YouTube Transcript:
Mike Cook on PCG - invited talk - AIBU

Skip watching entire videos - get the full transcript, search for keywords, and copy with one click.

Video Transcript

Video Summary

Summary

Core Theme

Procedural generation (PCG) is presented as a creative craft akin to gardening, emphasizing the iterative process of controlling and introducing "wilderness" to create unique and engaging content, rather than solely focusing on algorithms or final output.

uh please let me introduce you uh to the

audience we have

uh this is a hybrid meeting so we have

uh both students here uh in the lecture hall

hall

uh and we also have students

participating on resume

so when it's time for questions which

we'll do after your talk

we will take questions from the audience

but i would like to welcome you mike to

stockholm uh are you in london right now

or i am yeah yeah so mike is a

researcher at queen mary institute in

london uh where he works on ai for games

and in particular pct he's also the

founder of proc jam

and the author of one of the coolest programs

programs

that exists which is angelina who can

actually make her own games as being an ai

so it's a real honor to have you here

mike i'm so happy to to see you too um

um so

yes please uh take it away

thank you so much mario um it's really

wonderful to be here as well and i'm

very jealous to all of you that you get to

be in a class taught by miriam because

she's one of my favorite people in games

research um and that was why it was

impossible to say no to this opportunity

to talk to you all today

um so thank you for that intro miriam

and hello everyone

my name's mike um and miriam's given you

a little bit of background on me already

um but today i want to talk to you about

procedural generation of course and i

understand that you've learned a little

bit already

and i wanted to do something a little

bit different with this talk i hope you

enjoy it and we're going to come to it

from a slightly different angle and

think about um why we use procedural

generation and what it can do for us and

what it's like to actually use it in the

wild so

i'm going to start my stopwatch so that

i know when we're 15 minutes over time

um i'm only joking miriam don't worry

so a little bit of background miriam's

already said a little bit um but i i'm

an ai and games researcher and also a

game designer and i work in a field

called automated game design

where we're interested in using um pcg

techniques to build ai that can design

games so we tackle really all kinds of

different generative tasks

and i also work in a field called

computational creativity um

and in there we're really interested in

how ai is perceived by people so we're

not just interested in whether ai can

make a game that's good but we're

interested in how people feel about that

game and how they feel about the ai do

they believe that it's independent and

do they believe that it's creative so

there's a lot of really interesting like

philosophy in that too

i also make games as part of my research

and in my spare time like uh condition

unknown and rogue process that we're

going to see a little bit of later

and if you want to talk to me but you

don't get a chance to answer you know

ask a question today or maybe you're a

bit shy or anything please feel free to

contact me later i'll have all of my

details at the end of the talk

so i am here today to talk to you about

procedural generation um and when we

think about procedural generation maybe

it's completely new um for some of you

or maybe you've played some games like

minecraft which which this screenshot is from

from

um we think about certain kinds of

things some of us might even kind of be

negatively biased towards procedural

generation it's been used in very

specific ways in the past so far

i uh founded a game jam all about

procedural generation it's actually kind

of a community we all get together every

year and it's very laid back and relaxed

we share lots of resources and help

um and it's all about procedural

generation and the slogan of the

procedural generation jam

is make something that makes something

and i really like this slogan um i think

uh very proud of it and i think it's a

great invitation to sort of join the

community and start making things

and when i sat down to make this talk

this was the first slide i wrote um

but then i realized that although it's a

good invitation it doesn't really

help people kind of get into it and a

lot of the questions that we get are

things like well where do i start what

techniques do i use it doesn't tell you

what it's like to really make

a procedural generator

and so today i'm going to cover it from

a different angle with a different meta

message and a bit of a

stretched metaphor

but today i want to talk to you about

procedural generation as a form of gardening

gardening

and i want to think about what that

means with you and think about hopefully

how that can give you a perspective on

what i think procedural generation is

for and how to

go about

thinking about it yourself so i hope

this is fresh and interesting to you and

i hope that you you enjoy coming on this

uh weird metaphorical journey with me

so the inspiration for the gardening

metaphor actually comes from a couple of

places the first is max kraminsky's

amazing talk at the roguelike

celebration in 2018. all of these talks

are up on youtube and i really recommend

looking at them if you're interested in

this topic but max's is maybe one of the

best talks i've seen given at uh the event

event

and in it they talk about gardening from

the perspective of the player so they

talk about games in which you have a

sort of gardening like approach a

relationship with the game world so this

is a screenshot from animal crossing and

you're really encouraged to sort of

engage with these randomly generated worlds

worlds

but as i rewatched max's talk i began to

think about how gardening kind of

relates to how we build procedural

generators as well

and it made me think of another example

which is um emily shorts uh

emily shaw's annals of the paris

um and this is maybe one of the best

pieces of generated text in existence i

would say you can get it online at in thewalls.itch.io

thewalls.itch.io

and it is a 50 000 word travel guide and

there's lots of segments like this so

this is describing a place called

hasselchester and you have like these

poetic descriptions of the place

sunlight filtered through rustling

leaves and it tells you things like the

mythology of the area where you can stay

where you can get a hot meal

who is the lord of the area all sorts of

things and it's all generated by emily

and in fact it was written for something

called the national novel generating

month where the challenge is to generate

a 50 000 word document um in one month

using generative techniques and

and

the thing that's really interesting

about it is um is that emily um includes

