Most, if not all, AI systems do not imitate humans. Some of them out-perform humans. Examples include using AI to play a game, classification problems, auto-driving, and goal-oriented chatbots. Those tasks usually come with an easily and clearly defined value function, which is the objective function for the AI to optimize.

My question is: how is deep reinforcement learning, or related techniques, to be applied to an AI system that is designed to just imitate humans but not outperform humans?

Note this is different from a human-like system. Our objective here is to let the AI become a human rather than a superintelligence. For example, if a human consistently makes a mistake in image identification, then the AI system must also make the same mistake. Another example is the classic chatbot to pass the Turing test.

Is deep reinforcement learning useful in these kinds of tasks?

I find it is really hard to start with because the value function cannot be easily calculated.

What is some theory behind this?


1 Answer 1


There is an interesting discussion the progress achieved in this field by far in the paper of Francois Chollet - https://arxiv.org/abs/1911.01547. At the present time, many architectures are able to outperform the human in particular tasks, because they have a strong priors coded into them and ability to process a huge amount of data.

However, when it comes to generalization, or even more difficult task of doing the things, that the coder has not put into the model, the present algorithms do no perform well. There is a rather sophisticated and developed mathematical definition of what is required from the system to be called intelligent. And in a nutshell, it is ability to develop a sensible behaviour and enough accuracy for new tasks without putting strong priors and large amount of experience.

In the end of the paper, there is proposed a benchmark to measure the intelligence of AI system.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .