Hang tight while we fetch the video data and transcripts. This only takes a moment.
Connecting to YouTube player…
Fetching transcript data…
We’ll display the transcript, summary, and all view options as soon as everything loads.
Next steps
Loading transcript tools…
Julian Togelius on AI for Games - Invited Talk at AIBU Course | AI for Experience Design | YouTubeToText
YouTube Transcript: Julian Togelius on AI for Games - Invited Talk at AIBU Course
Skip watching entire videos - get the full transcript, search for keywords, and copy with one click.
Share:
Video Transcript
Video Summary
Summary
Core Theme
This presentation explores the symbiotic relationship between Artificial Intelligence (AI) and video games, highlighting how games serve as crucial testbeds for AI development and how AI is revolutionizing game design and content creation.
everything is there great
great
right should i start
yes take it away okay
okay
hey everyone um i don't know how many of
you are there or where you are who you
are you're in stockholm i guess
i'm julian taglias and i
came here to talk about artificial
intelligence and games i was asked to do
a relatively broad
intro not presupposing very much so here
we go
um so first of all i'm at new york
university and missing the logo i'm also
a company called model ai in copenhagen
that's not where i'm from so let me
start by introducing myself
so i'm originally from malmo um i was
about to say the beautiful city of but
that would be a lie um i finished high
school there um even though i
very nearly
failed my mathematics classes
and i decided that i wouldn't absolutely
never be any kind of engineer or have
anything to do with mathematics i wanted
to study
philosophy and psychology in learn so
that i could understand the mind um that
i did but then um to my chagrin i
realized that i had to understand
understand
i mean really philosophy psychology
didn't really get me very far i had to
build the minds to to study them so i
did a masters in sussex england in
biological inspired ai and evolutionary
robotics and then i went to essex in england
england
and i wanted to
evolve neural networks to control robots
however what happened was that i
realized that
robots are really slow and they break
down all the time and they're noisy and
dirty and stuff like this so i decided i
could do the same thing in video games
and i discovered that well you could not
only use
uh games to sort of develop ai better
you could also use ai to make games better
better
and that's sort of these two things
became kind of my career throughout the
league the postdoc with jurgen schmidt
in lugano and i was faculty in idea at
the university of copenhagen for about
five years and now um um at new york
university for six seven something years
so i do
all kinds of things in the intersection
between ai and games i care about
using video games to make artificial
intelligence better and i care about
using um um
in in in and i care about using
artificial intelligence to make games
better so basically basically both
things so i'm basically going to do a
very sort of a
quick dip into some of these topics here
and then we can do more q a afterwards in
in
any language that i speak
it's not that many of them um
um
right i mentioned artificial
intelligence a lot uh
uh
in my intro what is artificial
intelligence you may you may or may not
have taken an artificial intelligence
class i was told that i really shouldn't
suppose that you had
so this may seem still to be a bit
mysterious here in the background is hal
from um
the 2001 a space adventure or space
odyssey story from 1964 still one of the
best movies ever made
it's a very good depiction of artificial
intelligence because we don't understand
anything about it there are lots of
definitions of them
of what artificial intelligence is
but the one i like best because it's so
dry and boring is
making computers able to do things which
currently only humans can do
which is what we do in ai we choose new problems
problems
and we think that oh we don't know how
to do this this is something mysterious
only humans can do if you can do this
you must certainly be intelligent and
then we sort of work really hard on
trying to make a computer do it and it
takes a lot of time
um and
uh then we come up with a computer
program that can sort of do it and then
people look at it and assess that
oh so now you can solve the problem that
means you don't need to be intelligent
to solve it
so what do humans do with games um i'm
not sure if i can ask questions out in
the room and have people answer miriam
can people talk back to me
can i ask questions in the room and will
people yes you can
it's going to be i can actually show
them to you as well
well actually no i can't i just realized
so i just wanted something
what do humans do with games
interact with it is one answer
that's good
that's good
yeah and one answer is playing games
exactly that's usually the first i get
almost always the first i get
and then then it's interesting how far
how how kind of like diverse the answers get
get
i guess learning learning vms
learning games yes indeed
learning to play them learning to
interact with them
having fun having
having [Music]
[Music]
we also analyze games a lot
very good
okay we get a pretty good sample here um
so i usually cluster the the different
answers i get in this play them i get
lots of variations to play them study
them let's analyze here it goes into
this i mean we have all these people who
are game studies people who study games
like you bought literature or something
um and learning is a good one someone
said learning i will let's fit this into
playing now um
um
build the content form there's lots of
games where you can build content as a
user and sort of share it with others
and play it yourself have you have the
game analyze it and so on
and of course design and develop games
it's a pretty big business
weirdly enough
it's become really big in my home town