an appendix at the end um where she

describes the process of writing the

annals um and i think it's one of the best

best

segments uh about procedural generation

that's ever been written like it's one

of the best pieces of writing about

procedural generation um and if you only

look at one thing as a result of this

talk i really recommend you pick that

out and have a look at it

and there's a few sections in there that

really stand out to me and i've quoted

one of them here emily writes i spent

considerable time working on the machine

meaning the the procedural text

generator and its various features

before i produced anything that i

considered worth recording here

i would generate a few provinces worth

of content so that's descriptions of

like little towns

and then i would add that content to

this file and then i would read through

it and then i added to the engine new

content elements based on the

inspirations of the text generated so far

far

so emily had this really interesting

loop of working with her procedural

generator she would

write and edit things look at the text

that it generated add things to the

final output go back to the generator

and it reminds me a lot of the way

people nurture and tend to their gardens

um emily would

add some and some things to the

generator plant some new flowers prune

some dead foliage off that she didn't like

and then wait and see what grew what bloomed

bloomed

and then come back to the generator make

some changes and go through this sort of cycle

cycle

and i began to think about how

how

sort of satisfying this metaphor was not

just because of this cycle of growing

things but also

the sort of two extreme types of garden

really relate uh very well to um

types of content that we put in our

games and where procedural generators

start from so i want to tell you about

what i think that the two extreme types

of garden are and how that relates to

procedural generation

so in some gardens we have complete

control over absolutely everything in

the garden so this is an ornamental sort

of garden from iran where almost

everything you can see here has been

precisely and perfectly placed every

flower every tree the pathways the water

you know the hedges have been trimmed to

be a particular shape

and we can think of these sorts of

gardens as similar to when we make

content for our games by hand so when

we're designing a game um we place every

single uh individual bit of content in

the world in the soundtrack in the

storyline exactly where we want it to be

it is perfectly designed as we wish it

to be there's nothing uncontrollable

about it or unpredictable

and this is actually where a lot of

procedural generators start from so

if we are really good at level design

maybe some of you have made platforming

games before um

that might be where we want to start off

with our procedural generator we already

know how to make really good platforming levels

levels

and so what we often do is we start from

this point of

something that we know very well and

then we begin to think how can we give

up control over little parts of this how

can we begin to let some wilderness kind

of creep in and not have everything so

perfectly designed

and this is a very common path for for

many types of game designers so derek

you for example um whose name you you

may or may not recognize but um back in

the sort of early 2000s derek was

working on a number of indie projects on

his own

and he got a lot of experience designing

levels and all of these games are

completely perfectly controlled they are

tended to and ornamental gardens where

every single part of the level is

exactly where derek wanted it to be

but later on closer to the 2010s derek

was starting to make a game called

spelunky and in spelunky derek wanted

there to be some variation he didn't

want players to be able to memorize what

the levels were going to do he wanted

there to be unpredictability

in other words he wanted to take this

perfectly ornamental garden that he

would normally design

and then allow it to sort of grow

outside of the templates a little bit

and do things that he wasn't expecting

and so if you've not played spelunky

before this is what the game looks like

when you're sort of playing it you're

quite zoomed in but if we zoom out we

can see um the full scale of a level here

here

and if i add in some lines and what

you'll see is that this level is made up

of a 4x4 grid of smaller rooms and each

one of these rooms is based on a

template that derek designed himself

so derek's brought his knowledge about

designing levels the skills he already has

has

and then he's found a way to allow this

a very simple algorithm to take those

levels and

mix them up um to give up some control

that he would normally have over this

process of level design

and he actually layers on a little bit

more than that so each of these

templates he designs they actually kind

of look like this um and these these

tiles that i've labeled as 50

when they get loaded into a level like

this the game rolls some dice and either

that's going to be a solid block if 50

rolls high and it's going to be empty if

it rolls low and there's loads of little

rules like this for adding in traps and

enemies and treasure

and what it means is that the heart of

spelunky is designed by derek but the

game that people play is a bit like a

garden that's been slightly overgrown

there's some wild flowers poking out

through the stone walls and there's a

bit of moss on the side of the shed it's

not exactly as he as he designed it

there's a little bit of variation in there

one end of the spectrum is that we start

from total order and we allow a little

bit of wilderness to grow in the other

end of the spectrum is what i would call

total chaos so if you've ever kind of

come to a house that's not been tended

to for a while or you've wandered around

like an abandoned building site or

something like that you'll find spaces

which are completely overgrown there is

no control here whatsoever this is not

a manicured garden in any way there's

weeds and long grasses and odd

flowers that you've never seen before

and some procedural generators start

from these points as well

and so a common example is

the use of something called noise

functions so i don't know whether you've

come across these before but a noise

function is a way of essentially getting some

some

random looking numbers um that have

maybe some kind of regularity to them

but but not in a predictable way so what

we're actually looking at here is

two-dimensional noise there's actually

something called purlin noise and you

don't need to know exactly what that

means but

essentially if we give this function two

numbers um it will give us a number

between one and zero as the output

so every pixel on this image that you're

seeing here the x and y coordinates of

the pixel have been fed to this function

and it's given us a number between zero

and one

