Re: [agi] RSI - What is it and how fast?
sam kayley wrote: 'integrable on the other end'.is a rather large issue to shove under the carpet in five words ;) Indeed :-) For two AIs recently forked from a common parent, probably, but for AIs with different 'life experiences' and resulting different conceptual structures, why should a severed mindpart be meaningful without translation into a common representation, i.e. a language? If the language could describe things that are not introspectable in humans, it would help, but there would still be a process of translation which I see no reason to think would be anything like as fast and lossless as copying a file. And as Hall points out, even if direct transfer is possible, it may often be better not to do so to make improvement of the skill possible. Well, one relatively easy way to get at least part way around this would be for the two AGIs to define beforehand a common format for the sharing of skill data. This might allow for defining lots of things such as labeling inputs/outputs, what formats of input/output this skill module uses, etc. If then one AGI exported the skill to this format, and the other wrote an import function then I think this should be plausible. Or if an import function is too hard for some reason it could run the skill format on a skill format virtual machine and just feed it the right inputs and collect and use the outputs. Would such a chunk of bits also be able to be further learned/improved and not just used as an external tool by the new AGI? I'm not sure, but I would lean towards saying yes if the import code used by the 2nd AGI takes the skill format bits and uses those to generate an integrated mind module of its own special internal format. I think getting too far into these technical details is going beyond my own skills, so I should stop here and just retreat to my original idea: because the bits for a skill are there in the first AGI, and because two AGIs can transmit lossless data bits directly between themselves quickly (compared to humans), this could create at least hypothetically a "direct" skill sharing functionality which humans do not have. We all have a lot of hypotheses at this point in history, I am just trying to err towards caution rather than ones that could be dangerous if proven wrong. -- Brian Atkins Singularity Institute for Artificial Intelligence http://www.singinst.org/ - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: [agi] RSI - What is it and how fast?
Brian Atkins wrote: J. Storrs Hall wrote: Actually the ability to copy skills is the key item, imho, that separates humans from the previous smart animals. It made us a memetic substrate. In terms of the animal kingdom, we do it very, very well. I'm sure that AIs will be able to as well, but probably it's not quite as simple as simply copying a subroutine library from one computer to another. The reason is learning. If you keep the simple-copy semantics, no learning happens when skills are transferred. In humans, a learning step is forced, contributing to the memetic evolution of the skill. IMO, AGIs plausibly could actually transfer full, complete skills including whatever learning is part of it. It's all computer bits sitting somewhere, and they should be transferable and then integrable on the other end. If so, this is far more powerful, new, and distinct than a newbie tennis player watching a pro, and trying to learn how to serve that well over a period of years, or a math student trying to learn calculus. Even aside from the dramatic time scale difference, humans can never transfer their skills fully exactly in a lossless-esque fashion. 'integrable on the other end'.is a rather large issue to shove under the carpet in five words ;) For two AIs recently forked from a common parent, probably, but for AIs with different 'life experiences' and resulting different conceptual structures, why should a severed mindpart be meaningful without translation into a common representation, i.e. a language? If the language could describe things that are not introspectable in humans, it would help, but there would still be a process of translation which I see no reason to think would be anything like as fast and lossless as copying a file. And as Hall points out, even if direct transfer is possible, it may often be better not to do so to make improvement of the skill possible. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: Re: Re: [agi] Goals and subgoals
(2) the "supergoals vs. subgoals" issue --- this is where I disagree with what you said. Even though you mentioned topics like "goal alienation", you still suggest that to a large extent it is the "supergoals" that determine the system's goal-oriented activities, while I believe the system's behaviors are to a larger extent driven by the derived goals, or your "subgoals". No, I don't disagree with you on this point at all, I must have just phrased things misleadingly somehow... ben - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: Re: [agi] Goals and subgoals
