4
$\begingroup$

Is randomness (either true randomness or simulated randomness) necessary for AI? If true, does it mean "intelligence comes from randomness"? If not, can a robot lacking the ability to generate random numbers be called an artificial general intelligence?

$\endgroup$
  • $\begingroup$ What do you mean by "necessary for AI"? I know it may seem obvious, but I think you should be more explicit and clear. $\endgroup$ – nbro Sep 23 at 12:33
  • $\begingroup$ This is a good core question, in that the statistical approach to AI, which allows machine learning, has eclipsed GOFAI, but it might be useful to elaborate a little in the question. Pseudo-random number generation may not be entirely sufficient, in that, given sufficient time/iterations, it would result in unintended bias. $\endgroup$ – DukeZhou Sep 23 at 19:30
  • $\begingroup$ Even if the claim - randomness being necessary for AI- is true, how does the question does it mean intelligence comes from randomness follow? $\endgroup$ – naive Sep 25 at 9:48
  • $\begingroup$ Being able to generate random numbers does not imply intelligence, not even close to it. Intelligence might follow from the why of random number generation. That why is generally coded in the models for AI such as in neural networks by a human - and that is for weights and bias initialization which in turn is essential for solving the optimization problem using gradient descent. $\endgroup$ – naive Sep 25 at 9:53
2
$\begingroup$

Is randomness (either true randomness or simulated randomness) necessary for AI

It depends on how you define Artificial Intelligence. If you regard it strictly as an intentionally created construct which demonstrates utility, then no. (For instance, Nimatron, potentially the first functioning AI, beat most human competitors at NIM. But Nimatron was classical AI, entirely rules based with no learning.) That said:

  • Randomness has proved a useful component in machine learning, and any feasible AGI would likely require ML.

Given sufficient computing power, aka time and space, it would absolutely be possibly to brute force anything, including AGI, but the resulting algorithm would be "brittle", unable to "compute" anything not previously defined. A learning algorithm, presented a problem outside of its domain of knowledge may initially degrade in performance, but it can learn from those outcomes, and gradually improve performance.

IBM brute forced Chess with Deep Blue, but Chess is a strictly narrow problem that turned out not to require general intelligence. AGI requires human level performance in all tasks engaged in by humans, which, even if they could be broken down to a set of individual narrow problems, it's an ever expanding set of problems.

Does it mean "intelligence comes from randomness"?

Not if the definition of intelligence is rooted in utility because deterministic processes can demonstrate utility.

  • In statistical AI, the intelligence arises from the analysis of random search or the fitness of the genetic algorithm, not the randomness per se.

In other words, if you have the randomness without the analysis, every decision is an unqualified guess.

My sense is that it is free will that would arise from randomness—effects unrelated to causes—because without true randomness, the universe and everything in it is purely deterministic.

$\endgroup$
  • $\begingroup$ > Randomness has proved an essential component of machine learning $\endgroup$ – lambda Sep 24 at 0:08
  • $\begingroup$ @lambda here I mean that, although one could create a learning algorithm that uses non-random search, such methods seems less effective than MCTS or informed search that utilizes some randomness, and the same seems to hold true for evolutionary algorithms, in most intractable problems we've applied them to. $\endgroup$ – DukeZhou Sep 24 at 0:16
  • $\begingroup$ "it would absolutely be possibly to brute force anything", I would specify "brute force anything computable". $\endgroup$ – nbro Sep 27 at 0:29
1
$\begingroup$

Yes, randomness is necessary to achieve generality in theory. Right now AIs we have are on the basis of seeking pattern and use them to predict future moves or outcomes. If we don't include randomness in data then machine might consider that as pattern and behave according to that (Which will be bias for us). Generating random numbers is a different story in itself and won't be a criterion alone to judge. While this might be one of the conditions for sure.

$\endgroup$
  • 1
    $\begingroup$ Even if randomness is connected to creative writing, it's possible to formalize to process of story telling to eliminate randomness. This would result into a straight fairy tail which is build upon logic but not feelings. $\endgroup$ – Manuel Rodriguez Sep 23 at 12:27
  • 1
    $\begingroup$ If any system is designed such a way that it picks a particular sentence to start with and ends with one, without any randomness. It would be having a limited number of stories to come up with. Number might be quiet large, but will always be bound by permutations. Reinforcement Learning uses policies which are nothing but some guidelines and there is an end goal. In this kind of system, randomness exist during learning phase though they are bound by logic. $\endgroup$ – Rajat Paliwal Sep 23 at 13:04
  • $\begingroup$ Brute force (exhaustion) would not more limited than random composition--in both cases there would still be the same number of potential configurations. Only the approach to forming them is different. (See "The Library of Babel") But there is still the "needle in the haystack" problem, and the statistical approach has proven effective where brute force is insufficient. $\endgroup$ – DukeZhou Sep 23 at 21:12
  • $\begingroup$ Why it will be a limited number of stories to come up with? Given word 0 and word 1 we can generate all (which is infinite) binary stories without randomness. $\endgroup$ – lambda Sep 23 at 23:50
1
$\begingroup$
  1. It might be too philosophical answer, but maybe first we need to answer the question whether a human way of thinking or his creativeness includes random elements. For example if an author writing a book uses some randomness in developing some side thread or some episodic character and I would say, that yes - sometimes we think up of something random.

  2. Some algorithms uses randomness at their basis, for example evolutionary algorithms for generating first population.

$\endgroup$
  • $\begingroup$ All machine learning algorithms that have proven effective seem to use the statistical approach, which definitely involves randomness. So not just genetic algorithms, NNs in all of their flavors, native Bayes, et al. $\endgroup$ – DukeZhou Sep 23 at 21:10
  • $\begingroup$ I think the randomness we observe in human decision may not be true randomness, it may just be a function taking history, environment as arguments which is too complicated that it behaves like random function. $\endgroup$ – lambda Sep 23 at 23:38

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.