Skills for AIs and robots

[Peter Knutsen]

Originally Posted by hal

Once GURPS equates a given brain with a given complexity, things will work out ;)

problem is - people doing research into the human brain are beginning to suspect that there is a LOT more going on beneath the surface than originally thought, and that the neuron connections, being three-dimensional, are more complex.

In the end? If the Game Master wants positronic brains that are equal to flesh and blood brains, that's the way the cookie crumbles. If the GM wants to develop their own rules to fill avoid left by the authors - you KNOW I'm all for it! But you also know, that I Loathe the idea of treating robots as characters and that they should be built on character points rather than a set of rules not unlike those found in GURPS ROBOTS. I loathe the idea of building starships on character points, cars on character points, etc. But, that's just me...

Instead of rating brains according to a single axis, GURPS' Complexity, they could be rated according to three axes, to make some kind of "cubic" function: Think vs Move/Spatial vs Social.

Our brains are pretty good at all three, while the primitive electronic "brains" we can make, here at late TL8, are quite poor at Think and suck utterly at both Move/Spatial and Social.

The required "computing power" is then equal to Think x Move/Spatial x Social, meaning that if you want to, you can invent and then build brains optimized for just one thing (as we are trying to do with Think, as of now) with much less effort and expense, compared to a brain optimized for any two things, let alone all three things.

Think mostly maps to GURPS' IQ!, Move/Spatial to DX and a bit to PR (or maybe the absense of disads such as No Depth Perception), and Social isn't easy too easy to express in GURPS, but the least bad version is some kind of Low Empathy, negative (or positive) Charisma and so forth. Tonnes of mental disads to represent social incompetence for primitive AIs, removed sequentially as the brain becomes less socially inept.

[Flyndaran]

Once you start down that road, you end up facing the fact that there is no such thing as intelligence as a single thing. Intelligence is simply a holding place term for "able to understand and solve problems "I" consider important".
Virtually all forms of problem solving is modular evolved for specific types of problems with intrinsic assumptions and short cuts.
These short cuts lead to problems when facing even slightly novel situations. Logic fallacies common to human thinking, sensory illusions, etc.

[hal]

Originally Posted by Peter Knutsen

Instead of rating brains according to a single axis, GURPS' Complexity, they could be rated according to three axes, to make some kind of "cubic" function: Think vs Move/Spatial vs Social.

Our brains are pretty good at all three, while the primitive electronic "brains" we can make, here at late TL8, are quite poor at Think and suck utterly at both Move/Spatial and Social.

The required "computing power" is then equal to Think x Move/Spatial x Social, meaning that if you want to, you can invent and then build brains optimized for just one thing (as we are trying to do with Think, as of now) with much less effort and expense, compared to a brain optimized for any two things, let alone all three things.

Think mostly maps to GURPS' IQ!, Move/Spatial to DX and a bit to PR (or maybe the absense of disads such as No Depth Perception), and Social isn't easy too easy to express in GURPS, but the least bad version is some kind of Low Empathy, negative (or positive) Charisma and so forth. Tonnes of mental disads to represent social incompetence for primitive AIs, removed sequentially as the brain becomes less socially inept.

I like that you're trying to think outside of the box. The question becomes one of...

What is artificial intelligence? GURPS already mapped it out as a function of complexity. GURPS ROBOTS already mapped it out as a function of complexity as well, but added a few extra features to make dexterity become that much more attainable for the design. But why not make it such that low complexity computers, incapable of running high complexity programs, require servo-motors etc, to have a decent dex, but also require a sufficiently well put together software package to gain full maximum use of said servos? That takes care of the two axis you already wanted to build. If you use the rules for complexity of sorts to have certain advantages or disadvantages be treated as a software package - whose complexity limits what can or can not be purchased as advantages/disadvantage - then you will have your triple axis to judge any AI and its accompanying body with it.

[Peter Knutsen]

Originally Posted by hal

I like that you're trying to think outside of the box. The question becomes one of...

Well, that's what my homebrew RPG project, Sagatafl, is about. I steal when I can ("kleptonic principles"), but my ambition level is such that often I find that I want to go places nobody else has gone before, and so I have to design my own rules structures. Think outside the box, as you call it.

The TxM/SpxS thing was just a random idea, though.

