Improving THORN: ways and means

178 views
Skip to first unread message

dr.mt...@gmail.com

unread,
Jul 9, 2025, 10:58:32 AMJul 9
to Shen
Well, I'm planning my pilgrimage to Glastonbury and the summer hols.
But before I do, a word on THORN.  

THORN has it's limits, but one limit it does not have is a large program.  
At 580 loc, THORN is very lightweight.  It's fairly decent on FOL problems 
without equality; so-so on purely algebraic problems and quite bad on
theories that mix the two because of combinatorial problems.  Can it be bettered?

Roughly you can split improvements into 2x2 classes.

1.  Improvements that involved changing things inside THORN.
2.  Improvements that involve putting THORN within something outside THORN.

And this second division.

A.  Improvements that improve the compile time behaviour of THORN.
B.  Improvements that improve the run time behaviour of THORN.

1A. Includes all forms of equational unification.  2A. includes planners
and heuristic reasoners of all kinds.  1B. would include compilation
for parallel processing and 2B. perhaps combining THORN with 1B.
subsumption testing and maybe JIT compilation.  Personally I think 
2A. is likely to give the best fast returns.

Mark

Woo

unread,
Jul 22, 2025, 9:16:11 AMJul 22
to Shen
Can THORN prove Perelman's theory about the shape of the universe and refused accept $1 million prize?

dr.mt...@gmail.com

unread,
Jul 22, 2025, 2:26:23 PMJul 22
to Shen
Probably not :).  THORN can prove some humanly unsolvable problems
like theorems from equivalential calculus that would defy any human
being to solve.   The leading example takes more than 4 million inferences
to solve and produces a proof of > 500 lines in seconds.  It can also find
proofs in Hilbert systems which I (not necessarily everybody) cannot find.
But in general THORN cannot do things that human mathematicians cannot.

However there is an experiment which I'm writing up to interface THORN to ChatGPT
which has yielded startling results and may change this picture.  

Mark

dr.mt...@gmail.com

unread,
Aug 11, 2025, 3:08:32 PMAug 11
to Shen
Said experiment is now up and readable.  In general, improvements to THORN
come from embedding it in a planner outside of it rather than tweaking the code
inside of it.  You can't really raise the bar on THORN itself very much.  But an
good planner can be awesomely effective.

M.

Woo

unread,
Aug 13, 2025, 7:42:23 AMAug 13
to Shen

words fali me
Reply all
Reply to author
Forward
0 new messages