On 12/7/06, Ben Goertzel <[EMAIL PROTECTED]> wrote: Pei, As usual, comparing my views to yours reveals subtle differences in terminology! It surely does, though this time there seems to be more than terminology. There are two issues: (1) the "implicit goals vs. explicit goals" issue --- we don't differ too much here. Now I see what you mean, which is like the "automatic behavior vs. deliberate behavior" distinction in psychology. In NARS, there are also "goals" (regulated activities) that the system doesn't represent or explicitly control. (2) the "supergoals vs. subgoals" issue --- this is where I disagree with what you said. Even though you mentioned topics like "goal alienation", you still suggest that to a large extent it is the "supergoals" that determine the system's goal-oriented activities, while I believe the system's behaviors are to a larger extent driven by the derived goals, or your "subgoals". Pei I can see now that my language of implicit versus explicit goals is confusing in a non-Novamente context, and actually even in a Novamente context. Let me try to rephrase the distinction IMPLICIT GOAL: a function that, from the perspective of an "objective observer" that is not the system, the system **appears to be** striving to maximize REFLECTIVELY EXPLICIT GOAL: a function that the system conceptualizes, to itself, as something it is trying to maximize REPRESENTATIONALLY EXPLICIT GOAL: a function that is explicitly represented in the system's knowledge representation, and that is (at least weakly) an implicit goal What you call "goals" in NARS, I believe, are what I would call "representationally explicit goals," which are generally going to be implicit goals only weakly. I.e., they are cognitive phenomena that the system -- may or may not explicitly conceptualize to itself, via its reflective awareness, as being goals (i.e. they may or may not be reflectively explicit goals) -- may or may not actually act as though it is maximizing, from a casual external observer's perspective -- However, the system **does** try to maximize them, from the perspective of an observer who is looking at the dynamics going on inside the software system. A difference between NARS and Novamente is that in Novamente many implicit goals will **not** be representationally explicit goals. What Richard seems to be arguing against is the assignment of a major role to representationally explicit goals. Because he does not like explicit knowledge representation I am not sure if Richard is claiming that "implicit goals" as I define them (which may be manifested inside a system as diffuse motivational processes -- e.g. a dog's goal of getting exercise, which it may well never explicitly represent in its mind) are not a useful tool for analyzing the behaviors of minds. -- Ben > Position statements: > > (1) The system's behaviors are driven by its existing tasks/goals. > (2) At any given time, there are multiple goals in the system. > (3) A goal is meant to be achieved by the execution of certain > operations. Usually, a goal is never fully achieved, but to a certain > degree. > (4) According to their origin, there are roughly two types of goals: > given and derived. The former either come from the system's initial > design (genetic code), or are imposed by the environment, while the > latter are derived from the former and the system's beliefs. > (5) A goal G1 and a belief B derives a goal G2, if the belief B states > that achieving G2 can help, more or less, the achieving of G1. > (6) Derived goals are "functionally autonomous", in the sense that the > system treat them in the same way as given goals. > (7) Goals are not necessarily consistent with each other. Even derived > goal G2 can turn out to be inconsistent with its parent goal G1, if > the parent belief B is wrong (which is always possible in a system > with insufficient knowledge). > (8) Existing goals compete for the system's available resources. The > system does not process goals in sequential order, but in parallel, > but distributing the resources among existing goals unevenly. > (9) The resources (mainly time, but also space) a goal gets is > determined by its priority, which depends on several factors, and > change from time to time. A goal is removed when its priority is too > low. > (10) In general, given goals have higher priority, and exist for > longer time, but it is not always the case. > > All of the above are already implemented in NARS. > > Now let's see our agreements and disagreements. > > > SUPERGOALS VERSUS SUBGOALS > > --- > > > > A supergoal is defined as a goal of a system that is not a subgoal of > > any other goal of that system, to a significant extent. > > This distinction is the same as my "given/derived" distinction. I > don't use your terms because it is often interpreted by people as > meaning that a "subgoal" is always taken to be a means to achieve a > su
Re: Re: [agi] Goals and subgoals
Pei, As usual, comparing my views to yours reveals subtle differences in terminology! I can see now that my language of implicit versus explicit goals is confusing in a non-Novamente context, and actually even in a Novamente context. Let me try to rephrase the distinction IMPLICIT GOAL: a function that, from the perspective of an "objective observer" that is not the system, the system **appears to be** striving to maximize REFLECTIVELY EXPLICIT GOAL: a function that the system conceptualizes, to itself, as something it is trying to maximize REPRESENTATIONALLY EXPLICIT GOAL: a function that is explicitly represented in the system's knowledge representation, and that is (at least weakly) an implicit goal What you call "goals" in NARS, I believe, are what I would call "representationally explicit goals," which are generally going to be implicit goals only weakly. I.e., they are cognitive phenomena that the system -- may or may not explicitly conceptualize to itself, via its reflective awareness, as being goals (i.e. they may or may not be reflectively explicit goals) -- may or may not actually act as though it is maximizing, from a casual external observer's perspective -- However, the system **does** try to maximize them, from the perspective of an observer who is looking at the dynamics going on inside