What's new at Institute.

technique that will take us to real artificial intelligence

one day.

So the idea in reinforcement learning

is that you are not told what is the correct answer

or what is the correct strategy.

Instead, the neural network explores

many different strategies.

And the only thing it gets as feedback

is some kind of reward that tells it

whether the strategy was more or less successful.

So that is very good, because then you can start actually

solving new problems, problems that you haven't solved before.

You don't know the good strategies.

You're only able to tell what is a good solution

versus a bad solution.

So I want to start by reminding you of the basic setup

that we are looking at right now.

It's one particular reinforcement learning

technique that is called policy gradient.

It's, say, one of the two big classes of reinforcement

learning.

And the setup is shown here.

You have some kind of agent that is

interacting with an environment.

The environment is the world around the agent.

The agent makes some observations.

Through the observations, it learns something

about the state of the environment.

And then a policy or a strategy consists

in mapping that state into an action.

So it asks the question, what do I

want to do next if I observe this particular state?

What should I do next?

And then it executes this action,

which will also involve changing the state of the environment.

Usually, you move around some object, for example.

Then you will do the next observation,

which will confirm that the state of the environment

has changed.

You map that state again into an action, and so on and so on.

And so you go along until there's a cutoff.

Maybe you have set a total time that

is allowed for executing the strategy,

or maybe there is some goal that is defined.

And once you have reached this goal, then you finish.

And so there could be several different ways

of implementing such a policy.

You could think of a deterministic mapping

from state to action.