of malmo developing games and i have
nothing to do with it maybe it's because
i left i don't know um so
so
these are things that you must do with
games now ai classically has very much
focused on playing games
so let's look a little bit about this so
video games have been ai test beds and
benchmarks for a long time so
we have a couple of games that are
um that all have been used in the main
conferences in ai games as benchmarks
and test buds
starcraft um a version of super mario
bros called the mario ai benchmark
a racing game called torx
and a um
first this is unreal tournament
first-person shooter these particular
games may be less used these days the
mario benchmark is always around some
people will still work with starcraft 1
but things have moved on there's a lot
of work with starcraft 2. there's a lot
of work and new racing games and i've
seen a bunch of interesting work on
counter-strike recently um
um
a camp strike source um
so basically the idea here is that you
come up with an ai that can play any of
these games because it is a hard
sorry
so let's look at this i i i love showing
this video um
this video here i hope you can see it
moving as well we don't know how the
screen sharing works
but um
this video here is from the first mario
ai competition that we launched back in
2009 um this was because
some dude called marcus parson later
known as notch he also later made a
famous game called minecraft and then he
became weird but
but
i don't know the temporal sequence of
this but he
he
um made this
freeware version of
super mario bros and we me and my
student adapted this into an ai
competition we thought we had a great ai competition
competition
um until
we saw that
until this guy robin vangarten from
imperial college in london
submits this bot and this spot as you
can see is an amazing mario player it's
way better than any of us could ever
play this game
it's like the pixel perfect landings and
sort of the amazing sort of optimization
of everything you sort of just like cuts through
through
scratches of enemies like it was butter
or something
um so we were both really excited
because this is really great we're also
really sad um because um the
we thought we had a really good ai
competition this guy just comes along
and crushes it
and you know what makes us even more sad
it's this is 2009. do you think that
this would be some kind of super fancy
um deep neural network algorithm that's
um that um i saw the game no
you wanna know what it is
it's a star search
i don't know if you've learned about a
star in your introductory data
data structures class
otherwise you will learn about it in
in your introductory artificial
intelligence class
it is
look it's 15 lines of code
in java it's maybe 30 months of code um
it's a simple search algorithm that's
been around since the 60s literally and
what he did was that he didn't search in
physical space of the game
searched in state space and had fast
forward models to play the game and this
is kind of shocking how do you become so
good so next year
we change the competition
to um and make the level generator
generate these overhanging ledges
and you see here mario can't really sort
of get out of this everything mario
needs to go do is just move back left
and jump up on the roof of this thing
that's stopping him but the a-star
search algorithm can't really handle it
and you see the same thing again he
never searches um back left he's just
try he keeps trying to go right then you
see this
merry dance i think there's
one later on which is really fun yeah
here is like basically dancing with a
spiky guy
amazing micro totally deficient macro um
uh so this isn't of course we did this
specifically to destroy this algorithm
so the winner next year of this modern
ai competition
was this algorithm cod realm what is
agent caldwell which is a complicated
bunch of things um
i don't know if it's happening here this
a-star agent at the bottom um that
executes a rule-based system that
decides where to go and this rule-based
system is created in real time by an
evolutionary algorithm
which is
just a lot of stuff going on you see how
the mario does really clever things here
and plays in a pretty good manner um
um
so you can learn several things about
this from this one is that the very same
game could offer very very different
challenges with small tweaks such as
what levels you play or something else
it also you can also
learn that very often hybrid systems
that are not very clean um
really do the job the best
now yet another thing you can learn is
you can start thinking about are you
really solving the right problem this is
from my former pc student a postdoc nor
shocker um
she trained a um
she trained an agent based on her
five-year-old at the time niece on her
niece's playing style um and um trained
the basin network to be they do like this
this
and what we used this for was
participating in a competition that we
organized we could we organized were
technically outside it which was called
a mario ai turing test where the
whole aim was not playing the game well
it was playing the game in a human-like
manner so that people could not tell who
was a bot and who was
right
so you've seen examples here in the
super mario bros but there's a lot of
development of artificial intelligence
to play specific games um and you know you
you
you you put all your effort into a
starcraft pod or super mario spot or a
racing game bot for twerks
the problem is that you know it can't
really play
any of these bots can't really play any
of the other games
um this is a long history this is john
mccarthy back in the 60s playing chess
he was one of the
one of the fathers of artificial intelligence
intelligence
and back then the world was black and
white people wore ties you know stuff
like this computers were the size of a
room i don't know if you wanna lost this
particular game of chess
um and people kept kept working on on
chess for a very long time until this
happened this was 1997.