the software system. A difference between NARS and Novamente is that in Novamente many implicit goals will **not** be representationally explicit goals. What Richard seems to be arguing against is the assignment of a major role to representationally explicit goals. Because he does not like explicit knowledge representation I am not sure if Richard is claiming that "implicit goals" as I define them (which may be manifested inside a system as diffuse motivational processes -- e.g. a dog's goal of getting exercise, which it may well never explicitly represent in its mind) are not a useful tool for analyzing the behaviors of minds. -- Ben Position statements: (1) The system's behaviors are driven by its existing tasks/goals. (2) At any given time, there are multiple goals in the system. (3) A goal is meant to be achieved by the execution of certain operations. Usually, a goal is never fully achieved, but to a certain degree. (4) According to their origin, there are roughly two types of goals: given and derived. The former either come from the system's initial design (genetic code), or are imposed by the environment, while the latter are derived from the former and the system's beliefs. (5) A goal G1 and a belief B derives a goal G2, if the belief B states that achieving G2 can help, more or less, the achieving of G1. (6) Derived goals are "functionally autonomous", in the sense that the system treat them in the same way as given goals. (7) Goals are not necessarily consistent with each other. Even derived goal G2 can turn out to be inconsistent with its parent goal G1, if the parent belief B is wrong (which is always possible in a system with insufficient knowledge). (8) Existing goals compete for the system's available resources. The system does not process goals in sequential order, but in parallel, but distributing the resources among existing goals unevenly. (9) The resources (mainly time, but also space) a goal gets is determined by its priority, which depends on several factors, and change from time to time. A goal is removed when its priority is too low. (10) In general, given goals have higher priority, and exist for longer time, but it is not always the case. All of the above are already implemented in NARS. Now let's see our agreements and disagreements. > SUPERGOALS VERSUS SUBGOALS > --- > > A supergoal is defined as a goal of a system that is not a subgoal of > any other goal of that system, to a significant extent. This distinction is the same as my "given/derived" distinction. I don't use your terms because it is often interpreted by people as meaning that a "subgoal" is always taken to be a means to achieve a supergoal, and is therefore is inferior. As I said above, this is indeed the case at the time of derivation, but not necessarily afterwards --- this is what "functional autonomy" means. > With this in mind, regarding creation and erasure of goals, there are > two aspects which I prefer to separate: > > 1) optimizing the set of subgoals chosen in pursuit of a given set of > supergoals. This is well-studied in computer science and operations > research. Not easy computationally or emotionally, but conceptually > straightforward to understand. This is the case if we ignore the insufficiency of knowledge/resources (which we should in CS and OR, but shouldn't in AI). > 2) optimizing the set of supergoals. This is a far, far subtler thing. > > Supergoal optimization must be understood from a perspective of > dynamical systems theory, not from a perspective of logic. Agree. However, the set of subgoals by itself does
Re: [agi] Goals and subgoals
Ben Goertzel wrote: Hi Richard, Once again, I have to say that this characterization ignores the distinctions I have been making between "goal-stack" (GS) systems and "diffuse motivational constraint" (DMC) systems. As such, it only addresses one set of possibilities for how to drive the behavior of an AGI. And once again, I have to say that I reject this crude binary classification of minds ;-) Oh, come now, that was not by any means a "crude binary classification of minds"!! ;-) For one thing, I said quite clearly (later in the message) that *both* types are needed. (And I am sure I have said that before, too.) My point was, and is, that most of the discussion that I ever see about "AGI goals" implicitly assumes the goal stack approach. It is still happening now: virtually all the discussion on this list is about points that are just downright meaningless in the context of a diffuse constraint driven system. My only reason for continually harping on about the issue is that it is frustrating to see so much wasted thought. What I tried to show (or at least to *claim*), by going through your post, was that several of the statements that you made about goals just did not make any sense when viewed from anything except a GS perspective. I am glad that you accept my point about the need for a diffuse constraint driven system, in your summary below, but I have to say that this did not seem to square with the analysis given in your original post. Richard Loosemore. By discussing goals, I was not trying to imply that all aspect of a mind (or even most) need to, or should, operate according to an explicit goal hierarchy. I believe that the human mind incorporates **both(( a set of goal stacks (mainly useful in deliberative thought), *and* a major role for diffuse motivational constraints (guiding most mainly-unconscious thought). I suggest that functional AGI systems will have to do so, also. -- Ben G - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: [agi] RE: [extropy-chat] Criticizing One's Own Goals---Rational?
