Re: [agi] AGI bottlenecks
I had similar feelings about William Pearson's recent message aboutsystems that use reinforcement learning: A reinforcement scenario, from wikipedia is defined as "Formally, the basic reinforcement learning model consists of: 1. a set of environment states S; 2. a set of actions A; and 3. a set of scalar "rewards" in the Reals. " Here is my standard response to Behaviorism (which is what the above reinforcement learning model actually is): Who decides when the rewards should come, and who chooses what are the relevant "states" and "actions"?If you find out what is doing *that* work, you have found your intelligent system. And it will probably turn out to be so enormously complex, relative to the reinforcement learning part shown above, that the above formalism (assuming it has not been discarded by then) will be almost irrelevant.Just my deux centimes' worth.I've been looking at something like this, but on a very generalized scale. Part 1 is infinite amount of states, incalculable,Part 2 is a large huge amount, that can be broken into parts and (must be) dealt with.Part 3 is what I am looking at currently.Is it possible for us to generate a model calculation of Worth, or Happiness Value or GoodState value.Just playing around with some ideas I have come up with:GoodValue = alive * (a*health + b*wealth + c*enjoyment + d*learning + e*friends) +pastplans -timeThis covers many simple motivations of humans currently, with a couple of these vague ideas that would need to be fleshed out. There is also a pastplans, which is a general preference for doing some things as patterns of past actions.This is general right now, but would it be possible to flesh out a comlpex, changeable, but still sizeably managable equation here that could be a controlling factor on AI motivations?James RatcliffThank YouJames Ratcliffhttp://FallsTown.com - Local Wichita Falls Community Websitehttp://Falazar.com - Personal WebsiteHosting Starting at $9.95Dialups Accounts - $8.95 __Do You Yahoo!?Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] AGI bottlenecks
Richard, Can you explain differently, in other words the second part of this post. I am very interested in this as a large part of an AI system. I believe in some fashion there needs to be a controlling algorithm that tells the AI that it is doing "Right" be it either an internal or external human reward. We receive these rewards in our daily life, in our jobs relationships and such, wether we actually learn from these is to be debated though.James RatcliffRichard Loosemore [EMAIL PROTECTED] wrote: Will,Comments taken, but the direction of my critique may have gotten lost in the details:Suppose I proposed a solution to the problem of unifying quantum mechanics and gravity, and suppose I came out with a solution that said that the unified theory involved (a) a specific interface to quantum theory, which I spell out in great detail, and (b) ditto for an interface with geometrodynamics, and (c) a linkage component, to be specified.Physicists would laugh at this. What linkage component?! they would say. And what makes you *believe* that once you sorted out the linkage component, the two interfaces you just specified would play any role whatsoever in that linkage component? They would point out that my "linkage component" was the meat of the theory, and yet I had referred to in such a way that it seemed as though it was just an extra, to be sorted out later.This is exactly what happened to Behaviorism, and the idea of Reinforcement Learning. The one difference was that they did not explicitly specify an equivalent of my (c) item above: it was for the cognitive psychologists to come along later and point out that Reinforcement Learning implicitly assumed that something in the brain would do the job of deciding when to give rewards, and the job of deciding what the patterns actually were and that that something was the part doing all the real work. In the case of all the experiments in the behaviorist literature, the experimenter substituted for those components, making them less than obvious.Exactly the same critique bears on anyone who suggests that Reinforcement Learning could be the basis for an AGI. I do not believe there is still any reply to that critique.Richard LoosemoreWilliam Pearson wrote: On 01/06/06, Richard Loosemore <[EMAIL PROTECTED]> wrote: I had similar feelings about William Pearson's recent message about systems that use reinforcement learning: A reinforcement scenario, from wikipedia is defined as "Formally, the basic reinforcement learning model consists of:1. a set of environment states S; 2. a set of actions A; and 3. a set of scalar "rewards" in the Reals. " Here is my standard response to Behaviorism (which is what the above reinforcement learning model actually is): Who decides when the rewards should come, and who chooses what are the relevant "states" and "actions"? The rewards I don't deal with, I am interested in external brain add-ons rather than autonomous systems, so the reward system will be closely coupled to a human in some fashion. The rest of post I was trying to outline a system that could alter what it considered actions and states (and bias, learning algorithms etc). The RL definition was just there as an example to work against. If you find out what is doing *that* work, you have found your intelligent system. And it will probably turn out to be so enormously complex, relative to the reinforcement learning part shown above, that the above formalism (assuming it has not been discarded by then) will be almost irrelevant. The internals of the system will be enormously more complex compared to the reinforcement