um ibm's deep blue computer one over
gary kasperov and this was like one of
these big events of course ibm used this
very very cleverly for their um
for their
marketing like a computer wins over the
best human in the world um the best
chess player in the world at the game
which meant for a long time was
considered like the sort of the essence
of intelligence
um and
then uh people
were like um thinking
um well maybe we just chose the wrong um
which was the wrong game maybe chess
isn't that hard maybe we should use
another port game
so there was a lot of work and go okay
right yes we can look at what happened
here um sorry what happened here um you
you would think that you know this this
computer that one over the best human in
the world that this hard game um
um
does it have some secret source that
makes it intelligent
no it's a version of mini max search
minimax search basically means that you
try all the possible um
um
moves you can make see where they take
you try possible counter moves you
assume that you want to play the best
for you
and the ad your adversary
to play what's worse for me
so one side minimizes the other maximizers
maximizers
there's lots of tricks and
stuff on this so you know deep blue was
a huge engineering achievement but the
basic algorithm behind it wasn't any
harder than this
so people kept working on this and
people kept working in go which is a
sort of the east asian cultural
equivalent of chess it's a very old um
very old game that
board game with quite simple rules and a
lot of depth
and here we had a deepmind then part of
google their alphago bot winning of
elysee doll one of the world
championship back in 2016
um and this was another one like you
know huge cultural event like wow we
thought that this would actually be a
really really hard game um
was alphago did it have some kind of
secret source to intelligence um no it
it there is a lot of cleverness in it
but the basic principle is another
research algorithm
this one's called monte carlo research
it um basically uses statistics to
figure out which kind of possible board
state to explore
next as it simulates playing it also had
a bunch of deep neural networks that
helped it with them
suggesting interesting moves to try and
helped it evaluate board positions and
so on there was a lot of cleverness into
this and a lot of engineering but the
basic principle behind it is again that
you're simulating possible games and
making these like
trees of possible moves
in memory so
so
we see this happening again and again um
um
ai researchers
accept their minds on some problem
for example playing a particular game
really really really well
and then they work hard on it and then
they win and people say that well this
isn't ai because this is just an
algorithm wow um
um
car racing um i was part of like setting
up some of the first um cars in competitions
competitions um
um in
in
driving a car as you know is not that
easy it requires motor control
and motor control
um it requires planning and adversarial
planning sort of optimized lab times and
overtake and so on um
um
this is something for a long time
sorry for all the sound on these ones so
we had um people submitting the best car
racing bots to this um competition
server and then it would drive against
each other generally they got very good
at driving alone on a track and could do
yeah that's a question of
we could do extremely well on their own but
but
they did not learn
in general good road manners or maybe
they were not really rewarded for good
road manners
here is trying to sort of you know
use the slipstream to overtake the green
car they're doing fairly well
until we get this nasty situation here
um i learned to drive on the streets of
manhattan so i i mean i can relate to this
this
this here uses something called temporal
difference learning which is a
reinforcement learning method where you
take actions and you observe the reward
you get from the environment like um
did you get punished for driving
off-road or um did you get
rewards for driving fast and then it
in
improves the agent's
actions you take right
right so
so
this has been going on for a long time
so at some point we asked ourselves can
we construct an ai that can play many
games like the same
literal agent
could do could play more than one
um because it seems like you can go on
and construct very very good um ai for
playing a particular game without
actually making progress on the general
intelligence problem so
so
we postulated this if you have an agent
that can play all the top 100 games on
app store or steam or something like this
this
would it then be actually intelligent
actually doesn't make sense to say it's
actually intelligent
we created this general video game
playing competition
back then this is like the software for
this is a bit old-fashioned this was
originally it was originally written in
python but it was too slow we moved to java
java
um and the version that um is has been
widely used is in java which a lot of
people disagree with but hey it's fast um
um
the idea here is that we write games in
what's called the video game description language
language
and then every time we run the
competition people submit their best
agents and these agents are tested on a
number of usually 10 unseen games so you
don't know which games you're going to
play when you write to the agent so the
agent will have to be general
um and the agent the description
language as a way of describing
uh say late 70s early 80s error games
this is what it looks like this is a
game that's being
written in this language
um it's somewhat python-like in its um
