# Looking for a good problem for a solution I have coded :)

1. Jan 9, 2016

### tomekd789

I have coded the following abstraction for breeding neural networks with a genetic algorithm: https://github.com/tomekd789/clogann
Now I am looking for a problem to solve using it. It would be ideal if it is non trivial, of practical significance, and matching the tool. So far I've tried:
• Numbers factorization (with a mediocre effect)
• Checking satisfiability of random 3-SAT terms taken from the phase transition area (http://www.princeton...s/gent94sat.pdf) (somehow dissatisfied; did not reach the goal of proving P = NP http://cdn01.codecall.net/public/style_emoticons/default/wink.png [Broken])
• Forex (with positive effects, but not worth the effort).
All hints welcome.

Regards,
Tomasz

Last edited by a moderator: May 7, 2017
2. Jan 10, 2016

### Svein

Solution looking for a problem is often not cost-effective.

3. Jan 13, 2016

### tomekd789

Anything more concrete? ;)
Meanwhile, I've been advised to take a sequence of "coin tosses" generated manually by a human (i.e. no real coins, no computer involved, just from one's head) and breed a network capable of predicting next tosses results. I was astonished: for 1000 "coin tosses" the network was 70% correct.
I.e. we humans are quite predictable, at least in this case. :)
Another surprise was the necessary minimal size of the network: just two input neurons, and three extra, one treated as output, too.

Still looking for other ideas.
Tomasz

4. Jan 13, 2016

### Staff: Mentor

Which ratio do you get if you simply predict "the opposite of the previous coin"? Is the neural net better than that?

5. Jan 13, 2016

### tomekd789

65.2%, then still slightly worse.

6. Jan 13, 2016

### Staff: Mentor

At first glimpse this isn't very significant. What does your algorithm? Can it solve linear systems or can it learn?
Without being an expert I'd try to google 'genetic algebras' and see what kind of problems there are.
Or try some decryption problems. You can start with an easy Cesar code. Just to see how it performs. And then more difficult ones.
(I'd have a bunch of tasks but they all involve solving linear systems with free variables. They are $O(n^3)$ which quickly makes it boring to solve them by hand.)