an important role in my replying to your questions. Similarly qualia, beliefs, desires, and other mental entities can exist in virtual machines that in principle are capable of being implemented in different ways, but each actual implementation will have real and causally efficacious components. They are not "mere" anything. >...Red is an abstraction that can be represented by light > of a certain wavelength, states of transistors, and a certain quale. > But this abstract notion of red must be physically realized in some > way even in our mind. What is the object or physical system that our > minds use to subjectively represent the abstract notion of heat? > Whatever it is is what it is "like", and systems of transistors and > their states, though similar, cause and effect wise, seem > fundamentally different to me state wise. Well of course the components of a high level virtual machine can be totally different from the components of a lower level virtual machine used to implement them. For example, the transistors used to implement my editor's buffer have a physical weight, and are composed of silicon and other physical elements. The vector of strings has no weight (NB that doesn't mean its weight is 0), and it certainly is not composed of silicon: it's a totally different kind of thing, to which these categories are not applicable. The concept of a virtual machine in which causes and effects can occur, and which can have causal relationships with other things outside the machine is one that many people find hard to grasp. But it is used very effectively every day by software engineers. Brent may have an ontology that is fundamentlly different from mine, and perhaps nothing I can say will change his, nor vice versa. However, lengthy face to face tutorial discussion can sometimes get people to see that they are using an ontology different from the one they think they are using. I've tried to elaborate on some of the above points in a paper I have mentioned in previous postings: Title: Semantics in an intelligent control system Available as: ftp://ftp.cs.bham.ac.uk/pub/dist/papers/cog_affect/semantics.ps.Z It will be published some time in 1994/5 in philosophical transactions of the Royal Society. Of course, it may be all wrong! Aaron -- Aaron Sloman, School of Computer Science, The University of Birmingham, B15 2TT, England EMAIL A.Sloman@cs.bham.ac.uk OR A.Sloman@bham.ac.uk Phone: +44-(0)21-414-4775 Fax: +44-(0)21-414-4281 From Mon Dec 19 02:29:56 GMT 1994 Newsgroups: comp.ai.philosophy References: <3cl8ov$dr8@cantaloupe.srv.cs.cmu.edu> <3cvc9r$1q4@newsbf02.news-fddi.aol.com> Subject: When is a simulation of a Y a Y? (Was Bag the Turing test) This is a first attempt to clarify the question, triggered by (among other things) by comments written by Hans Moravec. But I may be too sleepy to clarify anything, just now. hpm@frc.ri.cmu.edu (Hans Moravec) writes: > Date: 18 Dec 1994 04:23:56 GMT > Organization: Field Robotics Center, CMU > > jrstern@aol.com (JRStern): > > The cliche is that "map is not the territory". > > Oh but it is, if both are schemas of abstract relationships. > A simulation of arithmetic is arithmetic, for instance, and > a simulation of a mind is a mind. > > --Hans Moravec CMU Robotics We need a survey of cases where the schema A F of a Y is a Y is true. Examples of the schema, not all true, would be "A description of a Y is a Y", "A simulation of a Y is a Y", "A copy of a Y is a Y". Sometimes whether this sort of thing is true or not depends on a culture or legal system: e.g. A copy of a cheque (check in USA) or will or certificate would not necessarily be a cheque, will or certificate. In a logical notation the abstract schema might be expressed something like this (using "A" as the universal quantifier "E" as the existential quantifier and -> as some form of strict implication): Ax( Ey(Y(y) & F(y,x)) -> Y(x) ) Whether this is true or not in a particular case, is surely going to depend on the relation between the relation F and the predicate Y. E.g. if Y is "Cheque" and F is "Copy" the instance is Ax( Ey(Cheque(y) & Copy(y,x)) -> Cheque(x) ) which appears not to be true (not in this country anyway). Of course, colloquial usage can be confusing. E.g. when playing with a child one may refer to a toy model of a horse as a horse, and yet when asked by a philosopher if it really is a horse many people would answer No. (Not all, I am sure.) One of my former students (John W. Smith) many years ago wrote an essay arguing that there was a strong analogy between A picture of a picture is a picture and A simulation of a calculation is a calculation expressing views about minds very similar to those of Hans. How general is this? It is clearly not true in general that A description of a Y is a Y E.g. A description of a mountain is not a mountain although there are some special cases that are true, e.g. trivially: A description of a description is a description Sometimes, a simulation of a Y is definitely not a Y unless you believe that EVERYTHING is some kind of "schema of abstract relationships". E.g. various philosophers have argued that a simulation of a waterfall is not a waterfall, because there's nothing wet, etc. in the machine. They want not just an abstract pattern, but also the right CAUSAL powers and relationships. Others respond that for virtual agents in a virtual reality it WOULD be a waterfall, because they think being a waterfall is merely a matter of instantiating some suitable abstract schema, possibly involving other perceivers (like phenomenalist and idealist philosophers who define existence in terms of being experienced). I don't know if Hans wants to go that far. What he wrote suggested that he thought A simulation of a mind is a mind was a special case, not just an instance of the general claim A simulation of a Y is a Y, whatever Y is. I.e. when he says a mind or a map is a "schema of abstract relationships" I assume he thinks there's something SPECIAL about minds and maps. If he thinks EVERYTHING is merely a schema of abstract relationships then this needs to be defended explicitly. I can imagine this general claim being part of a coherent philosophical position, but the reasons for it need to be spelled out, and all the technical details about lookup