Now give us a punishment. Okay.

How much time does that leave us?

I gave you an example of perception using particle filters.

That just looks something like this select from Post order by created descending.

This is a beautiful framework.

Notice that here we're introducing a new variable.

We never did that before in planning.

So today is an exciting day.

Disorder is feared...

So the only way to do this is to watch videos.

Resources are things like this pile of nuts and bolts that are used somewhere in a plan.

So the moral is we need some feedback from the environment. We can't just plan ahead and come up with a whole plan.

Thrun So let me give you some examples of this method in action.

We defined AI to be the study and process of finding appropriate actions for an agent.

Narrator Planning under uncertainty.

So here is the situation that does work and it's related to information space or belief space.

That is, taking the action of perceiving something.

So each state, except for the absorbing states, we have to define an action to define a policy.

It goes south exclusively to gather information.

Here's one more topic.

It doesn't matter.

So by doing both, and by observing that you don't consume the bandwidth, it still stays with you, you can watch them.

Of course, resources could be handled just in the language of classical planning.

You can't express the notion of All in propositional languages like classical planning, but you can in first order logic.

But you'll find a single sequence of actions and states that might cut the corners, as close as possible, to reach the plus 100 as fast as possible.

We talked a little bit about machine learning, a lot about particle filters, and some about motion planning, which relates to the planning class that Peter taught you quite a while back.

One of the key problems is that these robots have to decide what to do next.

We never brought planning and uncertainty together, uncertainty and learning, or planning and learning.

Whereas if you need memory, it's partially observable.

Narrator We've been considering sensor less planning in a deterministic world.

We distinguished deterministic was the casting environments, and we also talked about photos as partial observable.

So we made a robot that went inside and built maps of those mines.

Now we have to figure out what actions look like and what the results of those actions look like.

Now, the trickiest part about situation calculus is describing what changes and what doesn't change as a result of an action.

This unit we'll return to the topic of planning, and we'll talk about 4 things that we left out last time we talked about planning.

So what you'll find in the last case, E, is that the dollar amount and the bandwidth is insufficient to make you knowledgeable about planning and scheduling at the same time.

This is a stochastic state transition which we can equally define for actions, south, west and east, and now we can see a situation like this conventional planning is insufficient.

Just as in problem solving, the belief state depends on what type of environment you want to deal with.

It's really different from planning than before because of the stochasticity of the action effects.

So we now learned about fully observable environments, and planning in stochastic environments with MDPs

We'll talk about planning under uncertainty, and it really puts together from the material we've talked about in past classes.

So far, we talked about the fully observable case, and I'd like to get back to the more general case of partial observability.

So for example it might be worth 10 to be in the state over here, 0 to be in the state over here, and 100 to be in a state over here.

In fact, you can use value iteration MDP's value iteration in this new space to find a solution to this really complicated partially observable planning process.

In the core classical planning, the belief state had to be a complete assignment, and that was useful for dealing with deterministic fully observable domains.