RE: [agi] Educating an AI in a simulated world
Hi Ben, If Novababies are going to play and learn in a simulated world which is most likely based on an agent-based/object-orientated programming foundation, would it be useful for the basic Novamente to have prebuilt capacity for agent-based modelling? Would this in be necessary if a Novababy is to process objects in their native format as suggested by Brad Wyble (eg. sprites in 3d coordinates). Cheers, Philip --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
RE: [agi] Educating an AI in a simulated world
Hi, This kind of built-in capability certainly isn't *necessary* but it might be useful. This kind of issue is definitely worth exploring... More thoughts later ;-) ben -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] Behalf Of Philip Sutton Sent: Saturday, July 19, 2003 11:24 AM To: [EMAIL PROTECTED] Subject: RE: [agi] Educating an AI in a simulated world Hi Ben, If Novababies are going to play and learn in a simulated world which is most likely based on an agent-based/object-orientated programming foundation, would it be useful for the basic Novamente to have prebuilt capacity for agent-based modelling? Would this in be necessary if a Novababy is to process objects in their native format as suggested by Brad Wyble (eg. sprites in 3d coordinates). Cheers, Philip --- 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] Educating an AI in a simulated world
It's an interesting idea, to raise Novababy knowing that it can adopt different bodies at will. Clearly this will lead to a rather different psychology than we see among humans --- making the in-advance design of educational environments particularly tricky!! First of all, please read Diaspora by Greg Egan. As a SF author, he excels in his informed approach to AI design, philosophy, and neuroscience. This book touches on this topic(AI's designed for multiple bodies) very directly. This VR training room initially seemed like a great idea to me, but on consideration, I'm not so certain it's worth the trouble. First of all, you are reducing the complexity of the environment by orders of magnitude. One could argue that it is a baby's physical interation with the world is the cornerstone on which all future intelligence resides. Now you've made pains to point out that you're not trying to recreate people, but intelligence. However, a Novamente grounded in a different reality will be difficult for people to interact with. So here are two possible issues: the VR world might actually slow down the intellectual growth of the Nova Baby. And even when intelligent, it will be more alien from us than it needs to be. A second point about this plan is that you are creating extra work for yourself both in designing a VR training paradigm, and then in bridging the gap from VR to the real world, which would be no picnic. So there are some possible negatives, the positives you've already listed. If this course is decided upon, consider giving the Novamente an ability to sense objects in their native format (sprites in 3d coordinates). If your intent is to simplify the world, don't add in the fuzz of the artificial visual input, which is often flawed (e.g. clipping errors). Give the Novababy access to the underlying framework of the world or it will be eternally confused as it tries to figure out why it can walk through trees, or why Mr. Smith's left toes are inside its own foot. -Brad --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
RE: [agi] Educating an AI in a simulated world
Brad wrote: It's an interesting idea, to raise Novababy knowing that it can adopt different bodies at will. Clearly this will lead to a rather different psychology than we see among humans --- making the in-advance design of educational environments particularly tricky!! First of all, please read Diaspora by Greg Egan. As a SF author, he excels in his informed approach to AI design, philosophy, and neuroscience. This book touches on this topic(AI's designed for multiple bodies) very directly. I read that book a few years ago, after Eliezer Y kindly sent me a copy (along with Fire Upon the Deep by Vernor Vinge). Those gifts from Eli reawakened my interest in SF, which had dwindled to just about zero in the early 90's, due to an apparently lack of decently interesting SF books beyond those I'd already read. I think Egan's speculative physics is pretty nifty [though not nearly crazy enough to be true, to paraphrase one of the founders of quantum theory], but, I wasn't so impressed by Egan's treatment of AI (though I agree it's not *idiotic* (unlike plenty of SF)). He has more uploaded humans than AI's, and even his AI's are pretty much derivatives of uploaded humans. He does play around with the difference between embodied and disembodied minds -- BUT, even his disembodied uploads seem to have pretty much the same old human-type sort of self that we have ... presumably because they're uploaded humans (or derivatives thereof) that haven't drifted too far from their roots... This VR training room initially seemed like a great idea to me, but on consideration, I'm not so certain it's worth the trouble. First of all, you are reducing the complexity of the environment by orders of magnitude. One could argue that it is a baby's physical interation with the world is the cornerstone on which all future intelligence resides. Now you've made pains to point out that you're not trying to recreate people, but intelligence. However, a Novamente grounded in a different reality will be difficult for people to interact with. So here are two possible issues: the VR world