Jef Allbright wrote: Ben Goertzel wrote: The relationship between rationality and goals is fairly subtle, and something I have been thinking about recently Ben, as you know, I admire and appreciate your thinking but have always perceived an "inside-outness" with your approach (which we have discussed before) in that your descriptions of mind always seem (to me) to begin from a point of pre-existing awareness. (I can think of immediate specific objections to the preceding statement, but in the interest of expediency in this low-bandwidth discussion medium, I would ask that you suspend immediate objections and look for the general point I am trying to make clear.) It seems to me that discussing AI or human thought in terms of goals and subgoals is a very "narrow-AI" approach and destined to fail in general application. Why? Because to conceive of a goal requires a perspective outside of and encompassing the goal system. We can speak in a valid way about the goals of a system, or the goals of a person, but it is always from a perspective outside of that system. It seems to me that a better functional description is based on "values", more specifically the eigenvectors and eigenvalues of a highly multidimensional model *inside the agent* which drive its behavior in a very simple way: It acts to reduce the difference between the internal model and perceived reality. [The hard part is how to evolve these recursively self-modifying patterns of behavior, without requiring natural evolutionary time scale.] Goals thus emerge as third-party descriptions of behavior, or even as post hoc internal explanations or rationalizations of its own behavior, but don't merit the status of fundamental drivers of the behavior. Does this make sense to you? I've been saying this for years, but have never gotten even a "huh?", let alone a "duh." ;-) - Jef This is identical to one of the points I was making when talking about diffuse contraint-driven motivational systems, though we are phrasing it differently. A system can be *relaxation* driven - it changes its state according to a large number of constraints that are always trying to do local gradient descent - in such a way that it looks approximately as if it were pursuing a kind of goal-seeking behavior. Thus: a boltzmann machine does not explicitly try to retrieve a previously matched associate of a pattern, it just relaxes its constraints until the pattern comes out. If a system had several relaxation mechanisms working simultaneously, each of these might seem to be a "goal". I dislike that word, as I have said before, precisely because it has connotations of explicitness that I don't buy, and because there is something else that really is an explicit goal (I intend to get in the car and go home later today: this is a real "goal"). Your point about people taking a perspective "outside" or "inside" the system is the same as saying that we should not be interpreting behavioral characteristics (in this case, movement towards "goals") as if they are directly represented inside the system by a mechanisms that explicitly encodes the goal and explicitly tries to achieve it. The early connectionists made this one of their big issues. (See the two PDP volumes for hundreds of repetitions of the same ideological statement). Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: Re: Re: [agi] Goals and subgoals
I believe that the human mind incorporates **both(( a set of goal stacks (mainly useful in deliberative thought), *and* a major role for diffuse motivational constraints (guiding most mainly-unconscious thought). I suggest that functional AGI systems will have to do so, also. Also, I believe that diffuse motivational constraints will often lead to what I call "implicit goals" as emergent patterns. Thus for instance animals with very limited deliberative, reflective awareness may still have powerfully clear implicit goals. -- Ben G - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: Re: [agi] Goals and subgoals
Hi Richard, Once again, I have to say that this characterization ignores the distinctions I have been making between "goal-stack" (GS) systems and "diffuse motivational constraint" (DMC) systems. As such, it only addresses one set of possibilities for how to drive the behavior of an AGI. And once again, I have to say that I reject this crude binary classification of minds ;-) By discussing goals, I was not trying to imply that all aspect of a mind (or even most) need to, or should, operate according to an explicit goal hierarchy. I believe that the human mind incorporates **both(( a set of goal stacks (mainly useful in deliberative thought), *and* a major role for diffuse motivational constraints (guiding most mainly-unconscious thought). I suggest that functional AGI systems will have to do so, also. -- Ben G - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: [agi] RE: [extropy-chat] Criticizing One's Own Goals---Rational?
Hi, It seems to me that discussing AI or human thought in terms of goals and subgoals is a very "narrow-AI" approach and destined to fail in general application. I think it captures a certain portion of what occurs in the human mind. Not a large portion, perhaps, but an important portion. Why? Because to conceive of a goal requires a perspective outside of and encompassing the goal system. We can speak in a valid way about the goals of a system, or the goals of a person, but it is always from a perspective outside of that system. But, the essence of human reflective, deliberative awareness is precisely our capability to view ourselves from a "perspective outside ourselves." ... and then use this view to model ourselves and ultimately change ourselves, iteratively... It seems to me that a better functional description is based on "values", more specifically the eigenvectors and eigenvalues of a highly multidimensional model *inside the agent* which drive its behavior in a very simple way: It acts to reduce the difference between the internal model and perceived reality. I wouldn't frame it in terms of eigenvectors and eigenvalues, because I don't know how you are defining addition or scalar multiplication on this space of "mental models." But I agree that the operation of "acting to reduce the difference between internal models and perceived reality" is an important part of cognition. It is different from explicit goal-seeking, which IMO is also important. Goals thus emerge as third-party descriptions of behavior, or even as post hoc internal explanations or rationalizations of its own behavior, but don't merit the status of fundamental drivers of the behavior. I distinguished "explicit goals" from "implicit goals." I believe that in your comments you are using the term "goal" to mean what I term "explicit goal." I think that some human behavior is driven by explicit goals, and some is not. I agree that the identification of the **implicit goals** of a system S (the functions the system S acts like it is seeking to maximize) is often best done by another system outside the system S. Nevertheless, I think that implicit goals are worth talking about, and can meaningfully be placed in a hierarchy. -- Ben - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
[agi] RE: [extropy-chat] Criticizing One's Own Goals---Rational?