and then we've mapped that to a color on

the gray scale between white and black

where like zero is white

one is black and then point 0.5 is gray

and what you get with this noise

function is you get a very kind of

cloudy image you you get a sort of

unpredictable but very textured um output

output

and what people use do is they use these

noise functions to create organic

looking things so they were used a lot

in textures back in the 90s to make like

fire or clouds but a really common

application of them is to make worlds so

if we map

these numbers to kind of height

we can create maps so this colored image

here the white is sort of the tips of a

mountain and then there's like the brown

hills the green grass yellow beaches

down into a shallow water and then ocean

and all we've done here is mapped the

noise function to colors and it's

created a kind of world map for us sort of

and this is also can also be done in 3d

and in fact it's actually how minecraft

uh creates its worlds now

now

there are often you know a noise

function on its own won't do it for us

so what often happens is like this is

not an interesting world to play in and

so when people build these build games

using this uh sort of approach what they

often do is they often like combine

different noise functions and by the way

this was an oversight by me if you're

blue green colorblind i apologize this

basically just looks like the background

image it's just recolored um but what

they'll do is they'll get multiple noise

functions and then they might kind of

stretch them or rotate them and lay them

over one another and then what we could

say is um this extra noise function that

we've got could decide where we're going

to put trees in our world

um but maybe it's a special kind of tree

that only grows up a mountain and so we

have this sort of rule that says

if the height map is above a certain

height and the tree noise function is

above a certain value then you can place

a tree there or something so

this is obviously i'm obviously waving

my hands a lot here but um these are the

kind of approaches that are used in

games like minecraft so minecraft uses

lots and lots of noise functions and

randomness and but then it has ways of

shaping them and lots of rules that sort

of prune and control exactly how they

come out

but the difference between the approach

we looked at before with spelunky is

that this is starting out with something

which is very unpredictable and then

we're sort of refining it over time so

minecraft's world generator is the

result of a lot of small tweaks and

changes by people until they get

something that they feel looks okay

and there's a lot of disadvantages and

advantages to using this approach over a

handcrafted approach um one disadvantage

is that this is very unpredictable but

it's also very organic it's very uh

interesting to walk around in which is

one of the reasons why it's so pleasant

so these are what i think of as sort of

the two starting points for most

procedural generators some of us start

out from something that we feel

confident making something that we have

a lot of control over and then we give

up some of that control

others start out with something

completely random that may be kind of

messy and doesn't what isn't what we

want and then they find ways of

filtering and shaping until they get to

something which which they can use

and to me procedural generation is about

everything in between those two extremes

it's about figuring out exactly which

parts of your garden you want to become

overgrown and which parts you want to be

tightly controlled

and that's where the joy of

experimenting in the same way that emily

short did with her travel guide comes

from i think

so i wanted to take a little detour here

to talk about

that first type of generator that we

talked about with spelunky and making

something from something that we're

already confident making so

i don't completely know uh all of your

backgrounds watching this talk but

one thing that i found a lot with

procedural generation is that a lot of

people are maybe a little bit hesitant

or they don't know how to start and a

lot of people think that they should

focus on the right algorithm they should

look for the right technique or technology

technology

and i wanted to talk about a very

different approach today that all of you

can go home today and do without

touching any code at all

and start thinking about what it means

to generate something start thinking

about those processes and

and

i've decided to stretch the gardening

metaphor even further by talking about

rewilding so

rewilding is a word from ecology and biology

biology

and it's essentially the process of

letting somewhere revert back to its

natural state so if you have somewhere

like farmland that's been very

controlled over many years um rewilding

is sort of allowing that farmland to

become run over with things that would

have grown there before lots of animals

return that kind of thing

and it's quite a big thing at the moment

in the uk funnily enough

so what i want to show you today is sort

of how you can go about the process of

rewilding something which maybe you're

confident in making so what i'm

imagining for this section is maybe you

are someone who hasn't used procedural

generation before maybe even as a bit skeptical

skeptical

but you are really good at writing or

level design or music composition or

something so how can you begin to think

about what what this technique means and

how it could be useful for you

so what you can do is you can get some

pen and paper um and uh in this case i

already had some cards so i kind of

cheated a bit but

what i did was i built um i drew some

buildings that um i thought were cute

um and

however you do this um you might decide

that you want to write some text or

compose some music or draw something

else like some people um at some point

you're going to need to find some places

to kind of cut them up into pieces um

it's great if they're kind of all

roughly the same size like i have here

with my convenient cards um

and then we're going to essentially

build a generator out of this process

so once you've done this

you can sort of flip over all of your

cards and shuffle them together and the

reason we flip them over is because we

want to think about how our code views

this content so our procedural generator

doesn't know anything about level design

it doesn't know anything about drama or

story writing it it just knows whatever

information we give it so we start out

with no information at all

shuffle the cards out and then think

about what structure we want to lay

these cards in so if you saw before i

had them kind of like a roof and then a building

building

and so i'm imagining this is kind of my

imagined rule for for generating some so

i'm going to place two cards on top of

each other from the shuffle pile and

then flip them over and see what i get

