Artificial Intelligence
<blockquote><i>The more equally attractive two alternatives seem, the harder it can be to choose between them -- no matter that, to the same degree, the choice can only matter less.</i>
— Fredkin's Paradox (as reported by Marvin Minsky)
<hr />
At the moment, the plan is to use a modified neural net as the basis of the !AI in FærieMUD. Stories are the basis of the system, which is designed to identify stories which are similar to stories within the !AI's "memory." The !AI also produces an emotion object?, which governs the action which will be taken by the character or entity being controlled by the !AI.
Perhaps a better term for the !AIs in FærieMUD are "artificial souls" because they model not a logical, thinking machine, but rather a moral, reacting machine. The goal is not to produce intellectual feats of intelligence, but rather to produce believably consistent (and inconsistent) behavior over a period of time.
Thus an artificial soul has to have "taste" as defined in Minsky's <cite>The Society of Mind</cite>: the ability to choose easily and quickly between equally attractive alternatives.
Technical Details
The primary difference between the FærieMUD !AI and other neural nets is in the complexity of the signals passed between the neurons? of the net. For this reason, the FærieMUD neural nets may be similar to the quantum holographic nets being propounded at this site. But we're not sure. Either they're not explaining everything or we're not understanding everything.
The more complex signals being passed by neurons in the FaerieMUD !AI are always in the form of elements? of the periodic table. They react with the neurons (which are also stored as elements) to produce emotion objects, which in turn produce both the learning law and the signal which is passed to the synapses? of the next neuron.
— Scotus - 03 Jun 2000
<hr /> *A note from Stillflame : The site mentioned above has recently reformatted itself, as well as increased the extent to which they provided information. My understanding of it currently says that their uniqueness is in the data storage and retrieval of any given neuron, rather than the complexity of the information being transfered between them. Their data storage is modeled after holographic data storage, a concept which allows multitudes of information to be stored in the same set of numbers, and, through its weirdities, allows for the bulk of the learning to occur as the data is being stored, rather than mutating it after the fact. Their methods use magicks from schools of math i have no experience with, and so cannot comment further without serious risk to both my mind and the minds of the readers.
— Stillflame - 02 Apr 2001
<hr /> *A reply from Scotus : You are certainly right that And Corporation does store more in each neuron than just whether it is on or off. It was also my understanding that the connections between neurons pass more than just I-stimulate-you/I-do-not-stimulate (on/off) messages. I don't see how else they could have meaningful state that is more complex than on or off (along with the state of their rules).
I'm not sure what stillflame means by "increased the extent to which they provided information," but I'm finding less there than I once did. They used to have a bunch of material quite similar to the PC AI article where I learned of them.
I think a better definition of "holographic memory" would be "memory which stores data over a wide range of memory loci" (or addresses). In this sense And Corp. is certainly working on holographic systems, as is any other neural net (which by definition spreads data out over neurons).
FaerieMUD's neural nets have the ability to store more than one story in the same set of neurons. What I am suggesting about And Corporation is that they may be doing a similar thing. (The article in PC AI and their web site hint at this, but do not answer the question unambiguously.)
<hr />
I'm reading a book called Catching Ourselves in the Act, by Horst Hendriks-Jansen, which makes a strong case for modelling !AI by setting up structures which exhibit what he calls "Interactive Emergence". I'm fairly certain this is exactly what connection networks with periodic nodes (or more accurately, Pattern object nodes) will accomplish. More as I read more.
— Ged The Greys Hain - 26 Sep 2001
<hr />
Including of swarms and stories as discussed during Oct. - Nov. 2001
Just to make sure this info isn't lost to the sands of time, i'm writing it down in whatever form i can remember it, if only to serve as a placeholder for the information which actually explains all this stuff.
A swarm intelligence is just a bunch of little guys flying around in some environment. for our purposes, that environment will be a set of stories - those stories which the intelligence feels are part of its current identity/role. it will model its behavior after the decisions the swarm make within that environment.
At some point, it is possible to change the environment within which the swarm operates - when the intelligence perceives conditions such that it sees itself as belonging to another role which has a different set of stories which govern its behavior. the switch is then made, completely changing the behaviors of the entity.