Ben Goertzel wrote: > The relationship between rationality and goals is fairly > subtle, and something I have been thinking about recently Ben, as you know, I admire and appreciate your thinking but have always perceived an "inside-outness" with your approach (which we have discussed before) in that your descriptions of mind always seem (to me) to begin from a point of pre-existing awareness. (I can think of immediate specific objections to the preceding statement, but in the interest of expediency in this low-bandwidth discussion medium, I would ask that you suspend immediate objections and look for the general point I am trying to make clear.) It seems to me that discussing AI or human thought in terms of goals and subgoals is a very "narrow-AI" approach and destined to fail in general application. Why? Because to conceive of a goal requires a perspective outside of and encompassing the goal system. We can speak in a valid way about the goals of a system, or the goals of a person, but it is always from a perspective outside of that system. It seems to me that a better functional description is based on "values", more specifically the eigenvectors and eigenvalues of a highly multidimensional model *inside the agent* which drive its behavior in a very simple way: It acts to reduce the difference between the internal model and perceived reality. [The hard part is how to evolve these recursively self-modifying patterns of behavior, without requiring natural evolutionary time scale.] Goals thus emerge as third-party descriptions of behavior, or even as post hoc internal explanations or rationalizations of its own behavior, but don't merit the status of fundamental drivers of the behavior. Does this make sense to you? I've been saying this for years, but have never gotten even a "huh?", let alone a "duh." ;-) - Jef > To address the issue I will introduce a series of concepts > related to goals. > > SUPERGOALS VERSUS SUBGOALS > --- > > A supergoal is defined as a goal of a system that is not a > subgoal of any other goal of that system, to a significant extent. > > With this in mind, regarding creation and erasure of goals, > there are two aspects which I prefer to separate: > > 1) optimizing the set of subgoals chosen in pursuit of a > given set of supergoals. This is well-studied in computer > science and operations research. Not easy computationally or > emotionally, but conceptually straightforward to understand. > > 2) optimizing the set of supergoals. This is a far, far > subtler thing. > > Supergoal optimization must be understood from a perspective > of dynamical systems theory, not from a perspective of logic. > > A strongly self-modifying AI system will be able to alter its > own supergoals So can a human, to an extent, with a lot > of effort > > EXPLICIT VERSUS IMPLICIT GOALS > > > Next, I think it is worthwhile to distinguish two kinds of goals > -- explicit goals: those that a system believes it is pursuing > -- implicit goals: those a system acts like it is pursuing > > Definition: a "coherent goal achiever" is one whose implicit > goals and explicit goals are basically the same > > What is interesting, then, is the dynamics of coherent goal > achievers that are also strongly enough self-modifying to > modify their supergoals In this case, what properties > control the evolution of the supergoal-set over time? This > is closely related to Friendly AI, of course > > META-GOALS > > > Next, there is the notion of a "meta-goal", a supergoal > designed to coexist with other supergoals and to regulate the > process of supergoal creation/erasure/modification. > > For instance, a friend of mine has a metagoal of streamlining > and simplifying his set of supergoals. I have a metagoal of > making sure my various sometimes-contradictory supergoals all > cooperate with each other in an open and friendly way, rather > than being competitive and adversarial. > > RATIONALITY AND GOALS > --- > > To me, rationality has two aspects: > > 1) how effectively one achieves one's explicit goals, given > the constraints imposed by the resources at one's disposal. > > 2) how coherent one is as a goal-achiever (implicit goals = > explicit goals) > > IMO, revising one's supergoal set is a complex dynamic process that is > **orthogonal** to rationality. I suppose that Nietzsche > understood this, though he phrased it quite differently. His > notion of "revaluation of all values" is certainly closely > tied to the notion of supergoal-set refinement/modification > > Refining the goal hierarchy underlying a given set of > supergoals is a necessary part of rationality, but IMO that's > a different sort of process... > > In general, it would seem important to be aware of when
Re: [agi] Goals and subgoals
Ben, Very nice --- we do need to approach this topic in a systematic manner. In the following, I'll first make some position statements, then comment on your email. Position statements: (1) The system's behaviors are driven by its existing tasks/goals. (2) At any given time, there are multiple goals in the system. (3) A goal is meant to be achieved by the execution of certain operations. Usually, a goal is never fully achieved, but to a certain degree. (4) According to their origin, there are roughly two types of goals: given and derived. The former either come from the system's initial design (genetic code), or are imposed by the environment, while the latter are derived from the former and the system's beliefs. (5) A goal G1 and a belief B derives a goal G2, if the belief B states that achieving G2 can help, more or less, the achieving of G1. (6) Derived goals are "functionally autonomous", in the sense that the system treat them in the same way as given goals. (7) Goals are not necessarily consistent with each other. Even derived goal G2 can turn out to be inconsistent with its parent goal G1, if the parent belief B is wrong (which is always possible in a system with insufficient knowledge). (8) Existing goals compete for the system's available resources. The system does not process goals in sequential order, but in parallel, but distributing the resources among