and of course one of the problems is

that because i haven't given any

information to my generator um we can

produce things that don't make sense so

you can have a roof on top of another

roof so this is clearly not what we want

you might get similar results if you're

generating something else like a head on

top of another head or a body without

any feet or whatever that may not fit

the thing that you're looking for

so that's fine so at this point you're

going to get a pen

and we're going to flip those cards back

over and we're going to give ourselves a

little bit of information so

you could limit yourself just to sort of

doodling on the back with a particular

color some kind of tag or like a single

word but you can't write like a full

description of what it is it has to be

like one word something very simple

something just to give you a bit of

guidance um

and actually for this stage it's even

better if um once you've done this

instead of you dealing out the cards um

you actually give it to like a friend or

someone else someone who really doesn't

know what's on the other side of them

and once we've added this bit of data

this is kind of like giving information

to our our generative algorithm right

so we can now have a slightly different

um set of rules for when we're

generating this so instead of just

putting two cards we can now say okay

well we want the top card to say roof on

the back and the bottom card to say building

building

and this is an additional bit of

information we're using some of this uh

data that we've given our system

and now we get a nice building out of it

because we've we've created some

constraints and our generator sort of

knows a bit more about what to do and

we've actually got a cool output we've

got a bakery that has an observatory on

the top um which is like interesting you

know it's a bit different

and of course you might not want this

you might want this this might be kind

of inappropriate for what you need your

buildings for so you might need to add

more uh labels or more information or

change the way you generate

but once we've got this down we can

actually think about

well now that i've got this basic system

i could create two-story buildings i

could create a building with a roof and

then two buildings underneath

and you begin to think about how you

could extend your generator

and then as you do this and this is

again going back to the cycle of like

extending and then looking at it and

thinking you can see here that oh the

second story actually has a door on it

which means if someone walked out of

this door they would fall one story down

and then hit their head on the floor so

this is no good so you know we go back

and we add some more labels maybe we add

a label to the bottom that says okay

this is a ground floor building this has

a door on it

and then we add some second story

buildings that don't have doors

so we have this cycle of going back and

forth and extending our generator adding

little bits of information and we don't

need to write any code we can think about

about

how are we changing the space of

possibilities with what we do

so every time we change our process for

dealing the cards out every time we

change the cards in our deck every time

we add a different label to the back of

one of these cards

we're changing the possible things that

this set of cards can represent

and that's really at the heart of

procedural generation is

thinking about all of these

possibilities and thinking about how you can

can

shape the ones that you want to make

them more likely and get rid of the ones

that you don't want to make them less

likely and think about how much surprise

and unpredictability you want in there and

and

really these paper-based approaches are

very effective often we want to do

something very complicated we want to do

something very fancy that uses the

state-of-the-art cutting edge stuff but

that doesn't always make us better at

using procedural generation and actually

derek use spelunky generator is still

probably the best example

of a procedural generator in any game

ever because it was exactly what that

game needed it was perfect for that game design

design

and so even though this approach seems

very simple it really is something that

you can play around with yourself and

because you will be experts in different

things to me um you know i don't know

exactly where the best place is to cut

up a piece of music or what the best

labels are for pieces of text

you might think that instead of chopping

a piece of music after the fourth bar it

might be better to chop it horizontally

along each instrument and connect

instruments together instead of bars and

you might have different ideas about

what labels to use to describe

descriptions of items or characters

but this basic thing of chopping

something up labeling it and

reassembling it like a jigsaw is the

foundation of a lot of really impressive

uh procedural generators and it's it's

not complicated it's it's something that

you can really get started with very

easily and start thinking about many of

i talked a little bit about why i think

gardening is a good matter for

procedural generation and how you can

think about this process of

sort of iteratively changing the

generator thinking about what it does

going back to it and changing it again

and i wanted to sort of back that up by

showing you some of the things i've used

procedural generation 4. some of them

complicated some of them simple and

and

i'm going to try and squeeze in some

demos if i can as well

so the first thing i wanted to talk to

you about was a game i was developing

called rogue process

rogue process was an action

platformer kind of in the vein of

spelunky where you broke into

skyscrapers and hacked security systems

to steal things from corporations um so

i wanted the skyscrapers that you broke

into to be procedurally generated i

wanted them to be different every time

um but i also wanted to be able to quite

tightly control the kinds of experiences

that you had in them and so

i ended up using an approach very

similar to spelunky's so what you're

seeing here is a catalog of loads of

different rooms that i designed by hand

and the the game would assemble kind of

like spelunky's levels by gluing them

together and making a whole level out of them

them

and just like the example we just saw

with the buildings and cards and writing

on the back of them

each one of these rooms is labeled and

the labels don't tell it sort of whether

it's a building or a roof or anything

but in this case they tell it where you

can exit each room from um so if i show

you a zoomed in example

you can see the room in the

sort of center of the screen you can

exit it from the bottom left and you can

exit it from the bottom right and those

are the labels that this room has

in case the level generator needs to

know kind of where can i place this room

what should be placed next to it