I'd be much more inclined to use something like GURPS' Complexity, in Sagatafl's computer rules (tentative plans are for something twice as fine-grained, so each step increases processing power by a factor of 3 or 3.3).

A three-dimensional model rather than one-axis does make for better simulation, and would work a bit better in Sagatafl where Charisma is its own attribute rather than a kind of Talent on IQ-based skills, but in the end, I think it might be better to model various kinds of AI brains in Sagatafl by using Species Packages and Attribute/sub-Attribute Cost Factors. I mean, without having a formal AxBxC system.

[Peter Knutsen]

Originally Posted by hal

What is artificial intelligence? GURPS already mapped it out as a function of complexity. GURPS ROBOTS already mapped it out as a function of complexity as well, but added a few extra features to make dexterity become that much more attainable for the design. But why not make it such that low complexity computers, incapable of running high complexity programs, require servo-motors etc, to have a decent dex, but also require a sufficiently well put together software package to gain full maximum use of said servos? That takes care of the two axis you already wanted to build. If you use the rules for complexity of sorts to have certain advantages or disadvantages be treated as a software package - whose complexity limits what can or can not be purchased as advantages/disadvantage - then you will have your triple axis to judge any AI and its accompanying body with it.

As far as I'm concerned, intelligence is intelligence. Doesn't matter if it sits in a biological or an eletronic brain, or even in an incorporeal spirit type being in a fantasy setting. Although ideally each of these three flavours of intelligence should have a different feel and style. Most of the time.



A quick and dirty system could be that instead of deriving IQ (and, as some may want to, DX) directly from Complexity, Complexity gives you a number of points, which you then use to buy the three items, and this allocation reflects the basic design of the artificial brain or the AI program, and cannot be changed later. It's the "genes" of the brain or program.

Each Complexity can buy you 1 step of Think, 1 step of Move/Spatial or one step of Social.

Each step of Think is then just +1 IQ!, from a default value of perhaps 3 or 4. Whatever works.

Each step of Move/Spatial can be +1 DX, but the first several steps should probably be used to buy off disads such as Klutz and No Depth Perception instead, ones that all robotic AIs should have by default. Maybe Colour Blind too.

Each step of Social buys off 5 CP worth of mental disad, such as No Empathy, Clueless and so forth, and when they're all gone, buys 5 CP of a narrow "Social Skill" Talent. Or just Charisma. I'm not sure what works best.

[Flyndaran]

Originally Posted by Peter Knutsen

As far as I'm concerned, intelligence is intelligence. Doesn't matter if it sits in a biological or an eletronic brain, or even in an incorporeal spirit type being in a fantasy setting. Although ideally each of these three flavours of intelligence should have a different feel and style. Most of the time.
....

But all purely mental aspects are the direct results of physical objects. They can't be separated except philosophically. If we're speaking of reality, of course which is usually where I go when discussing AIs and robots.

[Peter Knutsen]

Originally Posted by Anthony

The problem with complexity limits on skills is that an AI with the same complexity of the human brain and the same general performance should have the same skill limits, and GURPS doesn't put any limit on human skill levels (this is actually a problem with AI IQ, as well; human brain complexity varies by well under 1 point).

The problem here is that GURPS has skill costs "plateau out" at 4 CP/lvl. That may be very reasonable given the diminishing effects of a 3d6-based skill scale, where skill 24 isn't really that much better than skill 21 (in fact it is so much better - roughly - that the 12 CP cost difference is not unreasonable).

But it does man that there's no stopping effect.

There's no "force", acting on the character or on the player playing the character, or on the in-world institution that supports the character's improvement process (e.g. some kind of school, academy or university, or a guild system), to encourage eventually stopping, to say "OK, this is enough, now it's time to either go out and use this skill that I've learned (on an adventure), or else to put it aside and focus on learning another skill".

(Note also the bizarre rule in GURPS, which says that learning-by-doing is only four times less efficient than being taught actively full-time by a fully qualified teacher and being his sole student. Out here in the real world, where I live, the correct number would be closer to 20 than to 4. It might be 30 or 40!)

Many other systems have skill costs that continue to increase forever.

Ars Magica uses triangular costs. 2nd skill level costs twice as much as 1st, third skill level costs three times as much as first, 4th costs 4 times as much as first. And so forth. It never stops.