existing goals unevenly. (9) The resources (mainly time, but also space) a goal gets is determined by its priority, which depends on several factors, and change from time to time. A goal is removed when its priority is too low. (10) In general, given goals have higher priority, and exist for longer time, but it is not always the case. All of the above are already implemented in NARS. Now let's see our agreements and disagreements. SUPERGOALS VERSUS SUBGOALS --- A supergoal is defined as a goal of a system that is not a subgoal of any other goal of that system, to a significant extent. This distinction is the same as my "given/derived" distinction. I don't use your terms because it is often interpreted by people as meaning that a "subgoal" is always taken to be a means to achieve a supergoal, and is therefore is inferior. As I said above, this is indeed the case at the time of derivation, but not necessarily afterwards --- this is what "functional autonomy" means. With this in mind, regarding creation and erasure of goals, there are two aspects which I prefer to separate: 1) optimizing the set of subgoals chosen in pursuit of a given set of supergoals. This is well-studied in computer science and operations research. Not easy computationally or emotionally, but conceptually straightforward to understand. This is the case if we ignore the insufficiency of knowledge/resources (which we should in CS and OR, but shouldn't in AI). 2) optimizing the set of supergoals. This is a far, far subtler thing. Supergoal optimization must be understood from a perspective of dynamical systems theory, not from a perspective of logic. Agree. However, the set of subgoals by itself does not fully determine the system's future motivational structure, because it also heavily depends on the system's experience, which will determine the goals derived from the initial goal set. That is why I said in the AGIRI Workshop that "friendly AI" is mostly an education issue (experience control), rather than a pure design issue (initial goal selection). A strongly self-modifying AI system will be able to alter its own supergoals So can a human, to an extent, with a lot of effort Maybe, but since a mature AI system will mostly driven be derived goals, it is not always necessary to modify "supergoal", even if it is possible. EXPLICIT VERSUS IMPLICIT GOALS Next, I think it is worthwhile to distinguish two kinds of goals -- explicit goals: those that a system believes it is pursuing (in the sense of reflective, deliberative self-knowledge) -- implicit goals: those a system acts like it is pursuing (in the judgment of highly intelligent, unbiased observers) Do you mean something like intention vs. reality? Definition: a "coherent goal achiever" is one whose implicit goals and explicit goals are basically the same What is interesting, then, is the dynamics of coherent goal achievers that are also strongly enough self-modifying to modify their supergoals In this case, what properties control the evolution of the supergoal-set over time? This is closely related to Friendly AI, of course The definition looks fine, though I don't think such a system can be built --- no one can prove what he want will be what he get. META-GOALS Next, there is the notion of a "meta-goal", a supergoal designed to coexist with other supergoals and to regulate the process of supergoal creation/erasure/modification. For instance,
Re: [agi] Goals and subgoals
Ben Goertzel wrote: The topic of the relation between rationality and goals came up on the extropy-chat list recently, and I wrote a long post about it, which I think is also relevant to some recent discussions on this list... -- Ben Once again, I have to say that this characterization ignores the distinctions I have been making between "goal-stack" (GS) systems and "diffuse motivational constraint" (DMC) systems. As such, it only addresses one set of possibilities for how to drive the behavior of an AGI. My point in writing this reply is to make it clear that not everyone thinks that cognitive systems work this way. I have sketched some of the detailed differences below. Sorry if they are too compactly worded at times. *** SUPERGOALS VERSUS SUBGOALS --- A supergoal is defined as a goal of a system that is not a subgoal of any other goal of that system, to a significant extent. With this in mind, regarding creation and erasure of goals, there are two aspects which I prefer to separate: 1) optimizing the set of subgoals chosen in pursuit of a given set of supergoals. This is well-studied in computer science and operations research. Not easy computationally or emotionally, but conceptually straightforward to understand. When the target of the subgoals is defined as a set of complex constraints operating over a large portion of the system's knowledge, and when the system's knowledge is (at least sometimes) encoded in non-explicit ways (i.e. cannot be exactly stated in predicate form), the "optimization" of the subgoals becomes a type of problem that has not been studied by computer science and operations research. (We are talking about serious amounts of complex systems behavior here). 2) optimizing the set of supergoals. This is a far, far subtler thing. Not only subtler, but bordering on the incoherent. "Optimization" implies "making it work better," which in turn implies that the system has some concept of criteria that can be used to judge what counts as "better" and what counts as "worse". But if you had such criteria in hand, you could rephrase them like this: "I think that better behavior would result if I changed my aspirations consistent with these criteria" which could then be rephrased as "I aspire to improve my aspirations using these criteria" . and THAT, of course, is just a supergoal by any other name. But if that is a supergoal, is it higher than the others? And can it be optimized? And what then happens if you try to optimize that one? The concept of optimizing a supergoal seems to me to break down. [See also comments on "meta goal" below] Supergoal optimization must be understood from a perspective of dynamical systems theory, not from a