Original subtitles

This is the first segment of the artificial intelligence planning course.

In this segment I will give you an introduction and overview to the problem

we are addressing in the field of A.I. planning.

This will include some examples and introduction to the basic techniques we

will be using to solve planning problems. So the first question I have to answer

is, what is planning?" And more specifically, what do we mean by planning

in the context of artificial intelligence?" I will answer this

question by informally describing the planning problem that is the problem we

are trying to solve in this field. I will then argue why this problem is

important for artificial intelligence as a whole.

And then continue to describe some techniques that will be used to solve

this problem. So let us start by looking at human

planning and acting. Humans rarely plan before acting in

everyday situations. Ask yourself, when was the last time I

sat down and made a plan before acting? Chances are this will have been some time

ago. This is because humans act without prior

explicit planning quite often. There's a number of situations where this

is the case and here are some examples. When the purpose of my action is

immediate, I don't need to make an explicit plan. For example, to record

this lecture, I needed to switch on this computer. I know how to do this, so I

just did it. I didn't need to make an explicit plan.

The purpose of the action was immediate. When performing well-trained behaviors, I

also don't need to do explicit planning. For me, this would be driving a car.

I know how to drive a car. I've done this many times, so I don't

need to make a plan before I switch gears or before I turn the steering wheel.

It's a well-trained behavior, I don't need to plan.

When the course of action can be freely adapted, I also don't need to plan.

This would be when I go shopping in the supermarket.

I don't need to plan in which order I go through the different aisles,

because I can always adapt my my acting to what I've missed in previous aisles

and just go there again. So the, the course of action can be

freely adapted, means I don't need to plan.

A number of situations make it possible to plan, though,

and here are some examples where planning is necessary,

that is explicit planning. So, when I'm addressing a new situation,

something that I haven't done before or haven't done often,

then I need to do explicit planning. An example of this would be moving a

house. Everybody who has done a big move with,

with furniture will know what this means. You need to organize a van.

You need to organize people. You need to have an explicit plan in

place before you can successfully move from one place to another.

another situation is when the task you're trying to achieve is very complex.

So, for example, when I was planning this course, I was doing explissive planning.

This is quite a complex task, it involve ten hours of lecturing and many other

things, so explicit planning was necessary.

Another type of situation where acting happens only after planning is when the

environment imposes a high risk or a high cost.

So if I'm the manager of a nuclear power station, I will do a lot of planning

before I act, because it's very important what I do and the potential damage I can

do with wrong action is high. So,

I will do explicit planning to counteract that.

Also, when I'm collaborating with others, explicit planning can be extremely

helpful. So, think of people who are trying to

build a house. That's the people who are trying to put

up the walls, trying to put in the plumbing, and of, the electricians.

They all need to coordinate their activity and that means they all need to

have an explicit plan for when they do what and in which order.

So the main lesson here is that people only plan when it's strictly necessary.

We don't do planning when we don't have to.

We only plan when we feel there's a benefit to it.

And this is because planning is a complicated and time-consuming process.

There is a basic trade-off here. If we plan, we normally come up with a

course of action that leads to better results,

but there is a cost. So, if there is no benefit to be had from

planning, we're often better off not planning.

That is, often we seek only solutions or plans that are good enough for what we

are trying to achieve, not optimal plans. So people only plan when it's strictly

necessary. Here is the definition for what we mean

by artificial intelligence planning. Let me read this out for you first.

Planning is an explicit deliberation process that chooses and organizes

actions by anticipating their outcomes and that aims at achieving some

pre-stated objectives. So I will try to take this apart for you

now. What this says is, planning is an

explicit deliberation process. What this means is, to plan, we need to

think. It's a mental process where we think

about the actions that we are trying to do.

It also needs to be explicit thinking, which means, it's conscious.

It's not a subconscious process that's going on,

we are aware that we are doing this planning so we are thinking about

planning. In this thought process, we choose and

organize actions. So, choosing means, we have some options

available, things that we may be able to do.

And we choose some of these actions, and we discard others, as part of the

planning process. We also organize these actions into a

structure. That is,

we could choose which actions to do before which other action,

which actions to do in parallel, what the outcomes of each action will be,

etcetera. So we organize them into some structure.

And, the way we do this is by anticipating the outcomes of the

different actions that we have available as options.

So we think about, what will the world be like if we do this action?