so this would never try and have you

exit from the top left because there's a

wall there you can't get there so the

labels here help the procedural

generator know where exactly it can

place it physically in the um in the building

building

and the nice thing about this is it

means i have complete control over

everything inside the room and the

algorithm just needs to know about how

to connect them up so it really is just

like a jigsaw

and much like spelunky we also took an

approach where there's sort of random

variation inside so the traps that you

can see in the bottom right and the

boxes that you can see in the bottom

left they're all randomly added to this

room so they're not always here and they

will be here sometimes and they won't be

here other times or there might be

something different there

and this is also great because it means

that i can easily play the game see how

a room feels go back to the editor

change it and then come back again

so this was a very simple approach it

was because the game did not need

anything more complicated um ultimately

this was an action platforming game um

it it just needed a little bit of

procedural generation just a little bit

of wilderness and unpredictability

um i did have uh just a little uh this

is from like an early trailer for the

game um just in case you wanted to see

although this was actually when the game

had a very different um procedural

generator ironically i couldn't find any

footage of um the game like after that point

point um

so procedural generation here is not the

focus as you can see the focus is you

know hacking things and moving around

and so

the game's procedural generator is

simpler and it kind of sits in the

background and that's often all the game needs

needs

now if we want to look at slightly more

interesting uses of procedural

generation i guess um i wanted to talk

to you about some games i'm working on

now so this is a series of games

nothing besides remains which i made a

few years ago condition unknown which i

made in 2019

and stargrave which i started last year

and i'm still working on so these were

all game jam entries but stargrave i

couldn't finish so um i'm i'm sort of

working on it every now and again

uh when i have time

so these three games are

they all have a similar core approach

and it's what my

incoming phd student florence calls

generative archaeology games so florence

is an archaeologist

themselves and is bringing all of this

knowledge to their phd when they start

next month and we'll actually be looking

at these games from a perspective of

archaeology and kind of combining that

with procedural generation

so what do we mean by generative

archaeology games

so the way these games work is we

generate um a kind of a very small world

so it could be like a village or a

spaceship or a research station

somewhere but it's quite it's quite

small and contained

and then we run a simulation of this

world and record what happens

and the important thing about the

simulation is that it always ends so

you'll see what i mean by this in a

second but basically it's not possible

for it to go on forever there is

definitely going to be a reason why this

simulation finishes

and then once it's done we put the

player at the in the world that results

and we let them kind of wander around

and uh explore and kind of understand

what happened

so if you've played information games

like return of the obra din or

the outer wilds

it's a little bit like those but with

some procedural generation behind them

so we're using procedural generation to

create this uh big messy world but it's

a big messy world that has

a real history like everything is

simulated so the player can go and see

and learn about what happened

and i've been trying to experiment with

how exactly we can have that kind of

make fun gameplay

so in nothing beside remains nothing

beside remains looks pretty ugly i made

this in 48 hours so there was there is

no art um but

basically i simulated like a very simple

farming village

and i simulated things like the health

of the ecosystem how good their farms

were um the

the

kind of

whether they were getting attacked by

wild animals things like that

and eventually the population of the

village kind of dwindles down to zero

people leave or there's sort of a famine

or something like that and then hundreds

of years later you turn up and can

explore the village

and as you explore um the village sort

of tells a story of what happened so if

you can find a church for example

um the people that lived in the village

will have engraved things that represent

kind of the struggles that their village

was facing so in this example um there's

an engraving of fields of of crops

growing and that means that this village

probably had problems with their farming

so they were sort of praying for for

better and harvests

so this is a very basic idea it was

really just how would people kind of

respond to um

sort of exploring the aftermath of a

simulation um

and people were quite interested and

they even wrote some reports sort of

hypothesizing what they think happened

to the villages that they explored

so the next year i wanted to take this a

little bit further um and what i did was

i built something called condition

unknown so if you know about the movie

the thing um which is a great movie it's

uh based on this essentially there's a

research station in the arctic there are

six crew members researching there and

one day a monster gets unleashed and

they the crew either sort of perish

fighting it or some of them might escape

and you turn up after all of this is

finished um and in the style of return

of the obradin you have to identify all

six crew members and figure out what

happened to them

um and the way that we did this is i

built a much more detailed simulation so

if everyone in the research station has

like um a routine that they go through

and this routine can be interrupted by

things so if they spot something like a

fire or they spot the monster or

something like that their routine will

change but crucially it's not just that

they'll react it's not just that we're

stimulating their behaviors

but they will do certain things that

leave traces behind for the player to

find later intentionally so in condition

unknown then the crew members radio each

other and we record every radio

conversation they have basically it's

all templated text it's very simple but

because it represents real events um it

allows the player to do kind of

detective work later on

so i think what i'm going to do here is um

i'm going to

experiment by attempting to show you this

this

let's see

hopefully you can see this let me know

if you can't um

um and

and

this has some kind of interesting

consequences that i didn't realize so

you you start out i have no idea what

this is by the way because um every

single one of these is procedurally

generated and one of the problems is

[Music]