Sagatafl has each skill level cost 150% of the previous, up until something called the Plateau Value is reached, and after that, usually each skill level costs twice as much as the previous. Initial Learning Speed, as well as what skill level the PV occurs at, is determined by the character's Aptitude for the skill, which is derived from the weighted average of relevant Attributes and sub-Attributes.

This creates a distinct "stopping effect" in Sagatafl. If the skill isn't really important to you, you stop at PV (or even at PV-1, if it a tertiary skill). If it is a bit more important than that, you keep going to PV+1 and then you stop. Or even PV+2 or PV+3 if it is very important (I'm sure veteran ninja have Stealth at PV+3. More than anything, Stealth is what they do). But there's increasing pressure, and that pressure is felt both by the character and by the player playing the character (and by any institution supporting the character's learning process, e.g. his ninja clan, who is increasingly thinking that he ought to leave the training track and go out and do missions or the clan).

There's a similar albeit weaker effect in Ars Magica. But note that both systems use a skill scale that's different from GURPS 3d6-based one. Both have skills starting at zero (although Ars Magica's don't truly start at zero, in that effective skill equals Attribute + Training) and have a much weaker diminishing returns effect when increasing high skill levels (all opposed rolls use something similar to comparative margin-of-success, so more skill is always better there).

Also, the attributes, which in both systems are inborn (they can't be trained; their adult value having become pre-determined no later than early childhood), play a role in this discouragement process.

An Ars Magica character with Intelligence +5 and Magic Theory skill 7 has an effective skill of 12. That's very good. But training-wise he has spent a lot of time getting his skill to 7, so he's unlikely to improve that further. He (and his player) would do much better spending time on some other kind of self-improvement (or, you know, going out and adventuring - Ars Magica characters actually do that, some of the time. They don't all sit in their tower labs all the time). Contrast a character with a more mundane intellect, Intelligence +1, but Magic Theory skill 8. He has more training than the genius character (an expensive 8 vs a slightly less expensive 7), but his effective skill is much lower (9 vs 12), and it is unreasonable to except that he'll ever exceed the genius in effective skill (he's need to train his skill from 8 to 12!!).

Same in Sagatafl, albeit with genetics playing a slightly different role. If you have a high Aptitude for a given skill, or more likely for a given category of skills (since categories of skills tend to be based on fairly similar weightings of Attribute values, eg.. Science skills or Knowledge skills, or Social skills, or Melee Weapon skills), then your initial Learning Speed will be higher (easily 2-5 times faster than an average person, and in some cases much more than that), but you will also reach the Plateau Value much later, in terms of skill level, so that you'll have "easy going" for more skill levels.

Again, this creates a "soft cap" effect. If a character has below-average Dexterity and average Agility and Spatial Intelligence, then he's not going to become very good with melee weapons or unarmed combat. Nor ranged weapons, although he can compensate a bit there if he has high Perception.

Nothing makes it absolutely impossible for him to keep training, and training, and training, and after many, many decades, he can become fairly good.

But it is unlikely in all sorts of ways. Very unlikely.

In terms of in-world demographics. In terms of in-world social structures (who will provide him with a place to sleep, food to eat, clothe to wear, while he wastes all those decades on non-productive endavours for which he is obviously not suited?). In terms of the player decision making process during the character creation process. In terms of in-character decisions made after game start.

Same goes for non-Human brains, including animal brains, computer hardware or software brains, and spirit being minds. Those will have physical or mental Attributes or sub-Attributes, from which skill Aptitudes are derived, and that creates the same "soft cap" effect, but not necessarily at the same "skill values" as for Humans.

For instance, primtive (as in GURPS TL9'ish) AIs may have a very low Charisma Attribute, and a very low Interpersonal Intelligence sub-Attribute, and also a low Perception (but not very low - although Colour Vision could be very low). One consequence of that is that Aptitude for many (probably all) social skills will be low, leading to low Learning Speed and the early arrival of the Plateau Effect. So they're not going to "be good with people".