And the result is either what we want or don't want and that's what the next point

is. The process aims at achieving some

pre-stated objectives. So there are things that we want to have

true in the world, these are our objectives,

and by anticipating the outcomes, we can compare the world states as they will be

when we execute an action to the ones in which the objectives we try to achieve

are satisfied. So that is what we mean by planning.

Planning is an explicit deliberation process that chooses and organizes

actions by anticipating their outcomes and that aims at achieving some

pre-stated objectives. Artificial intelligence planning now is

the computational study of this deliberation process.

So what we're interested in is the thinking about plans, the reasoning about

actions that takes place when we are planning and we are trying to build a

computational model of this process. Now that I've defined what we mean by

planning, I want to explain to you why it is so

important to study planning in artificial intelligence.

The goal of artificial intelligence is really twofold,

there's a scientific goal and an engineering goal.

The scientific goal of A.I. is to understand intelligence,

and the key observation here is, that planning is an important aspect of

intelligent behavior. So, if we observe some intelligent

behavior, we assume that there is an underlying plan and we assume that this

plan is the result of some planning. So, to understand intelligence, we need

to understand planning, which is part of intelligence.

In that sense, understanding planning directly contributes to the scientific

goal of A.I.. The other goal of A.I.

is the engineering goal, which is to build intelligent entities,

that is we want to build robots or other entities that exhibit intelligent

behavior. And if this is to be intelligent to us,

this needs to involve actions that are carefully chosen and organized as we do

in planning. So what we do in planning is we build

models of how this planning works and these models are software models, so we

can build them into our intelligent entities as components.

So planning directly also contributes to the engineering goal of A.I.

And just as a side remark, the robot you see here is the Shakey robot that was

built in the late 60s and that was one of the first robots that used an actual

planner to come up with its actions. There are really two different types of

planning, domain-specific and domain-independent planning.

In domain-specific planning, we use specific representations and techniques

that are adapted to each problem we are trying to solve.

There are a number of important examples for this type of planning,

domain-specific planning, for example, path and motion planning.

If we are trying to navigate a robot through a two-dimensional or

three-dimensional space, we need to come up with a path through that space, that

gets the robot from one location to another.

And to do so, a number of algorithms have been developed to, to make sure that the

robot doesn't bump into other objects or will fit through narrow passages.

All these algorithms are highly specific and very efficient.

Another example is perception planning. If we try to understand a given situation

a robot may have to wander around in a scene and observe different aspects of

different angles to understand what is going on.

And again, there are highly specific algorithms that have been developed for

this type of problem. Manipulation planning is another such

problem where we are trying to, for example, assemble an object from

different parts and that needs to happen in a specific order for it to work.

also, natural language generation uses highly specific algorithms for planning,

namely the planning of utterances that lead to communicating,

as given subject. The point is in all these domains, we have specific

algorithms that we use to efficiently solve a specific problem.

On the other hand, there's domain-independent planning.

And there, we use generic representations and techniques to solve the generic

planning problem. The advantage of this is that it saves

effort, so we don't need to reinvent the same

techniques for different problems all the time.

We can always reuse the same planning algorithms.

The disadvantage is that, this means planning from first principles and is

often relatively slow, but it also leads to a general

understanding of planning and as I've just explained, that's the scientific

goal of artificial intelligence. The important lesson here is that

domain-independent planning complements domain-specific planning.

Domain-specific planning is good for specific problems where highly efficient

solutions are required. Domain-independent planning is good if we

need to plan from first principles for the type of situation I've explained

earlier, situations we have never seen before for

example. So the two types of planning complement

each other. But in this course, we will focus on

techniques for domain-independent planning.

So here's a little quiz to test your understanding so far.

The following five statements are either true or false.

Please tick the box for the statements that are true.

The first statement, people only plan when they have to because the benefit of

an optimal plan does not justify the effort of planning is true.

The second statement for humans planning as a subconscious process, which is why

computational planning so hard is false. The reason is that planning is not a

subconscious process. We have defined planning as the explicit

deliberation process, so it needs to be conscious.

Third statement, planning involves a mental simulation of actions to foresee

future world states and compare them to goals,

that statement is true. fourth statement, in artificial

intelligence, planning is concerned with the search for computationally optimal

plans, that statement is false.

We're not only after optimal plans, we also want to sometimes find out

whether a plan exists at all, whether it's optimal or not.

Finally, domain-specific planning is used when efficiency is vital,

whereas domain-independent planning is good for planning from first principles.

That statement is true again.