part I described. But it won't make that irrelevent. What goes on inside a PC is vastly more complex than the system that governs the permissions of what each *nix program can do. This doesn't mean the permission governing system is irrelevent. Like the permissions system in *nix the reinforcement system it is only supposed to govern who is allowed to do what, not what actually happens. Unlike the permission system it is supposed to get that from the affect of the programs on the environment. Without it both sorts of systems would be highly unstable. I see it as a necessity for complete modular flexibility. If you get one of the bits that does the work wrong, or wrong for the current environment, how do you allow it to change? Just my deux centimes' worth. Appreciated. On a more positive note, I do think it is possible for AGI researchers to work together within a common formalism. My presentation at the AGIRI workshop was about that, and when I get the paper version of the talk finalized I will post it somewhere. I'll be interested, but sceptical. Will --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED] ---To unsubscribe, change your address, or temporarily deactivate your subscription, please go to
Re: [agi] AGI bottlenecks
James, It is a little hard to know where to start, to be honest. Do you have a background in any particular area already, or are you pre-college? If the latter, and if you are interested in the field in a serious way, I would recommend that you hunt down a good programme in cognitive science (and if possible do software engineering as a minor). After about three or four years of that, you'll have a better idea of where the below argument was coming from. Even then, expect to have to argue the heck out of your professors, only believe one tenth of everything they say, and discover your own science as you go along, rather than be told what the answers are. A lot of the questions do not have answers yet. All thinking systems do have a motivation system of some sort (what you were talking about below as rewards), but people's ideas about the design of that motivational system vary widely from the implicit and confused to the detailed and convoluted (but not necessarily less confused). The existence of a motivational system was not the issue in my post: the issue was exactly *how* you design that motivation system. Behaviorism (and reinforcement learning) was a suggestion that took a diabolically simplistic view of how that motivation system is supposed to work so simplistic that, in fact, it swept under the carpet all the real issues. What I was complaining of was a recent revival in interest in the idea of reinforcement learning, in which people were beginning to make the same stupid mistakes that were made 80 years ago, without apparently being aware of what those stupid mistakes were. (To give you an analogy that illustrates the problem: imagine someone waltzes into Detroit and says It ain't so hard to beat these Japanese car makers: I mean, a car is just four wheels and a thing that pushes them around. I could build one of those in my garage and beat the pants off Toyota in a couple of weeks! A car is not four wheels and a thing that pushes them around. Likewise, an artificial general intelligence is not a set of environment states S, a set of actions A, and a set of scalar rewards in the Reals.) Watching history repeat itself is pretty damned annoying. Richard Loosemore James Ratcliff wrote: Richard, Can you explain differently, in other words the second part of this post. I am very interested in this as a large part of an AI system. I believe in some fashion there needs to be a controlling algorithm that tells the AI that it is doing Right be it either an internal or external human reward. We receive these rewards in our daily life, in our jobs relationships and such, wether we actually learn from these is to be debated though. James Ratcliff */Richard Loosemore [EMAIL PROTECTED]/* wrote: Will, Comments taken, but the direction of my critique may have gotten lost in the details: Suppose I proposed a solution to the problem of unifying quantum mechanics and gravity, and suppose I came out with a solution that said that the unified theory involved (a) a specific interface to quantum theory, which I spell out in great detail, and (b) ditto for an interface with geometrodynamics, and (c) a linkage component, to be specified. Physicists would laugh at this. What linkage component?! they would say. And what makes you *believe* that once you sorted out the linkage component, the two interfaces you just specified would play any role whatsoever in that linkage component? They would point out that my linkage component was the meat of the theory, and yet I had referred to in such a way that it seemed as though it was just an extra, to be sorted out later. This is exactly what happened to Behaviorism, and the idea of Reinforcement Learning. The one difference was that they did not explicitly specify an equivalent of my (c) item above: it was for the cognitive psychologists to come along later and point out that Reinforcement Learning implicitly assumed that something in the brain would do the job of deciding when to give rewards, and the job of deciding what the patterns actually were and that that something was the part doing all the real work. In the case of all the experiments in the behaviorist literature, the experimenter substituted for those components, making them less than obvious. Exactly the same critique bears on anyone who suggests that Reinforcement Learning could be the basis for an AGI. I do not believe there is still any reply to that critique. Richard Loosemore William Pearson wrote: On 01/06/06, Richard Loosemore wrote: I had similar feelings about William Pearson's recent message about systems that use reinforcement learning: A reinforcement scenario, from wikipedia is defined as Formally, the basic reinforcement