might actually slow down the intellectual growth of the Nova Baby. And even when intelligent, it will be more alien from us than it needs to be. A second point about this plan is that you are creating extra work for yourself both in designing a VR training paradigm, and then in bridging the gap from VR to the real world, which would be no picnic. So there are some possible negatives, the positives you've already listed. Agree about the possible negatives. However, I'm more worried about the VR worlds may be too simple to really be useful problem than about the too alien to communicate with us problem, because a) If the VR world is one in which humans can play too, I imagine communication will be possible in the context of that world. b) Humans can interact in the context of virtual worlds dissimilar from physical reality, and I suspect that similarly an AGI bred in a virtual world will be able to interact in the physical world ... *if* it really lives up to the G in AGI Of course, there will be a significant transition here, but I doubt it'll be an unbridgeable one. If this course is decided upon, First, I stress this is not something the Novamente engineering team will be working on right now. We are not yet at the stage where building a great training environment is our big problem. However, it may be that some additional resources to build this environment may present themselves -- because volunteers who don't have the training to work on AGI research proper, may possibly have the training to work on AI-training-world construction... Anyway, the course is not decided upon yet -- what's decided upon is merely that I'm going to spend a bit of time fleshing out the possibility. We wouldn't decide upon a major course of action such as this without far more careful evaluation... consider giving the Novamente an ability to sense objects in their native format (sprites in 3d coordinates). If your intent is to simplify the world, don't add in the fuzz of the artificial visual input, which is often flawed (e.g. clipping errors). Give the Novababy access to the underlying framework of the world or it will be eternally confused as it tries to figure out why it can walk through trees, or why Mr. Smith's left toes are inside its own foot. Very interesting suggestion, thanks!!! Why not, indeed, supply it with native format as well as useful but more derived/processed formats.. I want to add that we've also evaluated the possibility of using actual robotic systems for NM teaching/training, but have concluded that right now this would require significant financial resources, plus the collaboration of a dedicated robotics team. The situation is not as bad as when I last explored this possibility in the late 90's -- I built a small mobile robot hoping to use it to play with AI, but damn if I
Re: [agi] Educating an AI in a simulated world
Ben, I think there's a prior question to a Novamente learning how to perceive/act through an agent in a simulated world. I think the first issue is for Novamente to discover that, as an intrinsic part of its nature, it can interact with the world via more than one agent interface. Biological intelligences are born with one body and a predetermined set of sensors and actuators. Later humans learn that we can extend our powers via technologically added sensors and actuators. But an AGI is a much more plastic beast at the outset - it can be hooked to any number of sensor/actuator sets/combinations and these can be in the real world or in virtual reality. My guess is that it might be useful for an AGI to learn from the outset that it needs to make conscious choices about which sensor/actuator set to use when trying to interact with the world 'out there'. Probably to reduce early learning confusion it might be useful initially to give the AGI only 2 choices - between an agent that is a fixed-location box and an agent that is mobile - but with similar sensor sets so that it can fairly quickly learn that there is a relationship between what it perceives/learns via each sensor/actuator set. (Biligual children often learn to speak quite a bit later than monolingual children - the young AGI doesn't want to have early leaning hurdles set too high.) What I've said above I guess only matters if you are going to let a Novamente persist for a long period of time ie. you don't just reset it to the factory settings every time you run a learning session. If the Novamente persists as an entity for any length of time then its early learning is going to play a huge role in shaping its capabilities and personality. On a different matter, I think that it would be good for the AGI to learn to live initially in a world that is governed by the laws of physics, chemistry, ecology, etc. So, although the best initial learning environment might be virtual world (mainly to reduce the need for massive sensory processing power), I think that world should simulate the bounded/non magical nature of the real world we live in. Even if an AGI chooses to live in a non-bounded/magical virtual world most of the time in later life it needs to know that its fundamental existence is tied to a real world - it's going to need non-magical computational power and that's going to need real physical energy and it dependence on the real world has consequences for the other entities that live in the real world ie you and me an a few billion other people and some tens of million of other forms of life. Cheers, Philip --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
RE: [agi] Educating an AI in a simulated world