Some beings, including some Humans, also have a post-PV multiplier higher than 2. Having one of 2.1 is not too unusual, as exemplified by highly intelligent children who cruised effortlessly through primary and secondary school and thus never learned to apply themselves. These tend strongly to quit, when the going gets hard (dropping out of college, et cetera). A higher value 2.2 or 2.3, would represent a notable disorder, probably a distinct neuro-structural problem (so something you're born with, rather than something that happened to you), but is still not outside the scope of Human variety

The effect of a raised pPVm is that the "stopping" effect is stronger, and gets stronger still the more you exceed the PV, for each skill. In a way, a highly intelligent character with a pPVm of 2.1 is "on track" to becoming a generalist, being very likely to eventually develop a large number of intellectual skills to a modest level, but never becoming truly good at any one of them.

One could get some interesting world demographic effects, if the pPVm for very primtive AI brains was 2.4 or even 2.5, especially if combined with low or even very low values in certain Attributes and sub-Attributes, thereby very strongly "encouraging" AI brains to stick to learning the kinds of skills they're good at learning, and ignoring those for which they are poorly suited. And also, very much so, stopping at PV or if not there than at PV+1.

[Peter Knutsen]

Originally Posted by Peter Knutsen

One could get some interesting world demographic effects, if the pPVm for very primtive AI brains was 2.4 or even 2.5, especially if combined with low or even very low values in certain Attributes and sub-Attributes, thereby very strongly "encouraging" AI brains to stick to learning the kinds of skills they're good at learning, and ignoring those for which they are poorly suited. And also, very much so, stopping at PV or if not there than at PV+1.

Another possible effect is reverse Natural Skill.

Natural Skill, in Sagatafl, represents having grown up in a culture that very strongly emphasizes one particular skill (or sometimes two). As such it is a Background trait rather than a genetic trait. The "case example" is horse nomads, such as Mongols and Sarmatiams, for whom Riding:Horse is a Natural Skill (or arguably both Riding and Animal Handling: Horse).

Also, according to some claims, iron age and medieval warrior nobility received intensive weapons training from early childhood. That could be Natural Skill too.

The effect of Natural Skill is to delay the Plateau Effect by one skill level.

Here, I'm pondering the reverse, though. We Humans have childhoods. We interact a lot with people. We observe adults a lot, and especially early on those adults aren't particularly aware that they're being observed (because, you know, infants and small children don't actually have brains). We observe other children, out on the street, or in kindergarten.

We pick up a lot from that. We also interact a lot with people, children, younge children (infants too) and older children, adults, we play a lot, we spend a lot of time in enviroments that are safe, where consequences are very trivial.

AIs never had that. That could give them a disadvantage-like trait that causes the Plateau Effect to arrive one skill level earlier, for many skills, including most or even all social skills. (More advanced AIs, of course, will have had simulated childhoods, or may have undergone extensive "learning game processes" with computer scientist teachers, and at higher TLs still, superintelligent AIs will have so high Aptitudes that it more than compensates for the reverse Natural Skill effect.)

It's all a question about how much character change matters, whether small or great importance is ascribed to how fast and how much characters change.

It matters a lot in Ars Magica (one of the typical goals of an Ars Magica session is to get hold of a high-Quality book, so that you can spend the next 3 months reading it, and in the process change by improving one of your skills really quickly), and it matters a lot in Sagatafl too, whereas it seems to be - at best - a secondary thing in GURPS. Which then in turn has some consequences for world simulation.

[Langy]

Originally Posted by Peter Knutsen

(Note also the bizarre rule in GURPS, which says that learning-by-doing is only four times less efficient than being taught actively full-time by a fully qualified teacher and being his sole student. Out here in the real world, where I live, the correct number would be closer to 20 than to 4. It might be 30 or 40!)

I doubt it. I've had courses where I had to teach myself everything - either because the course was designed that way or because I didn't feel any need to show up to class - and in neither case did it feel like I'd have been able to learn everything 20 times faster if I had a full-time tutor rather than working on it on my own. GURPS's default of a 2:1 ratio (not 4:1 - that's on-the-job training) seems fine to me.

I also am a personal tutor, and the big difference between teaching as a personal tutor and people learning on their own is that people learning on their own largely don't do the work or don't realize that they are lacking in some basic skills that they'll need to learn prior to tackling the harder stuff. Other people just can't learn by reading and instead can only learn via direct vocal coaching.

[Flyndaran]

I find that the things I can get alone, I get relatively quickly. But there are many things that I simply cannot understand without external help.