sort of unsolvable

problems there we go okay here's our weigh-in

weigh-in

so we can see here for example that this

this section of wall has been destroyed

um and this barrel here there are

there's descriptions in the bottom left

but it doesn't work very well at this

resolution but it says that this uh

barrel has been destroyed which means

that there was probably a fire here that

exploded and we can see that like

this person is wearing a lab coat they

have burns on them that suggests they

perish in a fire so

if we look at sort of the list of the

crew we might guess that this is

probably one of the researchers and that

can be part of our final report but the

important thing about the simulation is

that um if we find a terminal here and

we read a message um

we can see here that like this crew

member victor hansen sent a message at

10 49 that they had seen something

invading the station that they said it

was outside the corridor in lab one and

none of this is

you know fake this was real so we can

know that when they sent this message

they were probably standing somewhere

where they could see lab one and the

players can do like quite sophisticated

kind of um

uh they can ensure quite sophisticated

things just by

the knowledge that this is a sort of a

real um simulation

simulation so

i really you know i got even more

encouraged by the response to that

people really um liked this even though

it was it was a jam game i mean it

didn't really it didn't function

properly um there was a lot of things

wrong with it um but the year after i

began to work on a game called stargrave

and the way that stargrave works is we

simulate a spaceship with all of its crew

crew

and something bad happens that leads to

the spaceship being abandoned so it

could just be that they kind of crash

into an asteroid and they have to

abandon ship or it could be that there

was a mutiny or something like that

um and you turn up and your only

objective is to fix the ship and then

sort of fly at home to sell it

so what we're trying to do here is um in

nothing beside remains the procedural

generation was

kind of the focus but the player didn't

really have to do anything i just said

you know explore and have fun and see

what you think

in condition unknown there was an

objective tied to the procedural

generation so you had to identify the

six crew members you had to figure out

what happened to them and the problem

with that was that that wasn't always

possible because these are simulations i

don't have any kind of drama management

there's no ai deciding you know what is

appropriate or what's solvable

and it meant that some players just

couldn't solve their particular mystery

so in stargrave i'm trying to find a

balance here where the ship is always

fixable there's always a way to kind of

get get the solution to fixing the ship

and the crew are sort of an optional

mystery you don't need to find out what

happened to the crew but you can if you

want you can find logs from them they

are obviously usually responsible for

the ship crashing and it might help you

fix the ship but it's not it's not a

compulsory part of the system

so it's interesting in that sense

because i'm trying to balance how much

involvement the player has to have with

this procedural generator

and to go back to our idea of kind of

wilderness um these games are

unconstrained simulations there's no

there's no filters on them or anything

like that so it's a lot closer to our

wilderness example than our ornamental

garden example

and that's why i'm having to be very

careful with how i designed the players

relationship with this content because

we don't want to force them to sort of

we don't want to test them on how good

this this content is because we don't

know it's very chaotic it's very wild in

the same way that minecraft doesn't like

challenge you to get from a to b because

it doesn't always know if that's

possible instead it just leaves you in

this generated world and it trusts you

to sort of explore and find your own

kind of fun

so i'm looking forward to working more

on this um but it's uh it's a it's much

harder than the previous two

and the last thing i want to talk to you about

about

is puck so um i

you are the first group of people i've

talked to about this um miriam at the

beginning said that uh i

have built a software called angelina

which is true um and the next version of

the software angelina's getting a name

change so

um angelino will be known as puck and um

my my version of park will be called

angelina probably um but i'm hoping that

puck will be something that you can

download and have your own version of

that designs games and that you can have

your own kind of

connection to and customization of

now puck is still

in sort of testing i would say there's

still quite a lot of work to do

but um

i'm taking a very different approach

with it this time and i'm really trying

to focus more on

um an ai that

you as game developers could maybe

relate to as like a co-worker or a

colleague so something that works much

slower something that is more talkative

more communicative

and in general it it works in a way that

we don't always think ai of as kind of

working it's much more focused on how it

relates to other people in its in its

creative community people like you hopefully

hopefully

so um i don't have time to tell you a

lot in depth about park but i did just

want to talk a little bit about the

background of automated game design so

these are some games that my previous

systems angelina have made so

there's arcade games in the top left

hand corner then we tried to make games

about news like platformers based around

news articles in the bottom left

angelina entered a game jam um it made

some 3d games in the top right

and then in the bottom right i began to

work on

like a version of it where you could

watch it design games and this is sort

of a screenshot of what that looked like

so it could actually stream on twitch

and you could watch it sort of build

games test rule sets uh invent levels

and then kind of ask it questions in the chat

chat

and this is something which i never

fully finished with angelina and so i've

started with puck and this is sort of

like the main focus so the idea is that

puck is something that you watch work

it's not something that kind of lives on

my laptop like it used to hopefully it

will be something which you will be able

to interact with yourself fingers crossed

so the reason why i wanted to talk about

this is because i wanted to encourage

you to

kind of get a bit silly with the

procedural generation if you want to um

so we think about procedural generation

we think about generating like levels we

think about generating maybe music um

stor a bit of text

