Posts by Simon Strandgaard

1) Message boards : News : Project update: 110k programs milestone reached! (Message 658)
Posted 6 Dec 2022 by Simon Strandgaard
Post:
110k, Crazy!

110 is the sum of three consecutive squares, 110 = 5^2 + 6^2 + 7^2.

Thank you to all the miners for making this possible.
2) Questions and Answers : LODA Language : What's this all about? What? How? Why?? (Message 463)
Posted 1 Jul 2022 by Simon Strandgaard
Post:
Martin, your explanation is spot on. I may copy parts of it.

ML algorithms. Currently it's brute force, trying mutating existing programs and see if the candidate program is an entirely new program, or an improvement to an existing program.

In LODA-RUST uses: bloomfilter, pagerank, ngrams, Locality-sensitive_hashing.

This is what the "trigram.csv" with instructions look like
count;word0;word1;word2
20976;lpe;mov;STOP
17672;add;lpe;mov
15456;mov;lpb;sub
14987;START;mov;mov
11889;mov;mov;lpb
11836;lpb;sub;add
11041;lpb;sub;mov
10899;START;mov;add
9100;add;add;add
9094;add;lpb;sub
9062;sub;lpe;mov
8906;START;add;mov
8405;START;mov;lpb
8346;mov;add;lpb
SNIP 4100 rows SNIP


I want to experiment with transformers. I'm not sure how to represent the instructions/registers/values.
3) Questions and Answers : LODA Language : What's this all about? What? How? Why?? (Message 450)
Posted 29 Jun 2022 by Simon Strandgaard
Post:
Proposals for improved texts for the website are much welcome.

I'm not a native english speaker, so explaining what LODA is about, in a easy readable way, is hard for me.

Below is my attempt at describing what LODA is.

---

LODA is a math AI, that takes integer numbers as input, and outputs a corresponding formula.

LODA needs your help training the AI. Please join the BOINC community for this.

Is LODA correct?
No, is the short answer.
The long answer, it depends on the training data. The OEIS database is used as training data.

There is math everywhere, such as crystal structures, how many spheres can be fitted in a cube, how quickly can rabbits multiply.
If LODA find a new formula that does things in a smarter way, that what was previously known, it may lead to new discoveries in the field of math/physics. It may reduce power consumption in computers.
4) Questions and Answers : Feature requests : True offline operation (Message 444)
Posted 27 Jun 2022 by Simon Strandgaard
Post:
My experimental loda-rust can mine in offline mode, and accumulates candidate programs for later verification.

There is a ruby script for doing the verification and this has to be online. Underneath it uses loda-cpp and calls "lode check" and "loda eval" with a steps parameter.

At present, the loda-rust mining flow requires several manual steps, that I want to eliminate.

Help with automation is much welcome.
5) Questions and Answers : LODA Language : What is the research goal of LODA miner (Message 94)
Posted 14 May 2022 by Simon Strandgaard
Post:
Disclaimer: I know nothing about proving math and little about AI.

Only the OEIS terms data are being used as training data for LODA. The OEIS name and formulas are currently being ignored.

If there are few terms (less than 40) registered in OEIS, then the LODA program will output those terms, but may likely diverge after the last term.

If there are many terms (more than 10000) registered in OEIS and the LODA program outputs these terms, then the chance of correctness is higher.

A program satisfies all the many terms in OEIS and compute things in a novel way. Proving correctness of a such program, that is the challenge.

Currently it's being done manually.
https://oeis.org/search?q=loda&language=english&go=Search

Can this be automated?

Approach A. Transformer architecture, that takes LODA programs+terms as input. Given a prompt of 20 terms and see what program is being generated. Is it correct/incorrect and decide on what to do about it. Are there patterns that can be proven.

Approach B. Huge AI trained with OEIS terms + OEIS metadata + documents being referenced. Lots of math training data.




©2024 LODA Language