perspective of logic. In the case of the DMC approach, I don't see dynamical systems theory getting any traction either. A strongly self-modifying AI system will be able to alter its own supergoals So can a human, to an extent, with a lot of effort In general, probably true. But the ease and extent of the self modification differs greatly between GS and DMC approaches. This is very much an open question. EXPLICIT VERSUS IMPLICIT GOALS Next, I think it is worthwhile to distinguish two kinds of goals -- explicit goals: those that a system believes it is pursuing (in the sense of reflective, deliberative self-knowledge) -- implicit goals: those a system acts like it is pursuing (in the judgment of highly intelligent, unbiased observers) This is something I agree with, but that is because I would see a cognitive system as having two separate systems: the motivational system, which uses DMC ideas, and a more mundane GS type of mechanism that deals with the day to day process of tracking explicit goals. In other words, I don't disagree with the need for something like a GS mechanism, but it plays a secondary role to the motivational system. One problem arises, however: there is not necessarily a sharp cutoff between the two, so goals can have different degrees of explictness. (See next comment for an illustration). Definition: a "coherent goal achiever" is one whose implicit goals and explicit goals are basically the same The "coherent goal achiever" concept would simply not make any sense in the DMC type of motivational system I have been describing. The two types of goals are implemented in completely different ways, and cannot be integrated to become one thing. However, if what is meant by this concept is a minimization of the inconsistencies between sets of goals (some of which are explicit and some less so), then, yes, a DMC system would be trying to do this as best it could. What is interesting, then, is the dynamics of coherent goal achievers that are also strongly enough self-modifying to modify their supergoals In this ca
Re: Re: [agi] Goals and subgoals
Another aspect I have had to handle is the different temperal aspects of goals/states, like immediate gains vs short term and long terms goals and how they can coexist together. This is difficult to grasp as well. In Novamente, this is dealt with by having goals explicitly refer to time-scope. But indeed, supergoals with different time-scopes are prime examples of supergoals that may contradict each other in practice (in terms of the subgoals they generate), though being in-principle consistent with each other. Your baby AGI currently is pursuing only goals externally given to it, but soon it would need to handle things like limited resources over time, and deciding on better goals for a longer term vs short term, and balancing the two. Agree ... we are not dealing with those things yet... Also how is your AGI handling the reward mechanism, is it just a simple additive number property that you are increasing via a 'pat on the head' or 'good boy' reward mechanism, or is it something internally created or? At the moment it's just a 'good boy' reward mechanism, which rewards concrete behaviors in the sim world. Most of the system's internal activities are not regulated by specific goal-achievement-seeking, but rather by the intrinsic activities of the system's cognitive processes. -- Ben - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
Re: [agi] Goals and subgoals
That sounds good so far. Now how can we program all of that :} Another aspect I have had to handle is the different temperal aspects of goals/states, like immediate gains vs short term and long terms goals and how they can coexist together. This is difficult to grasp as well. Your baby AGI currently is pursuing only goals externally given to it, but soon it would need to handle things like limited resources over time, and deciding on better goals for a longer term vs short term, and balancing the two. Also how is your AGI handling the reward mechanism, is it just a simple additive number property that you are increasing via a 'pat on the head' or 'good boy' reward mechanism, or is it something internally created or? James Ben Goertzel <[EMAIL PROTECTED]> wrote: The topic of the relation between rationality and goals came up on the extropy-chat list recently, and I wrote a long post about it, which I think is also relevant to some recent discussions on this list... -- Ben *** SUPERGOALS VERSUS SUBGOALS --- A supergoal is defined as a goal of a system that is not a subgoal of any other goal of that system, to a significant extent. With this in mind, regarding creation and erasure of goals, there are two aspects which I prefer to separate: 1) optimizing the set of subgoals chosen in pursuit of a given set of supergoals. This is well-studied in computer science and operations research. Not easy computationally or emotionally, but conceptually straightforward to understand. 2) optimizing the set of supergoals. This is a far, far subtler thing. Supergoal optimization must be understood from a perspective of dynamical systems theory, not from a perspective of logic. A strongly self-modifying AI system will be able to alter its own supergoals So can a human, to an extent, with a lot of effort EXPLICIT VERSUS IMPLICIT GOALS Next, I think it is worthwhile to distinguish two kinds of goals -- explicit goals: those that a system believes it is pursuing (in the sense of reflective, deliberative self-knowledge) -- implicit goals: those a system acts like it is pursuing (in the judgment of highly intelligent, unbiased observers) Definition: a "coherent goal achiever" is one whose implicit goals and explicit goals are basically the same What is interesting, then, is the dynamics of coherent goal achievers that are also strongly enough self-modifying to modify their supergoals In this case, what properties control the evolution of the supergoal-set over time? This is closely related to Friendly AI, of course META-GOALS Next, there is the notion of a "meta-goal", a supergoal designed to coexist with other supergoals and to regulate the process of supergoal creation/erasure/modification. For instance, one metagoal would be to streamline and simplify one's set of supergoals. Another metagoal would be to make one's various sometimes-contradictory supergoals all cooperate with each other in an open and friendly way, rather than being competitive and adversarial. RATIONALITY AND GOALS --- To me, rationality has two aspects: 1) how effectively one achieves one's explicit goals, given the constraints imposed by the resources at one's disposal. 