Motivational system was Re: [agi] AGI bottlenecks
On 09/06/06, Richard Loosemore [EMAIL PROTECTED] wrote: Likewise, an artificial general intelligence is not a set of environment states S, a set of actions A, and a set of scalar rewards in the Reals.) Watching history repeat itself is pretty damned annoying. While I would agree with you the set of environmental states and actions are not well defined for anything we would call intelligence. I would argue the concept of rewards, probably not Reals, does have a place in understanding intelligence. It is very simple and I wouldn't apply it to everything that behaviourists would (we don't get direct rewards for solving crossword puzzles). But there is a necessity for a simple explanation for how simple chemicals can lead to the alteration of complex goals. How and why do we get addicted? What is it about morphine that allows the alteration of a brain to one that wants more morphine, when the desire for morphine didn't previously exist? That would be like bit flipping a piece of code or variable in an AI and then the AI deciding that bit-flipping that code was somehow good and should be sort after. The RL answer would be that the reward was variable altered. If your model of motivation can explain that sort of change, I would be interested to know more. Otherwise I have to stick with the best models I know. Will --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] AGI bottlenecks
Richard, I am a grad student and have studied this for a number of years already. I have dabbled in a few of the areas, but been unhappy in general with most peoples approaches as generally too specific (expert systems) or studying fringe problems of AI. I have been spending all my time reading papers and studying up on some sort of Strong AI, or general AI, but am dissappointed with many people here and in other places spending so much time on side issues, like calculating the actual date (which all seem wildly off) of the 'Singularity'.Aside: I have done some really interesting research with statistical extracting of facts from a large corpus, over 600 novels, and have had some really interesting results, but now am backing out to try and describe an overall AI agent sytem that has some consistency.James RatcliffRichard Loosemore [EMAIL PROTECTED] wrote: James,It is a little hard to know where to start, to be honest. Do you have a background in any particular area already, or are you pre-college? If the latter, and if you are interested in the field in a serious way, I would recommend that you hunt down a good programme in cognitive science (and if possible do software engineering as a minor). After about three or four years of that, you'll have a better idea of where the below argument was coming from. Even then, expect to have to argue the heck out of your professors, only believe one tenth of everything they say, and discover your own science as you go along, rather than be told what the answers are. A lot of the questions do not have answers yet.All thinking systems do have a motivation system of some sort (what you were talking about below as "rewards"), but people's ideas about the design of that motivational system vary widely from the implicit and confused to the detailed and convoluted (but not necessarily less confused). The existence of a motivational system was not the issue in my post: the issue was exactly *how* you design that motivation system.Behaviorism (and reinforcement learning) was a suggestion that took a diabolically simplistic view of how that motivation system is supposed to work so simplistic that, in fact, it swept under the carpet all the real issues. What I was complaining of was a recent revival in interest in the idea of reinforcement learning, in which people were beginning to make the same stupid mistakes that were made 80 years ago, without apparently being aware of what those stupid mistakes were.(To give you an analogy that illustrates the problem: imagine someone waltzes into Detroit and says "It ain't so hard to beat these Japanese car makers: I mean, a car is just four wheels and a thing that pushes them around. I could build one of those in my garage and beat the pants off Toyota in a couple of weeks!" A car is not "four wheels and a thing that pushes them around". Likewise, an artificial general intelligence is not "a set of environment states S, a set of actions A, and a set of scalar "rewards" in the Reals".)Watching history repeat itself is pretty damned annoying.Richard LoosemoreJames Ratcliff wrote: Richard, Can you explain differently, in other words the second part of this post. I am very interested in this as a large part of an AI system. I believe in some fashion there needs to be a controlling algorithm that tells the AI that it is doing "Right" be it either an internal or external human reward. We receive these rewards in our daily life, in our jobs relationships and such, wether we actually learn from these is to be debated though. James Ratcliff */Richard Loosemore <[EMAIL PROTECTED]>/* wrote: Will, Comments taken, but the direction of my critique may have gotten lost in the details: Suppose I proposed a solution to the problem of unifying quantum mechanics and gravity, and suppose I came out with a solution that said that the unified theory involved (a) a specific interface to quantum theory, which I spell out in great detail, and (b) ditto for an