AI in the 80s? How a Simple Animal Guessing Game Pioneered Machine Learning

It seemed like a neat trick at the time. There was also a crude CRUD database that worked the same way, retaining up to 50 names and phone numbers.

Actually that disk had a lot of really cool programs, now that I think about it. A biorhythm plotter, Woz's "Little Brick Out" Breakout clone, and a few other demos by luminaries like Bruce Tognazzini and Bob Bishop. And of course, ELIZA, the mother of all LLMs... only it was called FREUD for some reason.

Another portion of the article says more explicitly:

  Limitations:
    Maximum number of nodes: 200.
    The structure is stored only in memory (no disk saving).

If I recall, there is a way to significantly reduce the number of matchboxes needed by taking advantage of mirrored board conditions. Somewhere I have a Perl script that generated HTML for all the matchbox covers and possible next states.

And what happens to those who don't remember the past? Ask George Santayana.

The other major issue is that it isn't machine learning—it's a classic example of an expert system, even if there is a bit of a gray area around whether a binary decision tree qualifies as ML.

The worst part, of course, is that it takes less time to slap "GUESS THE ANIMAL" on a stock image of a glass terminal than it does to prompt a diffusion model to generate the same thing with passable fidelity... and it still wouldn't be an accurate depiction of what the program actually did.

It would be a prank--LLM writing about AI, badly.

But the Medium account behind it has only articles like this.

The reason is that the way decision trees, well, make decisions, is the way that expert systems of the '80s made decisions. The first expert system was called DENDRAL (capitals mandatory), derived from the Greek word for "tree", "dendron" (also see: "dendrite"; little tree). DENDRAL and similar systems implement deductive inference procedures that traverse a tree (either bottom-up or top-down, with backward or forward chaining). In traditional expert systems the edges in the tree are production rules of the form IF-THEN-ELSE and the nodes are intermediary goals of what is essentially a proof of a propositional logic formula (the decision tree; it's a boolean formula).

Expert systems' Achilles heel was the so-called "knowledge acquisition bottleneck" which refers, in short, to the difficulty of constructing and maintaining a huge rule-base as needed for an expert system that can do something more interesting than play 20 questions like the animal identification system above. Because of this difficulty, expert systems never quite fulfilled their promises and they suffered from the brittleness that is one of the first thing anyone hears about expert systems today.

And what did people come up with in order to overcome the knowledge acquisition bottleneck? Why, machine learning, of course. The early days of machine learning as a field --remember, we're back in the '80s, a bit before the AI winter hits-- were exactly systems that tried to learn the IF-THEN-ELSE rules for expert systems' rule bases.

Decision tree learners, like ID3 and C4.5 by J. Ross Quinlan, and similar algorithms, that were wildly popular even well after the ImageNet moment of deep neural nets come from that era. Instead of learning only the rules though, decision tree learners, learn an entire decision tree; in short, an entire expert system, from top to bottom.

I think ZX Spectrum was what I saw most often, and there would be a couple of things that just didn’t work in CP/M basic, gwbasic or whatever I was using at the time - I imagine file manipulation was avoided for that reason also.

But I did find this BASIC version for Atari computers, from Antic magazine in April 1986. The article specifies:

"It's vital to save your computer's knowledge base to disk or tape - this knowledge is what makes the game fun. To save, type [CONTROL] [S] and you'll be prompted for a device to save the file to, and a filename. Cassette users should simply type C: at the prompt. Disk drive owners type D: and then the filename. To LOAD a previously saved knowledge base, press [CONTROL] [L] and respond to the prompt as explained just above."

The .BAS file at the end doesn't seem to be a plain-text listing, it's probably a saved Atari file. But viewing it in a hex editor, I see strings in the binary file like "Shall I SAVE this data?"

https://www.atarimagazines.com/v4n12/Animal.html

Rebalancing is not possible not due to code restrictions but due to the nature of the game. How could the program know what new labels to invent for the inner nodes of the rebalanced tree? Unlike integers, animals are not totally ordered.

Consider the book where this Animal appeared: David H Ahl was aiming at a wide audience. The content was previously in Creative Computing Magazine.

It was up to the reader to adapt the BASIC programs to work on their machine.

The robot illustrations are the unforgettable charm of that book. It's on the net.

Speaking as someone who was there at the time: no, it wasn't. Opening a file, reading/writing data to it, and closing it was a straightforward set of operations on the C64 I had. The I/O was slower, of course.