In some learning algorithms, we don't directly train models by datasets with labels to predict, but rather we create 2 competing models and let them fight/compete against each other. As the many millions of epochs pass, the models fight each other, and every time each model improves itself (further optimise its weights) to win. After many epochs of the models smashing each other, eventually they become really strong super-hero models that can totally blow any human out of the water. This approach seems to be often used with machine learning models that are tasked to play multiplayer games. Instead of letting them play with the slow humans, they fight with each other to death for many many epochs to become way stronger than any human can naturally be.

What is the name of such kind of machine learning approach?

  • 2
    $\begingroup$ You can actually cut down a lot of your question. Are you talking about actor-critic approaches? There are also GANs (Generative Adversarial Networks) $\endgroup$
    – N. Kiefer
    May 5, 2021 at 9:42
  • 1
    $\begingroup$ @N.Kiefer - BTW someone here seems to be calling it "multi-agent reinforcement learning". What do you think? Is it not the same thing? $\endgroup$
    – caveman
    May 5, 2021 at 9:47
  • 1
    $\begingroup$ Here’s a link about what they call “competitive self-play.” Also, another link from the same source mentioning multi-agent reinforcement learning. $\endgroup$
    – S2673
    May 5, 2021 at 10:17
  • 2
    $\begingroup$ actor-critic or adversarial models. $\endgroup$ May 5, 2021 at 14:50
  • 1
    $\begingroup$ I wouldn’t say there is competition in actor-critic models. The critic is guessing returns to help actor updates, not trying to become better than the actor. $\endgroup$
    – S2673
    May 6, 2021 at 10:56


You must log in to answer this question.

Browse other questions tagged .