For the purposes of this question, a technique is considered adversarial if it develops two systems (or two kinds of system), the first trying to process certain kinds of data well, and the second trying to provide the first with hard-to-process data. Examples include:

  • Adversarial machine learning, in which the processor tries to identify security risks, while the agent trying to fool it seeks to bypass security checks;
  • Generative adversarial networks, in which the processor is discriminative (i.e. it models a conditional probability distribution) and the agent trying to fool it is generative (i.e. it models a joint probability distribution).

Judging by the dates discussed in the above links, AML dates to the early 2000s while GANs date to 2014. These both seem very recent, and I suspect the general meaning of "adversarial" I specify above has been used for much longer. Clearly, the idea of requiring continual adaptation by both agents to shifting goalposts need not be tied to such specific examples as security or Bayesian models. Any simulation of two competing species evolving, e.g. a predator and its prey, will be "adversarial" in the sense of this question.

So my question is: when was some kind of adversarial technique first used, in the hope it would lead to machines learning to solve problems more effectively than would happen without adversarial techniques? (A simulation that illustrates how biological evolution works, without wanting to solve a specific data-driven problem of human interest, would not qualify.)


1 Answer 1


I am unsure as to the parameters of what is being asked, so this may not answer it. First, "data-driven problem" is a term linked to the recent rise of data science and projecting it too far into the past risks modernizing people's intentions. Second, biological evolution need not work by "competition" between two or more species, natural selection applies even to a single species in its "competition" with the environment. And in artificial selection (real or simulated) the environment can be intentionally manipulated to make it an "adversary".

Lotka and Volterra developed predator-prey models in 1920-s, but there were no "data machines" then and this probably would not be "wanting to solve a specific data-driven problem of human interest". Closer is perhaps Turing's celebrated 1950 paper Computing Machinery and Intelligence, where he proposed his "imitation game" where a machine tries to fool a human interrogator, the origin of the "Turing test" of machine intelligence, and even suggested "evolutionary learning" as a means of achieving it:

"We have thus divided our problem into two parts. The child-programme and the education process. These two remain very closely connected. We cannot expect to find a good child-machine at the first attempt. One must experiment with teaching one such machine and see how well it learns. One can then try another and see if it is better or worse. There is an obvious connection between this process and evolution... Equally important is the fact that he is not restricted to random mutations. If he can trace a cause for some weakness he can probably think of the kind of mutation which will improve it... It is probably wise to include a random element in a learning machine... Since there is probably a very large number of satisfactory solutions the random method seems to be better than the systematic. It should be noticed that it is used in the analogous process of evolution."

It is not clear to me that Turing proposes automating both "adversaries" here, but this paper is often cited as seminal on machine learning and genetic algorithms, where candidate solutions to optimization or search problems are treated as an evolving population of "individuals" with "traits", and generational selection is implemented according to "fitness". It is believed that Fraser's 1957 simulations of artificial selection incorporated all the requisite elements of such algorithms, but again, I am not sure if this was "data-driven". Bremermann in the 1960s used "evolving programs" specifically to solve optimization problems, see e.g. Revisiting Bremermann’s Genetic Algorithm by Fogel and Anderson.

As for the more narrowly understood agent vs agent machine learning that is indeed recent, although computer programs playing against each other dates back to 1980-s.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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