As far as this discussion is concerned, "play" is an activity a system carries for its own sake, rather than as means to other ends. As a by-product, play always serves as an exercise for the relevant skills, as well as provides certain information about the environment (so it is indeed rewarded by evolution), but as soon as the system carries out the activity with those goals in mind, it is not playing anymore --- "play" must be "for fun", not "for money", "for career", etc.
Even though the advantages of playing can be justified, there is also an obvious cost --- the time/energy/resources used in the activity. That is one reason why few people can justify their research if the goal is to design a system that can play but do nothing else. Also, an activity is called "play" only when it is irrelevant to the primary goal of the system, so Deep Blue is not "playing" in this sense when it plays chess. For AGI, to be able to "play", in the above sense, is not only necessary, but also possible, even inevitable. For a system with insufficient knowledge and resources, goal derivation always leads to "alienation", in the sense that what start as means become ends. If I have a goal A, and I belief it can be achieved by achieving B first, I'll take B as my goal, and begin to get internal reward from progress towards it. In the long run, B may turn out to be irrelevant and even opposite to A, though I have no way to completely and absolutely rule out this possibility at the beginning. The same is true for an AI system --- as far as the goal derivation process is based on insufficient knowledge and resources, eventually the system will have many derived goals that do not really serve the initial or original goals from which they are derived. When pursuing these derived goals, the system is, more or less, playing, because it is rewarded (or gets pleasure) from these activities themselves, rather than using them to achieving other goals. The goal derivation mechanism in NARS already works in this way. See my publications (http://nars.wang.googlepages.com/) for more details. Pei On 7/6/07, Bob Mottram <[EMAIL PROTECTED]> wrote:
I think the purpose of play is that it allows the system to search the space of possible actions in a broad yet shallow way, and characterize the landscape under various fitness criteria. So at a later time when some more serious task needs to be undertaken the system can quickly jump to an area (or areas) of the space which it knows is likely to be appropriate. There are also reward systems associated with this kind of search, such that enjoyment is gained by continuing to characterize the space. This reward system seems to be particularly active in humans, who are always discontent and seeking to expand their envelope though leisure activities or knowledge/career advancement. As far as I know there aren't any AI systems which "play" in a proper sense. I've seen robots which appeared to be playing, but this was usually just anthropomorphisation of a rather incompetent system struggling to perform on a single narrow task. On 06/07/07, a <[EMAIL PROTECTED]> wrote: > >a> Sure, I can write a program to differentiate between a square and a circle, > >a> but it is not AGI. I need the program to automatically train and > >a> recognize different shapes. > > > >This is the most important question you have to ponder before > >doing anything specific (and useless!). > >Even if you implement something that can "automatically train itself" > >to do this particular thing, would it scale to do anything? Would it > >teach you something useful about hypothetical way to implement an AGI? > > > Harry Foundalis' thesis is to specific. It does not look like AGI. It only classifies. It does not manipulate. > > I just thought of a way to make my program train itself. It learns by itself by playing. Playing is exploring. Playing is a product of evolution. Playing lets you try "risky" things in order to learn. Playing is learning by trial and error. That's the perfect thing my program needs. Play is driven by a psychological addiction. But coding addiction to every subsystem in the program is too holistic. We need specialized non-emotional subsystems in order to speed it up. Emotion is aggravating to AGI because there is no need for emotion for AGI. But addiction is emotion. Addition is a motive. > > Initially, we need the program to do some random things such as randomly playing. If it does a specific thing, it gets addictive "chemicals". Then, it is addicted to do that specific thing. For example, it will get addicted to solve tests if it gets addictive "chemicals" after it passed the test. > > > I believe that passing an IQ test requires AGI so my program will have AGI if it scores high on the test. > > > > > > > > ____________________________________________________________________________________ > Get the Yahoo! toolbar and be alerted to new email wherever you're surfing. > http://new.toolbar.yahoo.com/toolbar/features/mail/index.php > > ----- > 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/?&; > ----- 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/?&;
----- 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/?member_id=231415&id_secret=13261461-57d2c2
