I like the thoughts here. My hunch is that the human ability to learn new activities is based on conceiving all of them as goal-seeking journeys, in which we have try to find the way.to our goals, using a set of basic paths and series of basic steps, [literally steps, if you're walking somewhere], which can be, and usually are, added to later on. We learn every activity (that I can immediately think of), in this way, by first acquiring a very rough idea of the kind of journey that has to be made. When we actually perform the activity, we normally have only a vague plan, and simply head off in the general direction of the goal, see where we've got to after a while, and then change direction accordingly, and sometimes the goal, and sometimes the activity. We do NOT follow a precise program, or set of instructions. We make it up as we go along, stringing together many routines as we do so. This is true of how we learn to, and do, converse, write essays or reports, think about problems, have sex, eat meals, etc. etc. We are highly improvisational and adaptive.
[The language of goals, finding the way, journeys, steps, etc. - way to go, no way, what is the way to...etc - permeates our whole culture and our descriptions of all activities]. Each movement or move within an activity can also be, and often is, conceived as a mini-journey - in which case it doesn't have to be copied exactly. All this depends, it seems to me, on being able to conceive of "goals", "move", "step" in very, very general and abstract ways - so that you have no problem recognizing that in one activity a step is literally a step, in another a kick, in another a punch, in another crawling etc. Infants clearly have this ability very early - our capacity to use the words "go to" and "move" very early on exemplify this ability, for they are enormously general concepts that can be instantiated by many different forms of movement, and the infant, I think, already understands that. It also depends on the capacity to map an activity - to have some kind of map of the journey from start to goal. How quickly you can program a robot to acquire this capacity to generalize - how many stages are involved - I have no immediate idea. Obviously animals don't have the same capacity to learn new activities and generalize that humans do. But they may well have mapping capacities very early in evolution. If bees can learn how to get from the hive to the honey via some bees' dance, (although according to a new book, Animal Architects, it's a complex language), then surely we can similarly "brief" robots - even if it's only at first to take fairly simple journeys around fairly simple environments. I'd be interested to know, JSH, why your early teleological AI efforts failed. ----- Original Message ----- From: Bob Mottram To: [email protected] Sent: Tuesday, April 24, 2007 5:15 PM Subject: Re: [agi] Torboto - the Torture Robot On 24/04/07, J. Storrs Hall, PhD. <[EMAIL PROTECTED]> wrote: Mining plus matching, analogy, and interpolation/extrapolation. The key to making it work is to form the abstractions that allow the robot/AI to interpret the actions as "grasp broom; lower until it touches floor" instead of "move hand to (x,y,z); close hand; lower hand 6.3 cm". The level at which the data mining takes place is an important issue. Some mining might be done close to the surface - that is, close to the level of direct sensing. For example when aligning with a waste basket simple visual line detection might be sufficient. Other things may only make sense at a deeper level, such as creating a 3D grid map and then being able to discover relationships between the map and the robots current state. A really serious step toward AGI would be to program a robot so that it could watch a human doing a task and then do the task. Back in the day, as a grad student in AI in the 70's, I worked on a system to do act interpretation, i.e. "watch" (be fed a sequence of predicate expressions describing) a series of actions by someone and try to figure out what their goal was and the teleological structure of their activity. The inverse of planning, as it were. The system was frankly junk and many of my notions of the right way to build an AI are informed by its shortcomings -- but we were working on (one of) the right problem(s). To be capable of doing this the robot would need a mirror system, mapping its own actions to the typical actions of a human then bootstrapping its own internal planning system by superimposing it onto other animate entities. This is quite a sophisticated and powerful skill, which allows the system to develop a sense of self versus others which is useful when living in a social group. ------------------------------------------------------------------------------ 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/?&; ------------------------------------------------------------------------------ No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.463 / Virus Database: 269.5.10/774 - Release Date: 23/04/2007 17:26 ----- 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&user_secret=fabd7936