It's an interesting idea, to raise Novababy knowing that it can adopt different bodies at will. Clearly this will lead to a rather different psychology than we see among humans --- making the in-advance design of educational environments particularly tricky!! On the other hand, creating a virtual world that obeys the laws of physics, chemistry and so forth is simply not very plausible given the current state of physics, chemistry etc. Of course, we can make better or worse approximations, though. And, just as with the multiple-bodies idea, I think a multiple-virtual-worlds plan makes sense. The AGI will then grow up without identification to a specific body, and also without identification to a particular type of world! more later, Ben -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] Behalf Of Philip Sutton Sent: Saturday, July 12, 2003 9:56 AM To: [EMAIL PROTECTED] Subject: Re: [agi] Educating an AI in a simulated world Ben, I think there's a prior question to a Novamente learning how to perceive/act through an agent in a simulated world. I think the first issue is for Novamente to discover that, as an intrinsic part of its nature, it can interact with the world via more than one agent interface. Biological intelligences are born with one body and a predetermined set of sensors and actuators. Later humans learn that we can extend our powers via technologically added sensors and actuators. But an AGI is a much more plastic beast at the outset - it can be hooked to any number of sensor/actuator sets/combinations and these can be in the real world or in virtual reality. My guess is that it might be useful for an AGI to learn from the outset that it needs to make conscious choices about which sensor/actuator set to use when trying to interact with the world 'out there'. Probably to reduce early learning confusion it might be useful initially to give the AGI only 2 choices - between an agent that is a fixed-location box and an agent that is mobile - but with similar sensor sets so that it can fairly quickly learn that there is a relationship between what it perceives/learns via each sensor/actuator set. (Biligual children often learn to speak quite a bit later than monolingual children - the young AGI doesn't want to have early leaning hurdles set too high.) What I've said above I guess only matters if you are going to let a Novamente persist for a long period of time ie. you don't just reset it to the factory settings every time you run a learning session. If the Novamente persists as an entity for any length of time then its early learning is going to play a huge role in shaping its capabilities and personality. On a different matter, I think that it would be good for the AGI to learn to live initially in a world that is governed by the laws of physics, chemistry, ecology, etc. So, although the best initial learning environment might be virtual world (mainly to reduce the need for massive sensory processing power), I think that world should simulate the bounded/non magical nature of the real world we live in. Even if an AGI chooses to live in a non-bounded/magical virtual world most of the time in later life it needs to know that its fundamental existence is tied to a real world - it's going to need non-magical computational power and that's going to need real physical energy and it dependence on the real world has consequences for the other entities that live in the real world ie you and me an a few billion other people and some tens of million of other forms of life. Cheers, Philip --- 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] Educating an AI in a simulated world
Hi Ben, I think this is a great way to give one or more Novamentes the experience it/they need to develop mentally, in a controlled environment and in an environment where the need for massive computational power to handle sensory data is cut (I would imagine) hugely thus leaving Novamente a fair bit of computational power to do the cognitive self-development/thinking work. You've probably thought of this already, but the simulated environment could be the way for a Novamente's carers and teachers to interface with the Novamente. Rather than trying to bring Novamente into our world we could enter its world via virtual reality - strictly both we and the Novamente(s) would enter each other's experience via a shared virtual reality world. So a Novamente would control the behavior of an agent in a simulated world and it's carers/mentors would do likewise. The playpen that you've often talked about would be a simulated world and both Novamente(s) and humans could be in there together. I'd be very keen to collaborate on the design of the simulated world and on the roles/goals that Novamente might be set in such an environment. I haven't got the skills to help with the development of Novamente internal architecture but I think I have something to offer in relationship to the project you are now contemplating. Cheers, Philip --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
Re: [agi] Educating an AI in a simulated world