tables and computers can't really be relevant because they would not apply to some of the other cases (e.g. waterfalls, earthquakes). Let's consider the weaker claim: If Y involves (merely?) being a schema of abstract relationships then it follows that A simulation of a Y is a Y What does it mean to say that being a Y is (merely?) being a "schema of abstract relationships" ? I presume it is something like this: There is some multi-way relationship R between a number of items such that, for every x, x is composed of N objects x1,...xN & R(x1,....xN) -> Y(x) i.e. being a Y is merely a matter of being composed of some number of objects in relationship R. Then, shortening that formula in the obvious way, we can express the claim that if Y involves only a schema of abstract relationships then a simulation of Y is a Y thus: For any predicate Y: Ax(( Composed(x, x1, ... xN) & R(x1,....xN) ) -> Y(x) ) -> Ax( Ey(Y(y) & Simulation(y,x)) -> Y(x) ) The claim about maps then is an example of this: Ax(( Composed(x, x1, ... xN) & R(x1,....xN) ) -> Map(x) ) -> Ax( Ey(Map(y) & Simulation(y,x)) -> Map(x) ) Whether you think that "Map(x)" satisfies the antecedent (i.e. that being a map is just a matter of satisfying an abstract schema of relationships), will depend on whether you think that everything that looks like a map (e.g. is a suitable 2-D configuration of marks on paper which could be interpreted as a map), or is isomorphic with some map, must actually be a map; or whether you think that being a map involves being a map OF an existing geographical area. The latter would require the relationship R above to be extended to include items that were not part of x, in the way that France is not part of a map of France though it is part of an extended (partly semantic) relationship involving the parts of a map and France itself. Some people would regard x being a map of z as requiring yet another thing to exist besides the components of x and z, namely some intelligent agent a FOR WHOM x is a map of z. The relation R might then, in the case of maps, involve something like this: Ea Ed R(x1, x2, ...xN, a, d) (I.e. there is an agent a, and a denoted object d such that R holds between the parts of x and a and d.) If this extended requirement were part of the concept of being a map then some simulations or other things that satisfied the simpler (internal) conditions for being capable of being a map might fail to be a map, e.g. because they are not maps of any real bit of the world, or because nobody actually ever takes them to be maps (e.g. they may be microscopic structures that nobody has ever seen, or portions of a wall in the shape of a map of France that nobody has ever marked out or seen as capable of being marked out, or objects like large collections of galaxies that are too big for anyone to observe and take to be a map). I think it is possible for something to be a simulation of a map which does not satisfy the extended conditions: e.g. it might be inaccessible to any agent, or it might not actually be used as a map of any actual place by any agent. So whether A simulation of a map is a map is true, depends on which concept of a map one uses. I am happy to accept that there's a generous concept of a map which merely involves having a structure that gives something the potential to be a map in the stricter sense. For the generous concept of a map the statement is correct. What about being a mind? Is there some complex relationship R between components of a complex entity such that the following is true? Ax(( Composed(x, x1, ... xN) & R(x1,....xN) ) -> Mind(x) ) from which we could derive Ax( Ey(Mind(y) & Simulation(y,x)) -> Mind(x) ) ? I.e. from which we could derive: A simulation of a mind is a mind? We would need to be able to show that a simulation essentially creates a new entity with components satisfying the relationships summarised in R. I.e. Ax( ( Ey(Mind(y) & Simulation(y,x) ) -> Ex1...ExN( Composed(x, x1, ... xN) & R(x1,....xN) ) A functionalist philosopher of mind typically believes this is true, though it may be hard to say what sorts of components and what sorts of relationships are involved in a mind (and therefore a simulation of a mind). I suspect that there are certain simulations of minds that don't have the "right" causal powers, because all they implement are "abstract" relationships between components of a mind, rather than real causal interactions. I started to discuss this in my review of Penrose's first attack on AI, but I am not happy that I have put the argument satisfactorily. The review, published in the AI journal in 1992, is also in ftp://ftp.cs.bham.ac.uk/pub/dist/papers/cog_affect/penrose.ps.Z Another problem is that some people (not all) think that being a mind is not merely a matter of having components related in a certain way, but also having some relationship to an environment, just as some people require a map to be related to something that it denotes. In the case of minds more may be needed, e.g. causal influences from the environment that can change the internal components, and also semantic relations linking the internal states to the environment. When we have these relationships properly described we may be in a better position to decide whether every simulation of a mind (at a certain level of detail) automatically satisfies those relationships. If so, then (that sort of) simulation of a mind is a mind. But we'd need to rule out some truly weak and shallow simulations. I have some sympathy for what Hans wrote, but I am not sure that a solid argument for that position can be produced. This is independent of whether the HLT could simulate a mind. One might argue that it can produce the appriate external behaviour without simulating the internal transitions. Whether one regards those as significant will depend on how one degines "mind", and why. I shall try to return to this topic when next I have time. In particular, instead of arguing over which of the formulas is TRUE of the notion of being a mind, we may be able to use the different cases to produce definitions of different sorts of mind. Then we might conclude that this particular simulation is a mind of type 5. I suspect there are many types. Cheers. Aaron --- [generated quite a lot of discussion]