Hacker News Reader

Comment by PaulHoule | Hacker News Reader

When people talk about AI today they are usually talking about machine learning systems based in neural networks.

The kind of symbolic AI described in this book went through several cycles of hype and disappointment to the point where many think it is obsolete. People often do it connect recent breakthroughs in SAT and SMT solvers with this history and for that matter production rules engines are dramatically better than they were in the 1980s but they’ve never made a breakthrough into general purpose use.

View on HN

Maybe if you define AI to be “technology that could plausibly lead to AGI”, but everything in this book is still terrifically relevant for many practical “how to get my computer to solve this semi-open ended search problem efficiently”. Which is far from uncommon.

> “technology that could plausibly lead to AGI”

My AI prof joked, I think, that it was "things that don't work yet" - clearly only a humanlike AI could do OCR... until it started working, etc.

But I actually think your hypothetically proposed definition fits the history even better.

Hi @Paul, I’m a newbie in this field. Are you saying that this branch of AI isn’t relevant when compared to machine learning that’s based on neural networks?

Norvig addresses this at the end of f http://norvig.com/Lisp-retro.html . Basically these old AI techniques are now considered regular programming and modern AI is focused on ML. Both are useful but for different tasks.

That is a great retrospective! Still very relevant 20 years after it was updated.

Symbolic AI never managed to scale to significant and general real world problems. Neither did Machine Learning, not until the advent of fast processors and massive datasets for training which broke the scale-barrier.

Rule-based expert systems (as opposed to SAT, etc) based on Symbolic AI also have the issue that for non-trivial problems, coding the rules themselves usually requires programming expertise in addition to the domain knowledge required to capture the business logic. Things have improved a lot since the 80s, but applications remain fairly niche.

I think ideas from symbolic AI are relevant here and there but they certainly aren't fashionable.

Almost any financial institution has a copy of IBM iLOG in there somewhere implementing policy in terms of production rules.

Some of the most interesting systems today combine ideas from machine learning with ideas from AI search. For instance there are many game playing programs like AlphaGo that use

https://en.wikipedia.org/wiki/Monte_Carlo_tree_search

which runs a large number of games to the end rather than searching the next few moves exhaustively. Using a machine learning model to play the game for the playouts but sampling a large number of moves with A.I. search turns out to be a winning strategy.

The most relevant field this would apply to is AI for games.

Mostly not, no.

Reply on news.ycombinator.com