I was training an AI to learn things during its lifetime such as find food and navigate a maze. Behaviors that might change during its lifetime.

But I hit upon a snag. Some behaviors, like avoiding poisonous snakes, cannot be learned in a lifetime since once bitten by a snake the being is dead.

That got me thinking about how to separate out behaviors that must be given to the AI at birth (either by programming or using some evolutionary algorithm) and which behaviors to let the AI learn in its lifetime.

Also, there is the matter of when a learned behavior should be able to overrule an innate behavior (if at all).

Is there much research into this? I'm looking for some method to determine a set of innate behaviors which can't be learned.

  • 1
    $\begingroup$ You think that certain behaviors, like avoiding poisonous snakes, are innate, but they are not. Humans (and other animals) usually learn to avoid poisonous snakes in a supervised fashion, rather than trial-and-error: in other words, someone tells you that certain things should not be done. So, my suggestion would be to try to implement a multi-agent system, where an agent also learns from other agents: for example, if an agent sees that after another agent ate some poisoning food it died, then it should try to avoid it too, given that it doesn't want to die. $\endgroup$
    – nbro
    Jul 20, 2019 at 21:44
  • $\begingroup$ However, even though I think your question arises because you thought that, if a behavior cannot be learned by trial and error, then it must be innate, my previous comment does not answer your question. So, please, maybe update your question. Do you still want to know about "innate behaviors"? $\endgroup$
    – nbro
    Jul 20, 2019 at 21:52


You must log in to answer this question.

Browse other questions tagged .