What is the input of the neural network during reinforcement learning and what do we have at the output?

Question

Neural networks

Good day.
When reading a book about reinforcement learning, the following questions arose:
1. What is the input? Environment conditions? Action value? If the former, how is the latter used? That is .. How the value of actions affects the choice, if we submit states to the input. And vice versa. How does the agent learn about the states of the environment if the input is the Value of actions. I'm sorry for the stupid question, but I'm just confused.
2. Do I understand correctly that there should be possible actions at the output?
3. What does a neural network look like? Is it a perceptron or something else?
Can you give a simple example from a game like Dendivian Ping-Pong? I don’t understand the example of a multi-armed bandit, at least kill ...

Answer 1

As input - the state of the environment, somehow encoded, plus a possible action. The output is the value of this action. In their first article, DeepMind calculated values ​​for all actions at once, so there was not one output, but N. But this is not necessary.
Plus, there are more complex actor-critic reinforcement learning models.
The architecture of the neural network depends on the way the environment is encoded. It can be a persptron, or it can be convolutional neural networks if we give a visual signal. And it can contain recurrent networks, if the input is text or some other sequence.

Answer 2

One of the easiest examples https://github.com/pytorch/examples/blob/master/re...