but procedural generation can be applied

to anything and you don't need to invent

a complicated technique often to make

that happen so let me tell you a little

bit currently about how puck designs

rules for games

so this is a description of naughts and

crosses um so the players take turns

placing pieces on a 3x3 grid the first

player to get three in a row wins

and just like we did before with the

level designs in spelunky or the

buildings written on pieces of card

we can just find places to cut

in this description we can just find

things that can be replaced by other

things so if i pull out these three

segments of text and i put in something

else instead

now it says players take turns dropping

pieces onto a 7x6 grid the first player

to get four in a row wins and now we've

suddenly transformed this into connect

four with very little change really um

and i'm sure that most of you could

think of you know loads of things that

we could spot into all of these

different slots to make up um completely

different rules we could pull ideas from

board games from video games and we can

just have the system kind of randomly

mix stuff up

and in fact just by looking at sort of

very very common folk games or board

games and that one at the bottom is kind

of a joke it's from go

but lots of rules are kind of simple um

and even a simple rule space is very

interesting to explore with an ai

now the problem with um designing uh in

spaces like this is that if you look at

spelunky's level generator basically any

combination of level segments is fun um

and it takes like a fraction of a second

to generate it's one of the best things

about the generator is that it can't

really fail

um however when you come to designing

something like game rule sets what you

find very quickly is that even a few

different rule variations result in a

possibility space that is just full of

bad designs

rules are very very delicate and again

this is something which you you develop

a feel for the more generation that you

do um and so what we end up with is

something which is actually full of a

lot of wilderness and it's very hard to

sort of filter for now

now

this is sort of a problem if you work in

sort of the big games industry because

you want reliability and that's one of

the reasons why you don't hear much

about automated game design from the

games industry because it's just it's

just not there yet it's not something

that they're super interested in

but wilderness can be kind of fun um it

can throw out a lot of things that you

don't expect

it can help you explore places that

people wouldn't explore normally

a lot of bad designs can become really

interesting designs if you make a few

small tweaks and

and

i think it's also worth pointing out

that building puck and building angelina

has made me a better game designer because

because

thinking about spaces of possibilities

and spaces of outputs

helps you think about what it means to

design a game it helps you think about

what a rule set is

why is knots and crosses or tic-tac-toe

designed the way it is

why are certain rules good in almost any

game and other rules um not what does it

mean for a game to be strategically

exciting or interesting

one of the fun things about working on

this version of puck is realizing that

lots of very famous games that we play

as kids just aren't very good on paper

like they shouldn't be very good but

they are fun to play like there's

nothing really exciting about connect

four if you were to compare it you know

if you were to ask a chess grand master

to look at it they wouldn't they

wouldn't understand why you thought it

was interesting but it's really fun to

play connect four

and so by thinking about these small

tiny problems um it helps you think more

about the bigger picture it helps you

focus on things that you that you might

ignore when you're stuck in traditional

game design mindsets

so one of the ways that that we used to

sort of

filter out that wilderness um with puck

is that we get puck to play these games

this is sort of a it was the only video

i could get

because puck makes a lot of bad games so

you're seeing kind of an animation here

of a very bad game being played but you

can see the the squares are changing

color as it plays the game and you can

see its score is going up on the right

and what puck does is puck records every

single game it plays and it keeps all

the data even the games that are really

bad even the games that are obviously

bad that that you can't even place

anything on the board and that's because

understanding what makes a bad game is

as important so this wilderness is

important for puck to go through i think

and i think in the long run it makes it

a better game designer or at least

that's what i'm hoping

um and i thought i'd show you just just

one more um uh quick demo i think i've

got time um

very briefly

uh which was something which puck came

out of um very recently

and this is a two-player game but puck

mixed in some of its rules from games

like bejeweled match three games so you

take turns placing um

uh icons on a board and your objective

is to have more pieces on the board than

your opponent

and obviously if we were just going to

play back and forth with no other rules

i would win because i went first but if

you make four in a row all of your

pieces disappear um

and so you play this very delicate game

of cat and mouse with the ai where

you're both trying not to make four in a

row and you're trying to force the other

person to make four in a row and it

turns out this is a really good game but

it's so counter-intuitive that when i

first saw it i thought puck had made a

mistake because it just doesn't seem

interesting it doesn't seem good um i'll

try and uh make a um

i'll try and get like a link to this

particular game for you so you can play

it later

because just looking at it doesn't

really get it across but play a game of

it see if you can beat the ai and see

what you think um

so park isn't at the point where it's like

making amazing games yet but it's coming

up with things which are interesting

things that i hadn't thought of um and i

put that down to kind of going through

the wilderness i put that down to that because

because

most people avoid that they don't want

to go through these weird ideas um but

ai are able to explore those spaces for

us and come up with interesting alternatives

alternatives

so i think i'm at a point where i should

begin my wrap-up so i've just got a few

slides that i want to leave you with

if you go online and you read things

about procedural generation you're going

to see a lot of like gdc talks about

consistency and player experience and

how people feel about the content and

things like that

and that is important if you're