2) how coherent one is as a goal-achiever (implicit goals = explicit goals) IMO, revising one's supergoal set is a complex dynamic process that is **orthogonal** to rationality. I suppose that Nietzsche understood this, though he phrased it quite differently. His notion of "revaluation of all values" is certainly closely tied to the notion of supergoal-set refinement/modification Refining the goal hierarchy underlying a given set of supergoals is a necessary part of rationality, but IMO that's a different sort of process... In general, it would seem important to be aware of when you are non-rationally revising a supergoal versus "merely" rationally modifying the set of subgoals used to achieve some supergoal. And yet, the two processes are very closely tied together. SUBGOAL PROMOTION AND ALIENATION One very common phenomenon is when a supergoal is erased, but one of its subgoals is promoted to the level of supergoal. For instance, originally one may become interested in science as a subgoal of achieving greatness, but later on one may decide seeking greatness is childish and silly, but retain the goal of advancing science as valuable in itself (now as a supergoal rather than a subgoal). When subgoal promotion happens unintentionally it is called subgoal "alienation." This happens because minds are not fully self-aware. A supergoal may be erased without all subgoals that it spawned being erased along with it. So, e.g. even though you give up your supergoal of drinking yourself to death, you may
[agi] Goals and subgoals
The topic of the relation between rationality and goals came up on the extropy-chat list recently, and I wrote a long post about it, which I think is also relevant to some recent discussions on this list... -- Ben *** SUPERGOALS VERSUS SUBGOALS --- A supergoal is defined as a goal of a system that is not a subgoal of any other goal of that system, to a significant extent. With this in mind, regarding creation and erasure of goals, there are two aspects which I prefer to separate: 1) optimizing the set of subgoals chosen in pursuit of a given set of supergoals. This is well-studied in computer science and operations research. Not easy computationally or emotionally, but conceptually straightforward to understand. 2) optimizing the set of supergoals. This is a far, far subtler thing. Supergoal optimization must be understood from a perspective of dynamical systems theory, not from a perspective of logic. A strongly self-modifying AI system will be able to alter its own supergoals So can a human, to an extent, with a lot of effort EXPLICIT VERSUS IMPLICIT GOALS Next, I think it is worthwhile to distinguish two kinds of goals -- explicit goals: those that a system believes it is pursuing (in the sense of reflective, deliberative self-knowledge) -- implicit goals: those a system acts like it is pursuing (in the judgment of highly intelligent, unbiased observers) Definition: a "coherent goal achiever" is one whose implicit goals and explicit goals are basically the same What is interesting, then, is the dynamics of coherent goal achievers that are also strongly enough self-modifying to modify their supergoals In this case, what properties control the evolution of the supergoal-set over time? This is closely related to Friendly AI, of course META-GOALS Next, there is the notion of a "meta-goal", a supergoal designed to coexist with other supergoals and to regulate the process of supergoal creation/erasure/modification. For instance, one metagoal would be to streamline and simplify one's set of supergoals. Another metagoal would be to make one's various sometimes-contradictory supergoals all cooperate with each other in an open and friendly way, rather than being competitive and adversarial. RATIONALITY AND GOALS --- To me, rationality has two aspects: 1) how effectively one achieves one's explicit goals, given the constraints imposed by the resources at one's disposal. 2) how coherent one is as a goal-achiever (implicit goals = explicit goals) IMO, revising one's supergoal set is a complex dynamic process that is **orthogonal** to rationality. I suppose that Nietzsche understood this, though he phrased it quite differently. His notion of "revaluation of all values" is certainly closely tied to the notion of supergoal-set refinement/modification Refining the goal hierarchy underlying a given set of supergoals is a necessary part of rationality, but IMO that's a different sort of process... In general, it would seem important to be aware of when you are non-rationally revising a supergoal versus "merely" rationally modifying the set of subgoals used to achieve some supergoal. And yet, the two processes are very closely tied together. SUBGOAL PROMOTION AND ALIENATION One very common phenomenon is when a supergoal is erased, but one of its subgoals is promoted to the level of supergoal. For instance, originally one may become interested in science as a subgoal of achieving greatness, but later on one may decide seeking greatness is childish and silly, but retain the goal of advancing science as valuable in itself (now as a supergoal rather than a subgoal). When subgoal promotion happens unintentionally it is called subgoal "alienation." This happens because minds are not fully self-aware. A supergoal may be erased without all subgoals that it spawned being erased along with it. So, e.g. even though you give up your supergoal of drinking yourself to death, you may involuntarily retain your subgoal of drinking (even though you started doing it only out of a desire to drink yourself to death). - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303