interface with geometrodynamics, and (c) a linkage component, to be specified. Physicists would laugh at this. What linkage component?! they would say. And what makes you *believe* that once you sorted out the linkage component, the two interfaces you just specified would play any role whatsoever in that linkage component? They would point out that my "linkage component" was the meat of the theory, and yet I had referred to in such a way that it seemed as though it was just an extra, to be sorted out later. This is exactly what happened to Behaviorism, and the idea of Reinforcement Learning. The one difference was that they did not explicitly specify an equivalent of my (c) item above: it was for the cognitive psychologists to come along later and point out that Reinforcement Learning implicitly assumed that something in the brain would do the job of deciding when to give
Re: [agi] AGI bottlenecks
On 01/06/06, Richard Loosemore [EMAIL PROTECTED] wrote: I had similar feelings about William Pearson's recent message about systems that use reinforcement learning: A reinforcement scenario, from wikipedia is defined as Formally, the basic reinforcement learning model consists of: 1. a set of environment states S; 2. a set of actions A; and 3. a set of scalar rewards in the Reals. Here is my standard response to Behaviorism (which is what the above reinforcement learning model actually is): Who decides when the rewards should come, and who chooses what are the relevant states and actions? The rewards I don't deal with, I am interested in external brain add-ons rather than autonomous systems, so the reward system will be closely coupled to a human in some fashion. The rest of post I was trying to outline a system that could alter what it considered actions and states (and bias, learning algorithms etc). The RL definition was just there as an example to work against. If you find out what is doing *that* work, you have found your intelligent system. And it will probably turn out to be so enormously complex, relative to the reinforcement learning part shown above, that the above formalism (assuming it has not been discarded by then) will be almost irrelevant. The internals of the system will be enormously more complex compared to the reinforcement part I described. But it won't make that irrelevent. What goes on inside a PC is vastly more complex than the system that governs the permissions of what each *nix program can do. This doesn't mean the permission governing system is irrelevent. Like the permissions system in *nix the reinforcement system it is only supposed to govern who is allowed to do what, not what actually happens. Unlike the permission system it is supposed to get that from the affect of the programs on the environment. Without it both sorts of systems would be highly unstable. I see it as a necessity for complete modular flexibility. If you get one of the bits that does the work wrong, or wrong for the current environment, how do you allow it to change? Just my deux centimes' worth. Appreciated. On a more positive note, I do think it is possible for AGI researchers to work together within a common formalism. My presentation at the AGIRI workshop was about that, and when I get the paper version of the talk finalized I will post it somewhere. I'll be interested, but sceptical. Will --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] AGI bottlenecks
Will, Comments taken, but the direction of my critique may have gotten lost in the details: Suppose I proposed a solution to the problem of unifying quantum mechanics and gravity, and suppose I came out with a solution that said that the unified theory involved (a) a specific interface to quantum theory, which I spell out in great detail, and (b) ditto for an interface with geometrodynamics, and (c) a linkage component, to be specified. Physicists would laugh at this. What linkage component?! they would say. And what makes you *believe* that once you sorted out the linkage component, the two interfaces you just specified would play any role whatsoever in that linkage component? They would point out that my linkage component was the meat of the theory, and yet I had referred to in such a way that it seemed as though it was just an extra, to be sorted out later. This is exactly what happened to Behaviorism, and the idea of Reinforcement Learning. The one difference was that they did not explicitly specify an equivalent of my (c) item above: it was for the cognitive psychologists to come along later and point out that Reinforcement Learning implicitly assumed that something in the brain would do the job of deciding when to give rewards, and the job of deciding what the patterns actually were and that that something was the part doing all the real work. In the case of all the experiments in the behaviorist literature, the experimenter substituted for those components, making them less than obvious. Exactly the same critique bears on anyone who suggests that Reinforcement Learning could be the basis for an AGI. I do not believe there is still any reply to that critique. Richard Loosemore William Pearson wrote: On 01/06/06, Richard Loosemore [EMAIL PROTECTED] wrote: I had similar feelings about William Pearson's recent message about systems that use reinforcement learning: A reinforcement scenario, from wikipedia is defined as Formally, the basic reinforcement learning model consists of: 1. a set of environment states S; 2. a set of actions A; and 3. a set of scalar rewards in the Reals. Here is my standard response to Behaviorism (which is what the above reinforcement learning model actually is): Who decides when the rewards should come, and