For a pretty cool subscription-based world of the type you are describing, check out www.there.com --Kevin - Original Message - From: Peter Voss [EMAIL PROTECTED] To: [EMAIL PROTECTED] Sent: Friday, July 11, 2003 2:42 PM Subject: RE: [agi] Educating an AI in a simulated world Hi Ben, For the past few months our project's focus has shifted to exactly that. Now that we have our basic framework and algorithms working, most of our effort goes into defining, setting up, and implementing various tasks in a (mainly) simulated environment. Our AGI engine interacts with the virtual world, learning stuff as it goes along. (It can also interact directly with the real world, or do both simultaneously.) Experience with this type of interactive learning to a large extend now drives our AGI's cognitive development. It is quite difficult coming up with a series of ever harder tasks that are in the correct sequence, are not too redundant, nor rely too much on hard-coded/ instinctual abilities. We've ended up with a set of test/tasks inspired by animal cognition/intelligence studies, Piagetian infant developmental stages, and of course, our own theoretical model of intelligence. Obviously, actually running the tests tells us a lot more about what to change or focus on. We see our project using mainly this approach to drive up levels of AGI. I'd be interested to hear your ideas. Attached is an internal document (cryptically) listing some current and near-term tasks framed in terms of dog-like abilites (call me AIGO). Peter -Original Message- Ben Goertzel ... One of the things I've been thinking about lately is the potential use of our (in development) Novamente AI system to control the behavior of an agent in a simulated world -- say, a very richly and effectively constructed massively multiplayer video game, or even a more futuristic VR-based simulation game I have developed some ideas about such an educational program, but I'd like to hear others' thoughts on this and related topics, if anyone should have any... --- 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] Educating an AI in a simulated world
..except your body is *not* supplied with sensors and actuators in There.. But if its a virtual world, why do you need sensors and actuators?? There is the presented visual display and control keys for moving around and conversing.. - Original Message - From: Ben Goertzel [EMAIL PROTECTED] To: [EMAIL PROTECTED] Sent: Friday, July 11, 2003 11:53 AM Subject: [agi] Educating an AI in a simulated world Hi, One of the things I've been thinking about lately is the potential use of our (in development) Novamente AI system to control the behavior of an agent in a simulated world -- say, a very richly and effectively constructed massively multiplayer video game, or even a more futuristic VR-based simulation game This is not something we're working on in practice right now, but it's something that could happen in the future, and it seems an interesting context in which to explore the experiential learning aspect of Novamente. The reader of the email is assumed to have some familiarity with the Novamente software design, minimally at the level of the overview yakkity-yakk on www.agiri.org Let's assume that the simulation world consists of a simulated 3D physical environment, and that Novamente is given control of an agent that is localized in a particular body within this environment. The body is supplied with sensors and actuators, and · Each sensor results in a stream of perceptual relationships being presented to Novamente, over time. These relationships are primitive perceptual relationship types, for example the output of a camera eye might be represented using a relationship (PixelAt n m c) where n and m are ints representing locations on the camera screen, and c is a list representing a perceived color. · Each actuator is represented by a function taking one or more arguments, e.g. move(v, a), where s is a float indicating speed and a is a list indicating direction. The particular set of sensors and actuators involved is very important for practical purposes, although the general approach described in the document works for essentially any set of sensors and actuators. We are thinking in particular of · Sensors such as: simulated camera eyes, microphones, · Actuators such as: movement devices that can move in a specified direction with a specified speed, sensor control devices (e.g. pointing a camera in a certain direction) We are not concerning ourselves here with the details of robot control - for instance, with the mechanisms of controlling a robot arm. This sort of thing can be handled in Novamente, but, it is not required in gamelike sim worlds, and anyhow we feel it's a less interesting area of focus than higher-level control. Regarding sensory processing, we are willing to make use of existing sense-stream processing tools - for example, if camera-eye input is involved, we are quite willing to use existing vision processing software and feed Novamente its output. We would also like Novamente to have access to the raw output of the camera eye, so that it can carry out subtler perception processing if it judges this appropriate. Next, we assume that there are particular goals one wants the Novamente-controlled agent to achieve in the simulated environment. These goals may be defined abstractly, but they should be definable formally, in terms of an Evaluator software object that can look at the log of Novamente' s behavior in the simulated world over a period of time and assess the extent to which Novamente has fulfilled its goals. While the end goals for Novamente may be extremely sophisticated, we consider it important to define a series of progressively more difficult and complex goals, beginning with very simple ones. The goal series must be defined so that, with each goal Novamente learns to achieve, its internal ontology of learned cognitive procedures is appropriately enlarged. Recall that the Novamente software design does not provide a full cognitive architecture, only a framework and a set of processes within which a cognitive architecture may emerge through experiential learning. The cognitive architecture itself then consists of a dual network (hierarchy/heterarchy) of learned procedures, appropriate for various sorts of activity in various sorts of context. For the cognitive architecture to build up properly, requires the right sort of experience. And so one needs an educational program ... a series of tasks to lead Novamente through... to progressively let it build up the right internal declarative and procedural knowledge for surviving, flourishing and achieving its goals in the environment... OK -- I'll leave off the email here. I have developed some ideas about such an educational program, but I'd like to hear others' thoughts on this and related topics, if anyone should have any... -- Ben G --- To unsubscribe, change your address, or temporarily deactivate your