— Stillflame - 04 Dec 2001
— --
Here are some links to AI-related research papers, software, etc:
- Lund University Cognitive Science - A treasure-trove of ideas on cognition. Some highlights:
- Balkenius/ Christian Balkenius] - He has some amazing papers on the effects of emotion, attention, and context on the cognitive process.
- Winter/ Simon Winter] - His paper <cite>Anticipation and Violin Strings</cite> is a must-read.
— --
As of April, 2002
Two types of artificial intelligences will be needed/used.
the first is for general characters in the world, and will be based on the swarm idea. a character will have the stories ze knows, these will not be stored for each character necessarily, but can be easily determined from the character's race, cuture, gender, job, &c. some of those stories (determined by the locations of the individual members of that character's "swarm" within the space of all stories) will be interpreted in the context of the current situation to give a response for that character. Fx: if the character knows and loves a story of a craftsman who works very hard every day, ignoring the emotional needs of his family and friends, and needs to make the decision of whether or not to go drinking with someone who invites hir out, will say no.
The positioning of the various members of a character's "swarm" is determined by the character's perception of the situation ze is in. while at work, a character will have one set of stories to look at, but once the neighboring village attacks hir town, a different set of stories will be considered. these sets of stories will sometimes be refered to as "roles", in that while considering a certain set of stories, a character will tend to act like a <role-name>. examples are blacksmith, mother, villager, high mage, &c.
The second to be developed later, as it is less necessary. logging in as a normal user, this AI will be COMPLETELY indistinguishable from a player character. passing the turing test in the context of faeriemud will be one requirement of this bot, as well as having some sort of goal oriented behavior guidelines. no discussion of how this will work has been done.
<i>See also</i> Phaedrus Artificial Intelligence.
— Stillflame - 13 Apr 2002
— --
<blockquote><i>"Online gamers are discovering the joys of real-time conversation already -- with other gamers -- but the game play isn't <b>determined by</b> those words and the two rarely <b>intertwine</b>."</i>
— "Discovering Games #18," Scott McCloud
McCloud? gives the following example from an online fantasy game:
Scout38: Daredevil is out on DVD.
TrollBoy: Does it still suck?
PalaceGuard: Halt! Who goes there?
FaerieMUD may risk a related problem because NPCs can remember conversations with PCs:
PalaceGuard: Do you think we need to watch the cliff? It would be hard for anyone to scale it and attack the castle.
Scout38: Rumor has it a we have a brave thief in the area. It would take a real daredevil to do it, but methinks this rogue may be up to it.
PalaceGuard: I heard the Daredevil DVD sucks.
Hopefully FaerieMUD PCs will learn to stay in character, cutting down on this problem. And we can also hope that the algorithm for determining what the palace guard remembers will not include too much of this kind of thing.
— Scotus - 30 Jul 2003
Good point.
Another factor that may help: The concept 'DVD' will not be in the ontology, so it may be possible to cause NPCs not to learn things that include unknown concepts.
Of course, there's always the risk (if we do AI right, anyway) that a PC will teach an NPC what a 'DVD' is...
— Ged The Greys Hain - 31 Jul 2003
— --
Iconic, Indexic, Symbolic
There are three main levels of inteligent abstraction of which a being is capable.
An Iconic reference is an association between two things based on similarity - a wax banana looks like a banana, that horse looks alot like the one that was here yesterday. The distinction here of something not actually being the same thing is maintained, but a clear relationship between the two is established. This is very easy to accomplish programmatically too.
An Indexic reference is built out of iconic references, and represents a spatial/temporal association between two things - push the red button and food comes out, "George Washington" is always under pictures of a particular man. This is also called learning something "by rote". This works by having first made a number of iconic references between similar situations, enough so the differences can be stripped away and the important remaining details can be made into an association.
A Symbolic reference is built out of indexic and iconic references, and is in fact a relationship amount other references. This is commonly understood as true understanding of something, as it does not rely on the statistical occurances associated with it ("dog" still means dog, no matter how often the word is used when no dogs are present). Learning symbolic references is much harder to do for most creatures, as it involves both learning and forgetting, and must also be done on entire logically closed sets of symbols at one time. The actual processes by which this is done are not entirely understood at this time.
— Stillflame - 08 Aug 2003
