Problem Solver in a Box Stuff

Gene Michael Stover (gene@acm.org)

created Tuesday, 2022 January 18
updated ????

Problem Solver in a Box = PSIB

Algorithms

algorithm	capabilities	limitations
symbolic	logic, space search, planning, can explain, can extrapolate	uncertainty must be explicit
connectionist	classifier, feature extraction, parallel	can't explain, can it extrapolate?
Bayes	classifier, feature extraction?	can't explain, can it extrapolate?
nearest neighbor	lazy	can it extrapolate?

General requirements & misnomers

common sense

I dislike this term. Seems to me that common sense is really the millions of little details about being a human that most humans acquire by the time they are, say, 5 years old. They do it by being human, which leads me to suspect that it's not common sense. Instead, it's...

fundamental domain knowledge

What we call common sense is the fundamental knowledge about being human, in a human body, with a human mind, interacting with other humans. If you want a robot to act like a human, or you want a PSIB to predict in accurate nuance how humans will react to its output, it'll need “common sense”, but PSIBs for other domains will require fundamental knowledge for the domains in which they work.

Would not hurt for a PSIB to have fundamental knowledge for multiple domains. For example, a PSIB that works on economics might require the fundamental knowledge about being human (a.k.a. common sense).

a body might be necessary to acquire domain knowledge

Body is metaphoric. For a robot butler or a sex bot, it might be anthropomorphic, but for a self-driving car, it's the car. For an economy planner, we might not be able to point to a body; it's too abstract, covers a whole country.
Senses allow the PSIB to acquire information.
Mobility allows the PSIB to affect the world directly (as opposed to affecting it by asking us to do something for it). Mobility might be metaphoric, might have nothing to do with moving the body around, but I don't know a better word. Mobility might be the most efficient way for a mind to test new ideas that it's forming.

consciousness

I define consciousness as the mind's awareness of its position in the world, how it relates to world.
Consciousness might imply a desire for self-preservation, or it might not. I'm not sure.
Consciousness isn't necessary for a PSIB in general, though some might benefit from it depending on the domain they operate on. Consciousness might be necessary for some domains, though I don't know for sure & am not convinced.

learning

When you increase the amount of knowledge or the quality of it, you're learning.
After learning, you're speed at solving problems might increase, the quality of your solutions might improve, or the difficulty of problems you can solve might increase.
You could do this by research: Reading, hearing someone tell you, or by observing others.
You could also learn by combining knowledge you already have. You'll need to test the ideas to increase confidence in them; that's what mobility is about.
Might be possible to implement knowledge base such that information added to some global repository (such as Wikipedia) is automatically new knowledge in the PSIB.

Requirements

easy setup

Put a PSIB on your desk (so to speak)
Possibly some programming to connect it to its body (so to speak).
Ask it a question. Maybe that's simply identifying a domain of expertise.
It'll learn on its own by doing research.
it'll provide new ideas, plans, suggestions.
By continued learning, it'll improve speed, quality of what it says, & difficulty of problems it tackles.
It might identify more problems to tackle.
It might ask you to perform experiements. In other words, some plans & suggestions are intended to help it provide more ideas later.

Hunches

natural language user interface is optional

Don't need natural language (English) if your users are specialists.
There might be domains in which natural language is a poor user interface. (Example: Mathematicians don't communicate their official idea in English. They use equations & specialized notation — often accompanied by natural language.)
If your users are not specialists, are casual, you might want natural language. Or maybe there are even better user interfaces.
The point is that it's not absolutely true that a PSIB must have natural language.

humans solve cerebral problems symbolicly

When a human solves a problem that requires a lot of thinking, they do it by expanding their knowledge, coming up with new knowledge (new perspectives on the problem), & searching the solution space. That's all symbolic (even though 2 of my friends say that they don't do that).

mashup of algorithms

Might combine symbolic, connectionist, maybe some other algorithms.

Use case examples

For these examples, I tried to pick use cases that would be reasonable for a PSIB & that are varied.

Find (design?) a drug that will cure a specific disease

Knowledge must be about the human body, maybe the human mind, chemistry, existing technologies for manufacturing drugs.
For more possibilities, could give it more fine-grained information about chemistry. For example, maybe “protein folding”.
If it knows about how drug manufacturing technologies work & are developed (not just what they do), it could design or suggest new technologies. At this point, we're getting into the area where the PSIB has broad, superhuman knowledge, is an artificial general intelligence.
Requires a lot of data, but on the first level, where it just has direct knowledge of chemistry & can't think up new drug manufacturing technologies, seems feasible if you are willing to wait. Also, seems that type of search for drugs is already being done.

Create outline for an original novel

Because we are asking for an outline (list of characters, their motivations, settings, plots & sub-plots), this doesn't require natural language.
Because engaging stories are built from & about the human condition, this could require fundamental knowledge about being human (a.k.a. “common sense”). So this is difficult.
In theory, assuming previous condition is met, the program could outline a best seller after analyzing many other novels (best sellers & not), but the only way for it to improve its performance in practice will be for it to learn how the novel was reviewed. So this becomes an example of when we need mobility.

Write source code from descriptions that a non-programmer provides.

For a non-programmer to use this PSIB, the PSIB probably must understand natural language. That's a huge can of worms that we will ignore for now. Yes, it's huge. Yes, it is all about artificial general intelligence. And we ignore it for the moment.
Imagining a session with this program raises interesting ideas & questions.

cybertiggyr.com