designing games it is but the one thing

that i want you to take away from today

is that building a procedural generator

is a craft it's a practice

and it can be something that you enjoy

doing even if you don't use what comes

out the other end i've built a lot of

generators that weren't used for

anything i just wanted to to have that

feeling of making something that makes

something um and i i liken it to

something like

painting or stitching or something like

that a creative art form that's just

sort of fun for the sake of it

and to stretch the gardening metaphor a

little bit further you know you don't

make you don't have a garden so that you

can grow flowers to pick and sell or

give to people

part of it is just the joy of engaging

with nature and learning about all the

different insects and flowers that are

in your garden and having that

experience about sort of just being

there while something grows you know and

then appreciating your hard work afterwards

afterwards

there's this thing in procedural

generation called the thousand bowls of

oatmeal problem

and it's used to criticize uh procedural

generators that aren't very uh good

good

so the idea is that you know

if you make a thousand bowls of oatmeal

they're all unique they're all different

but they all kind of look the same

they're very boring you know people

don't want that they want a procedural

generator that makes lots of different

types of content

but the thing i want to just leave you

with is the idea of what it would

actually take to build a machine that

made bowls of oatmeal like people spend

a lot of time perfecting their oatmeal

recipe and finding the exact balance of

ingredients that they like and i imagine

that if you built a machine to make the

perfect bowl of oatmeal you would learn

a lot about not just oatmeal but cooking

in general your own tastes

and it would be a kind of a large

project probably

and so when we think about things like

this when we think about procedural

generators it's easy to get tied up in

just thinking about the output but it's

also about the process of making and

that's kind of the thing i wanted to

give to you today it's

it's

so you know we talked about how

procedural generation is like tending to

a garden we talked about the two types

of garden total control total chaos um

this is uh someone's garden where

they've ended up getting rid of a bit of wilderness

wilderness um

um and

and

we also talked about thinking about

processes and thinking about

what you want to achieve with a

generator right so it's not about the

algorithms you use or anything like that

it's about your understanding of what

you're making and how you want to cut it

up and reshape it

and that's really the most important

thing is just playing around and practicing

practicing

and above all else i hope that you sort

of enjoy the process of making things

that make things so it's you know that

slogan for prop jam is not just about

getting in there and having the finished

result it's all about having the fun of

experimenting with generative systems

and learning something about the thing

that you're making um as you do so

okay that wasn't so bad for time um

thank you so much for listening to me um

there's more things from me on my

website possibilityspace.org including a

couple of tutorials about procedural

generation and i have little interactive

javascript like level generators that

you can click and change so you might

like that if you've liked some of the

stuff i've talked about today

and as i said at the beginning if you

want to talk to me about anything you

can tweet at me you can email me and i

also wanted to shout out both of my

amazing phd students eunice and florence

you should follow both of them on

twitter as well because they're doing

very exciting work in these sorts of

areas so if you like this talk you'll

like what they're doing as well um and

yeah that's all i've got to say i hope

it was uh i hope it was interesting and

i hope you enjoyed it and thank you so

much for having me

wow mike what an amazing talk thank you

i don't know should i should i keep that

slide up or do you want to see my face

this is recorded uh

anyway so

like i don't know how to watch look into

this camera but

uh this is recorded anyway so it's going

to be there ah the slide um

um how

how uh

do you have any questions like right off

the bat yes

is that something we should as artists

like looking through or

so i've i've not used houdini myself if

houdini is what i'm thinking of it's a

tool um there are a lot of kind of

visual effects tools and things like

that which

um allow you to sort of

make generative materials or shaders or

effects and things like that i think

they're really they're really cool tools

um and especially if you want to go into

the games industry i think you might

find it like a useful tool to pick up in

the same way that you might learn you

know photoshop or something like that um

there are if you do want to get into

this kind of thing though there are kind

of i don't know how expensive houdini is

um but there are kind of free ways that

you you can get started by looking at

sort of old-school attempts to generate

shaders for things um looking at those

noise function stuff looking at the demo

scene um i don't know if you're familiar

with this but there are a lot of people

it was it was more a thing i guess in

the 90s and the early 90s um of writing

the smallest programs you could that

made these beautiful 3d animations and

you might like that as well for the same

sort of thing um but if you can get a

copy of houdini i think it would be an

interesting thing for you to pick up

especially if you're a visual artist for sure

cool

the demo scene was huge at least in the

90s in stockholm i'm not sure like how

it is now

um well still there's still some people

going on with it yeah

yeah

Paste YouTube URL

Enter any YouTube video link to get the full transcript

Most transcripts ready in under 5 seconds

Get Our Chrome Extension

Get transcripts instantly without leaving YouTube. Install our Chrome extension for one-click access to any video's transcript directly on the watch page.

Add to Chrome — Free

Works with YouTube, Coursera, Udemy and more educational platforms

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube

←

→

↻

https://www.youtube.com/watch?v=UF8uR6Z6KLc

YoutubeToText

←

→

↻

https://youtubetotext.net/watch?v=UF8uR6Z6KLc

YouTube TranscriptPreparing your results…

YouTube Transcript:Mike Cook on PCG - invited talk - AIBU

Video Transcript

Summary

Core Theme

Paste YouTube URL

Transcript Extraction Form

Get Our Chrome Extension

Get Instant Transcripts: Just Edit the Domain in Your Address Bar!

YouTube Transcript:
Mike Cook on PCG - invited talk - AIBU