who chooses what are the relevant states and actions? The rewards I don't deal with, I am interested in external brain add-ons rather than autonomous systems, so the reward system will be closely coupled to a human in some fashion. The rest of post I was trying to outline a system that could alter what it considered actions and states (and bias, learning algorithms etc). The RL definition was just there as an example to work against. If you find out what is doing *that* work, you have found your intelligent system. And it will probably turn out to be so enormously complex, relative to the reinforcement learning part shown above, that the above formalism (assuming it has not been discarded by then) will be almost irrelevant. The internals of the system will be enormously more complex compared to the reinforcement part I described. But it won't make that irrelevent. What goes on inside a PC is vastly more complex than the system that governs the permissions of what each *nix program can do. This doesn't mean the permission governing system is irrelevent. Like the permissions system in *nix the reinforcement system it is only supposed to govern who is allowed to do what, not what actually happens. Unlike the permission system it is supposed to get that from the affect of the programs on the environment. Without it both sorts of systems would be highly unstable. I see it as a necessity for complete modular flexibility. If you get one of the bits that does the work wrong, or wrong for the current environment, how do you allow it to change? Just my deux centimes' worth. Appreciated. On a more positive note, I do think it is possible for AGI researchers to work together within a common formalism. My presentation at the AGIRI workshop was about that, and when I get the paper version of the talk finalized I will post it somewhere. I'll be interested, but sceptical. Will --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] AGI bottlenecks
Yan King Yin wrote: We need to identify AGI bottlenecks and tackle them systematically. Basically the AGI problem is to: 1. design a knowledge representation 2. design learning algorithms 3. fill the thing with knowledge The difficulties are: 1. the KR may be inadequate, but the designer doesn't know it 2. learning algorithms are hard to design, or are inefficient 3. not enough brute force (funding etc) to fill the AGI with knowledge One thing I suggest is for all/most AGI groups to agree on a common KR, but it seems that differences among current AGI architectures are difficult to reconcile. If the KRs are different, then we're all on our own to design custom learning algorithms. Then the next thing we can do is to share knowledge-filling. We may vote on a common domain to experiment with, and then jointly develop a large training corpus (sharing the costs / labor). This seems to be the more feasible option. YKY This is too simple by a long way. I can design a KR easily enough, but KRs are not libraries, they are used by something. The using part is what counts: it might take five days to design the KR, five (or fifty) years to design and build the system that makes use of the KR. I had similar feelings about William Pearson's recent message about systems that use reinforcement learning: A reinforcement scenario, from wikipedia is defined as Formally, the basic reinforcement learning model consists of: 1. a set of environment states S; 2. a set of actions A; and 3. a set of scalar rewards in the Reals. Here is my standard response to Behaviorism (which is what the above reinforcement learning model actually is): Who decides when the rewards should come, and who chooses what are the relevant states and actions? If you find out what is doing *that* work, you have found your intelligent system. And it will probably turn out to be so enormously complex, relative to the reinforcement learning part shown above, that the above formalism (assuming it has not been discarded by then) will be almost irrelevant. Just my deux centimes' worth. On a more positive note, I do think it is possible for AGI researchers to work together within a common formalism. My presentation at the AGIRI workshop was about that, and when I get the paper version of the talk finalized I will post it somewhere. Richard Loosemore --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] AGI bottlenecks
On 5/30/06, Yan King Yin [EMAIL PROTECTED] wrote: We need to identify AGI bottlenecks and tackle them systematically. Basically the AGI problem is to: 1. design a knowledge representation 2. design learning algorithms 3. fill the thing with knowledgeWhat do you expect it to do after that? Or do you also wish to program a behaviour into it? ;-)These are essentially narrow AI approaches and IMO a fundamental AGI architecture should be able to evolve sensory/abstract knowledge representation, various learning strategies, cognitive/associative functions, motivation/emotion and complex behaviour on its own (and possibly evolve self-representation and self-awareness too). There is no single solution or set formula. Best way is to evolve diverse systems and over them , super-systems, to evolve interconnections among them. Of course the lower systems can be programmed or designed and the intermediate ones can be biased according to our 'taste'. But the higher layers must be flexible and raw. No algorithms and no filling with human kind of 'knowledge'. Unlike bio-AGIs, it won't take millions of years. The first generation may take say few years, next will take a few days and later generations will take only a few seconds to evolve into better systems.Sanjay To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
