Is a Theory Generation Program possible?

Click For Summary

Discussion Overview

The discussion revolves around the feasibility of creating a theory generation program in physics, particularly focusing on complex problems like quantum gravity. Participants explore the potential for such a program to evolve theories based on experimental results and goals, as well as the idea of an idea generator to inspire insights. The conversation touches on the challenges of programming and the nature of theory generation.

Discussion Character

  • Exploratory
  • Debate/contested
  • Conceptual clarification

Main Points Raised

  • Some participants propose the idea of a theory generator that could evolve theories based on experimental data, while others express skepticism about its feasibility.
  • One participant suggests starting with a "proof-of-principle" using established theories, like the Special Theory of Relativity, before tackling unsolved problems.
  • There is a suggestion that the brain functions as a theory generator, leading to the idea that an AI could potentially replicate this process, although concerns about the timeline for developing such AI are raised.
  • Some participants mention the possibility of a theorem generator as an alternative approach, referencing existing examples.
  • Concerns are raised about the challenges of selecting relevant observations and ensuring that generated hypotheses are meaningful and not merely fitting data without real explanatory power.
  • One participant notes that the problem of quantum gravity may involve an excess of theories and a lack of experimental results, complicating the generation of new theories.
  • Another viewpoint suggests that generating unconventional hypotheses, even if incorrect, could lead to new insights and understanding in the field.

Areas of Agreement / Disagreement

Participants express a mix of skepticism and curiosity regarding the concept of a theory generator. While some agree on the potential benefits of generating new hypotheses, others question the practicality and validity of such an approach. The discussion remains unresolved with multiple competing views on the feasibility and methodology of theory generation.

Contextual Notes

Participants highlight several issues, including the difficulty of determining which observations are relevant, the challenge of ensuring that generated hypotheses are meaningful, and the need to validate any system against known problems and solutions.

Meatbot
Messages
146
Reaction score
1
Since many problems in physics are proving difficult, would it be possible to create a theory generator? It would be set to focus on one problem, like quantum gravity. It would be fed experimental results and goals and then be set free to evolve theories that produced those results. Has this been tried or is it just too difficult a programming task at the moment?

At the very least, how about an idea generator that might spark insight into the problem. It might contain general concepts and mathematical relationships that are combined randomly, like a band name generator.

Perhaps the idea generator can help seed the theory generator.
 
Last edited:
Physics news on Phys.org
Meatbot said:
Since many problems in physics are proving difficult, would it be possible to create a theory generator? It would be set to focus on one problem, like quantum gravity. It would be fed experimental results and goals and then be set free to evolve theories that produced those results. Has this been tried or is it just too difficult a programming task at the moment?

Shouldn't you set out to do a "proof-of-principle" first rather than going ahead and trying to solve an unsolved problem?

Try setting up a simple "theory generator" on, oh, something simpler like the Special Theory of Relativity that has been verified already. If you can't find that it can be done, what hope do you have for solving more difficult and unverified ones? This is how science, and especially physics, is done. A better "technique" should show that it can work in areas that we know already.

And oh, my opinion here is that there is no such thing as a 'theory generator'.

Zz.
 
ZapperZ said:
Shouldn't you set out to do a "proof-of-principle" first rather than going ahead and trying to solve an unsolved problem?
- Definitely.

ZapperZ said:
And oh, my opinion here is that there is no such thing as a 'theory generator'.
- Well, the brain is a theory generator so I would think an AI would be capable of it as well. If we could create AI smarter than we are, that should do the trick - but that will take too long. I don't think theory generation requires intelligence though, so we might be able to build a generator before we are able to create AI.
 
jim mcnamara said:
Maybe he means a theorem generator.

For example:
http://mizar.org/trybulec65/9.pdf

Yeah...something like that. Why not expand that into physics?
 
Meatbot said:
- Well, the brain is a theory generator <snip>

Non-psychotic humans are capable of inductive and deductive reasoning - whether or not they know what those terms mean.

Generating theories means taking relevant observations (knowing whether they are relevant or questionable is pretty much beyond AI algorithms now), and then constructing a general statement which explains all of those observations. Then you repeatedly test with new observations. ZZ and the OP covered that pretty well.

So, we have issues:
1. which observations to keep, which to pitch, and why were they pitched.
It is possible to construct a nifty, interesting, and wrong hypothesis by just tossing and keeping examples until you get a combination where everything fits a bogus hypothesis
-- listen to any politician or Rush Limabaugh for endless examples.

2. how to establish any hypothesis generated really accomplishes any meaningful explaining - another AI problem as well.

3. how to establish the system works correctly on known problems and solutions.
This was mentioned.
 
Meatbot said:
It would be fed experimental results and goals and then be set free to evolve theories that produced those results.
I thought the problem of QG was too many theories and not enough experimental results.
 
Last edited:
jim mcnamara said:
It is possible to construct a nifty, interesting, and wrong hypothesis by just tossing and keeping examples until you get a combination where everything fits a bogus hypothesis
--Yes, and actually that would be a good thing, since it might spark some new insight. The computer would probably come up with some wacky stuff. You'd have to explain why it was wrong and might learn something in the process. But it also might stumble into the truth.
 

Similar threads

  • · Replies 1 ·
Replies
1
Views
4K
  • · Replies 35 ·
2
Replies
35
Views
4K
  • · Replies 36 ·
2
Replies
36
Views
5K
  • · Replies 6 ·
Replies
6
Views
3K
  • · Replies 61 ·
3
Replies
61
Views
9K
Replies
21
Views
5K
  • · Replies 26 ·
Replies
26
Views
6K
Replies
1
Views
3K
  • · Replies 4 ·
Replies
4
Views
2K
  • · Replies 22 ·
Replies
22
Views
2K