Some documents on AM and EURISKO(blog.funcall.org)
blog.funcall.org
Some documents on AM and EURISKO
http://blog.funcall.org//lisp/2018/11/03/am-eurisko-lenat-documents/
10 comments
This bit about EURISKO exploiting its own heuristic scoring mechanism is cute:
> One of the first heuristics that EURISKO synthesized (H59) quickly attained nearly the highest Worth possible (999). Quite excitedly, we examined it and could not understand at first what it was doing that was so terrific. We monitored it carefully, and finally realized how it worked: whenever a new conjecture was made with high worth, this rule put its own name down as one of the discoverers! It turned out to be particularly difficult to prevent this generic type of finessing of EURISKO'S evaluation mechanism. Since the rules had full access to EURISKO'S code, they would have access to any safeguards we might try to implement. We finally opted for having a small 'meta-level' of protected code that the rest of the system could not modify.
> One of the first heuristics that EURISKO synthesized (H59) quickly attained nearly the highest Worth possible (999). Quite excitedly, we examined it and could not understand at first what it was doing that was so terrific. We monitored it carefully, and finally realized how it worked: whenever a new conjecture was made with high worth, this rule put its own name down as one of the discoverers! It turned out to be particularly difficult to prevent this generic type of finessing of EURISKO'S evaluation mechanism. Since the rules had full access to EURISKO'S code, they would have access to any safeguards we might try to implement. We finally opted for having a small 'meta-level' of protected code that the rest of the system could not modify.
It's not just cute, it's an open problem in the theory of recursively self-modifying AI design--there's no general solution to an AI hijacking its own reward channel, so far. https://www.lesswrong.com/posts/upLot6eG8cbXdKiFS/reward-fun... is the most recent thing I know of in the area.
What I find most incredible about EURISKO is it did all that on what, by today's standards, is a very small amount of computational power.
Just as outstanding is that, in all these years, in terms of open-ended problem solving, nothing has truly managed to surpass it.
The conjunction of those two things is a bit scary because it suggests to me that current methods are probably not making the most effective use of compute. What happens when that is corrected?
Just as outstanding is that, in all these years, in terms of open-ended problem solving, nothing has truly managed to surpass it.
The conjunction of those two things is a bit scary because it suggests to me that current methods are probably not making the most effective use of compute. What happens when that is corrected?
There is no credible documentation of EURISKO's achievements. No one has every actually seen EURISKO's source code except for Lenat.
Despite huge advances in technology, in computational power, and in the amount of engineers and grad students thrown at these problems, the magical reasoning feats have never been reproduced by anyone else - not even by Lenat's much more expensive project Cyc.
What's more, Lenat did not give a consistent account of the Traveler tournament. One time he claimed "ninety-six ships in Eurisko’s fleet, most of which were slow and clumsy because of their heavy armor", other times he would claim the winning strategy was "astronomical number of small ships like P.T. boats, with powerful weapons but absolutely no defense and no mobility".
The most probable explanation is that there's more myth than truth in the stories about EURISKO.
(on a less serious note, the jargon file entry for "bogosity" references Lenat; see http://itre.cis.upenn.edu/~myl/languagelog/archives/003515.h...)
Despite huge advances in technology, in computational power, and in the amount of engineers and grad students thrown at these problems, the magical reasoning feats have never been reproduced by anyone else - not even by Lenat's much more expensive project Cyc.
What's more, Lenat did not give a consistent account of the Traveler tournament. One time he claimed "ninety-six ships in Eurisko’s fleet, most of which were slow and clumsy because of their heavy armor", other times he would claim the winning strategy was "astronomical number of small ships like P.T. boats, with powerful weapons but absolutely no defense and no mobility".
The most probable explanation is that there's more myth than truth in the stories about EURISKO.
(on a less serious note, the jargon file entry for "bogosity" references Lenat; see http://itre.cis.upenn.edu/~myl/languagelog/archives/003515.h...)
IIRC there were two different Traveler entries in different years, and they changed the rules in between to fix the problems the first win had revealed. After the second win they said they'd cancel the tournament if he entered again.
There are more documents in this archive: http://nosuchlabs.com/pub/lenat-leaks.tar.gz
I hoped some source code would surface.
There is a link to a basic AM clone, "Pythagoras" by Steven L. Tanimoto.
http://web.cecs.pdx.edu/~mperkows/CLASS_ROBOTICS/LISP/tanimo...
It's really basic. When I read the code I felt disappointed, though maybe someone else will feel energized to expand it.
I thought I had seen another recreation, in Bratko's book on Prolog programming for AI, but if it is there then it got dropped in the second edition. Maybe my memory got confused between that book and Tanimoto's.
I thought I had seen another recreation, in Bratko's book on Prolog programming for AI, but if it is there then it got dropped in the second edition. Maybe my memory got confused between that book and Tanimoto's.
I was around and actually following AI at the time. After some early success with expert systems, the field had fallen into a trough of disillusionment. Even chess was still "unsolved" in the sense that computers had barely reached master level and were still several years before they'd become truly world class (let alone dominant). The very possibility of accurate speech or image recognition was being called into question. At that time, in sharp contrast to today, AI was a specialty most ambitious people in CS were running away from.
Thus, it was a bit of a shock when Eurisko won the Traveller "Trillion Credit Squadron" tournament. Traveller was a space-themed RPG, and TCS was a very detailed space-combat outgrowth of that. That a new kind of AI (what we now know as genetic programming was far from the AI mainstream) would excel at such an open-ended task was surprising. That it would do so at such an obviously military-relevant task in what was still the Cold War was a little scary. I for one was a bit relieved when Lenat turned his and his creations' attention to more benign tasks instead of disappearing into the military-industrial complex.
Lenat, whose ideas were initially viewed with skepticism and worse, did quite possibly more than anyone else at the time to keep AI interesting and relevant. If you're working on AI/ML and don't know the story, I highly recommend digging into it.