In this video nugget, we'll talk about replanning and execution monitoring, which is some new

things we have to add to the planning level to go online.

Remember, in the last video nugget, we actually talked about online search, which is kind

of at the atomic states level with belief states.

And now we raise that to the planning level.

So basically we can turn any planner P into an online problem solver by just adding a

plan action, replan with respect to some goal state G or set of goal states.

And of course, just by adding this level, we give ourselves a lot of freedom to do online

things.

For instance, just the plan that just says replan G, which basically gives us, makes

a complete plan, which is trivially complete because it kind of says, well, just plan again,

gives us for any goal, gives us a complete plan.

Of course, this is totally unhelpful because we can always do this.

So at the other end, if we have a plan with subplans for every contingency, what to do

if a meteor strikes or if I win the lottery or something like this may just be much too

costly and much too large.

That's kind of the other end where we plan for everything and don't replan at all.

Whereas here we kind of replan directly.

So there's kind of a trade off between pre-planning and replan.

Think about, for instance, if you drive into the desert with your car and there's the contingency,

of course, that your tire might blow and you cannot go on.

And in this case, you actually want to have a plan or want to have planned for that contingency

because the effects are actually catastrophic.

So you want to make sure you have water with you as of the plan.

And we navigate this trade off between pre-planning and replanning all the time.

If I make my way to the university, it might just be that there's a big puddle of water

in front of the main door, which I didn't know about and couldn't know about.

But I still don't have a plan of what do I do when I reach a big puddle in front of the

blue tower, but I'll just deal with it as I go along.

I might just not take the direct way from the bike to the front door, but kind of make

a detour or something like that.

We have to kind of find a way around this trade off.

We do it, but the planners also have to do it.

The other thing is, and that's actually just in a deterministic or fully observable deterministic

environment.

If we have an environment that is less good, stochastic or partially observable, we also

need to have some kind of a form of execution monitoring to determine when we kind of need

to repair a plan.

And that's something which I would like to show you, is if we have a plan which we might

need to repair because something goes wrong.

Say we have a plan, I call it whole plan, that goes from the initial state to the goal

state, and we start executing it.

One step, two step, three step.

I have a non-deterministic action here, and instead of putting me into a state E, which

I have expected, I land in a state O.

As for O, I'm somewhere else.

So if I do plan, if I do some kind of execution monitoring, my sensor tells me, oh, you're

in state O instead of state E, so I have to do something about it.

And of course, what we do as humans and of course also as planners is we have to find

a repair plan, something that gets me from state O to, well, where?