4
$\begingroup$

Set aside networks, image classification, gradients, and the strength of intelligence for a moment and consider the world before people lit fires.

Fires were started periodically just as they are now, when lightning struck dry deciduous matter. People probably ran for water. Perhaps some smart people learned that fire created warmth at night after a blaze and discovered how to preserve it. They couldn't let it go out. If they did, they'd have to wait through many cold nights until the next forest fire, which could take decades, and then try again.

How did someone invent how to start one without lightning?

The first plow, the first written law, the first bow, the first coin, the first water wheel, the first mechanical clock, the first circuit, the first logo, the first transistor, the first web site.

Now try to imagine it from the other end. There is something important. It has no name. It has no design. It has no method of procurement. If you create one, no one will know what it does or why it is there. But you do. You're its inventor. It began in your brain. You saw it before it existed. It was vision without the involvement of your eyes. You then must bring it into being with your speech or make it with your hands.

This is the rarest, most human, and most precious form of intelligence.

What kind of algorithm can invent? Genetic algorithms have not produced new designs of things that meant nothing to anyone when they appeared but mean something to the computer running the algorithm. Something is missing in the way we perceive intelligence. Maybe it is a kind of negative model, where the empty space represents the thing that needs inventing. How would one create such a model?

Is there something like a virtual die, where the empty space in the universe of utility is highlighted? What kind of algorithm can detect the absence of something without it ever having first existed?

Once that ability is understood, we can then approach the problem of running through a wide array of approaches to create this previously un-invented thing and check each for feasibility, but first we must learn how to artificially envision nameless things that fill previously unimagined niches.

What kind of algorithm can invent?

Addenda in Response to Excellent Comments

This question is dear to me because I've invented things, some for corporations, some for my own laboratory, and some that I failed to push hard enough and someone else invented something very close and developed it first. This last case is interesting, and I've seen it happen many times. It's also common in scientific history, where two people who don't communicate directly simultaneously come up with some scientific or industrial invention.

Reading Alonso Church's interview we find that Alan Turing didn't study under him as most historical accounts state. According to Church, he developed his lambda calculus and Turing developed his machine in parallel and without direct consultation. There is an environmental aspect to invention, as if the world around the inventors are subconsciously searching without a clear objective.

There is an accidental appearance to invention, but everyone I know or read about who invented something was poking around in the area of the invention in their mind, and not just casually. We obsessed over some imaginary search space, hunting for something novel and purposeful.

It is like a rat in a maze that smells cheese but does not know the path. We can't just try all the passageways marking each path to avoid duplicate trials to get to the cheese. We don't know it yet but there is a hole in the ceiling covered with a thin veil. Until we realize there is another level to the maze, we cannot find the cheese. The veil represents the discovery. The passage into the second level where the cheese resides is the novelty.

We could find the hole by accidentally hitting the veil when arbitrarily jumping up and down or with a stray ball when playing a game of toss with another rat, but it would sure be faster if we realized something. We smell cheese but failed to find it. When we doubt our method, we start looking for fissures in the surfaces we haven't yet stepped on. Doubt has something to do with it.

Yes, there is a requirement of some kind of understanding or model of the world that can be altered and tested in imaginary space. This is most obvious in the writings containing the thought experiments of Archemedes (buoyancy principle which led to the relative incompressability of liquids, the first conception of a screw, ...), the thought experiments of Isaac Newton (the two prisms in series, the cannon ball blasted into orbit which led to the entire field of Newtonian physics, ...), and the thought experiments of Turing (the imitation game, computability and the punch tape machine).

$\endgroup$
  • 1
    $\begingroup$ I might be wrong but most inventions I think were accidental, so a genetic algorithm can be the required algo provided it has a helper function which determines if the invention is of use. $\endgroup$ – DuttaA Nov 9 '18 at 7:36
  • 1
    $\begingroup$ @DuttaA: I think the key issue that GAs (and any AI that I know of) does not solve is: ". . . but mean something to the computer running the algorithm." Various kinds of AI can solve problems when interacting with an environment in a trial and error fashion, or can generate progressively better combinatorial solutions. None that I know of have anything resembling the level of internal self-assessment that constitutes "understanding" that would be necessary for an inventor to specifically choose some subset of their creativity and present it (communicate it?) as something new and useful. $\endgroup$ – Neil Slater Nov 9 '18 at 9:21
1
$\begingroup$
  • The Ludi System invented a game called (Yavaleth) that humans play.

See: Evolutionary Game Design (2011)

This can be taken as validation of computational creativity, even if the scope is limited to combinatorial games. Specifically, the system can only invent within a defined framework, but the same can be said of humans. (Most new abstract games invented by humans are evolutions of prior games, and new combinations of previous mechanics.) This can be extended to any field of human creative endeavor, where nearly all innovation occurs within previously defined models, and fundamental, paradigm-changing breakthroughs are the exception.

$\endgroup$
1
$\begingroup$

I think your on the right path in regards to looking at genetic algorithms for the ability to do this. Ken Stanley who was the main guy behind the NEAT neuroevolution algorithm also created a new type of fitness function, called novelty search that tracks output and uses knn to determine the novelty of new networks output, and then uses this as the fitness function. Ken and his team used this to create at a biped that learned to walk without the distance traveled being explicitly evaluated by the fitness function, but rather it just favored networks that behaved in new ways until one came about that took a step. I believe he also used this approach to generate some very cool weapons in a space ship battle video game. In the latter example it was literally inventing new weapons but that is still outside the realm of being a completely new invention seeing as weapons already exist. So really if you want an ai to invent completely new things and not be confined to the environment and task you assign it and rather wish to allow it to determine these on its own you first need to teach it to improve and modify its own code, so you would run neuroevolution to teach it nlp, then you write a new task where this net outputs new task/fitness evaluation/environment code then the evaluation function is running that code, and finally grade the novelty of the code (if it doesnt run it would make sense to give that a bad fitness score). Obviously this is a ten thousand foot view of how this algorithm would operate but i think something along those lines would be as close to what your describing as I can come up with.

$\endgroup$
0
$\begingroup$

In my opinion a large part of the answer has already been given by the comments on the original question by @Duttaa and @Neil Slater. If an algorithm just does random stuff and you let it run infinitely long, it will invent (or discover) everything that is possible in its domain. Think of a program that randomly creates musical pieces. Let it run infinitely long and it will write all musical pieces that are possible. The crucial point is how to assess the utility of the discovery? How can the program know what is a "good song" and thus try to improve it or what is a "bad song" and thus stop continuing to explore it?

Clearly the algorithm would need an internal measurement of the value of the discoveries. We humans also have this internal system. We might either be guided by a reward in form of hormones that are released in the brain at certain situations (this is more or less a hard coded value function), by acquired value functions (as we use in moral context or also in music), or by rational thoughts i.e. "I need to find a material that allows me to build lightweight cars that are still stable".

However, taken all this together invention in my opinion is nothing else but a random search that is guided by a value function.

$\endgroup$
0
$\begingroup$

I once set a genetic algorithm to work "inventing" three-element lens designs. The fitness criteria were a specific magnification with optimal resolution and minimum size. It was set to evolve independent populations. It ran overnight and came up with 12 different designs, all of which turned out to be either patented or old classical designs.

The outcome was pretty much what I expected, but it convinced me that the important part of invention is more the question (the problem to be solved), than the answer (the solution to the problem).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.