RE: [agi] Educating an AI in a simulated world
In the case you're talking about, it seems -- the actuators are the control keys -- the sensors are a function that simply returns the presented visual display as an array of pixels This is a very very simple set of sensors and actuators... Some existing games provide richer sensors and actuators than that. For instance, the GameCube controller (or auxiliary steering wheel, etc.) is a better actuator than control keys (as would be a simulation of these things). Some games provide multiple screens that you can scroll between, multiple views, etc. -- chat boxes and meaningful sounds as well as the visual display. Richer sensors and actuators lead a system to have multiple views of the same situation it's embedded in. This may be important, because it may help a system to learn to develop multiple views of situations in a more abstract sense. This is only one example of the potential value of a richer system of sensors and actuators for experiential interactive learning. -- Ben G -Original Message- From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] Behalf Of Kevin Sent: Friday, July 11, 2003 3:09 PM To: [EMAIL PROTECTED] Subject: Re: [agi] Educating an AI in a simulated world ..except your body is *not* supplied with sensors and actuators in There.. But if its a virtual world, why do you need sensors and actuators?? There is the presented visual display and control keys for moving around and conversing.. - Original Message - From: Ben Goertzel [EMAIL PROTECTED] To: [EMAIL PROTECTED] Sent: Friday, July 11, 2003 11:53 AM Subject: [agi] Educating an AI in a simulated world Hi, One of the things I've been thinking about lately is the potential use of our (in development) Novamente AI system to control the behavior of an agent in a simulated world -- say, a very richly and effectively constructed massively multiplayer video game, or even a more futuristic VR-based simulation game This is not something we're working on in practice right now, but it's something that could happen in the future, and it seems an interesting context in which to explore the experiential learning aspect of Novamente. The reader of the email is assumed to have some familiarity with the Novamente software design, minimally at the level of the overview yakkity-yakk on www.agiri.org Let's assume that the simulation world consists of a simulated 3D physical environment, and that Novamente is given control of an agent that is localized in a particular body within this environment. The body is supplied with sensors and actuators, and · Each sensor results in a stream of perceptual relationships being presented to Novamente, over time. These relationships are primitive perceptual relationship types, for example the output of a camera eye might be represented using a relationship (PixelAt n m c) where n and m are ints representing locations on the camera screen, and c is a list representing a perceived color. · Each actuator is represented by a function taking one or more arguments, e.g. move(v, a), where s is a float indicating speed and a is a list indicating direction. The particular set of sensors and actuators involved is very important for practical purposes, although the general approach described in the document works for essentially any set of sensors and actuators. We are thinking in particular of · Sensors such as: simulated camera eyes, microphones, · Actuators such as: movement devices that can move in a specified direction with a specified speed, sensor control devices (e.g. pointing a camera in a certain direction) We are not concerning ourselves here with the details of robot control - for instance, with the mechanisms of controlling a robot arm. This sort of thing can be handled in Novamente, but, it is not required in gamelike sim worlds, and anyhow we feel it's a less interesting area of focus than higher-level control. Regarding sensory processing, we are willing to make use of existing sense-stream processing tools - for example, if camera-eye input is involved, we are quite willing to use existing vision processing software and feed Novamente its output. We would also like Novamente to have access to the raw output of the camera eye, so that it can carry out subtler perception processing if it judges this appropriate. Next, we assume that there are particular goals one wants the Novamente-controlled agent to achieve in the simulated environment. These goals may be defined abstractly, but they should be definable formally, in terms of an Evaluator software object that can look at the log of Novamente' s behavior in the simulated world over a period of time and assess the extent to which Novamente has fulfilled its goals. While the end goals for Novamente may be extremely sophisticated, we consider it important to define