in syntax it says here these are the
things that exist and the properties
they have this is how the level maps to
the memory this is um what happens like
in in the things like if an avatar
collides with diamond then the diamond
disappears and the avatar gets two
scores and so on and you get a game well
this is the old sprite set that looks
something like this
so let's look at these things happening
this is a an in visual yellow version of
the overworld of the original zelda
you run around you um well it's a game
inspired by it we call it sunda you run
around you get keys open doors and you
kill monsters
as a union nice
nice
so let's imag how would an ai play this
this is um let's look at a random player
playing this game
doing random actions
it's very bad it's a random player it
dies quite immediately
now let's look at something smarter
um so we once again go back to monte
carlo research which is the um agent
that was
the core of alphago that beatles at all
that go
it's a very versatile algorithm it can
do so many things
um and you have a version of monte carlo
research playing zelda
and you see the agent doing really well
at killing these monsters
getting the key and then they roll out
so basically how far simulates doesn't
really see the door
because there's a lot of randomness
involved but then okay and at some point
it actually sees the door and and runs there
there now
now
here is the game called boulder dash
this is very closely modeled on the
actual game called boulders which some
of you may or may not have played is an
80s classic
unlike many 80s games it really still
holds up
you have um you need to dig away the
dirt to get the diamonds
you need to avoid getting smashed in the
head with a boulder you need to also
there's a puzzle element you need to try to
to
not sort of block yourself in by the boulders
boulders and
and
because it's very easy to do that
so um and you have these monsters here
on the side and the monsters will kill
you if you um
uh reach them
there's something that looks like a bat
in a cave here to the left and we all
know it's 20 21 we all know don't
disturb the bats in the caves that's um um
um
it's bad for you bad for the world
and then when you have ten dime once you
go to the to the exit
now here's a random player playing polar
dash will he do well
he takes random actions
he does not do well
he succumbed to the classic stone in the head
head
um here is the monte carlo research
agent doing this
and it's a very sophisticated nature to
see how he gets a few diamonds
um and um
how he behaves like an idiot and gets
killed this is a very hard game um
um
so basically you see um
you see um uh
you see this agent that actually plays
many of these games fairly well um this
particular agent can win
approximately half of the
160 games in a framework um but there's
still some some that are very very hard
and this just basically goes to show
that sure you could make an agent that
plays bolder specifically
it might not even be that hard the
problem is um one that plays baldash and
all the other games we have in there
versions of uh space invaders and mario
and all kinds of things
and zelda and stuff and that's really
really really really hard so the general
game playing
problem is really hard
okay now on to something completely
different actually
closely related but still
so i don't know if any one of you dream
dreams of being a game developer um i'm
trying i'm not trying to dissuade you
but modern game game productions have
grown to some
huge size because there's just so many
things that need to be done and
programming isn't the largest post
there's so much content they didn't
need to be created however
however
there is um for a long time there's been um
um
this uh
movement trend whatever we call it of
people using procedural content
generation games so here
procedural content generation means that
some part of the game world is generated
either during development time or during runtime
runtime
um and you see here some of the um uh
some of these sort of you know um
pioneers you have elite here in the
background um and uh spelunky
and we're gonna mention some of these as
we go on and civilization which are all
based on like some part of the world is
being generated as you play the game and
this is like the core of the game in a sense
so elite here
i used to play elite on my i mean i got
my first commodore 64 when it was 11.
this was 1990 so the machine was already
outdated um
but i used to play elite in this machine
um elite is a game where you um
fly around the space there's 3d
renderings of things these graphics
might not might be from the amiga
version i'm not sure or it's from the
commodore version one of them um and
there's like um
other spaceships some are just
freighters some are friendly some are
the police some are pirates stuff like this
this
um you have lots of stats in your
spaceship um there's a huge
universe with different star system each
star system has planets and the planets
have space stations in some cases around
them um and they're the the star chart
is enormous as you can see 4096
different um
star systems if you remember right
um and each of them has space stations
and you can land at the space station
space station there's different prices
for commodities you can buy and sell and
you can trade them to different to other
space stations
um and so a lot of what happens is
trading but there's also like lots of
missions in there you need to fly and um
fly and sort of um
or you can choose to basically carry out
various missions it's enormous there's
so much going on um
and this fits in memory in a commodore
64. um a commodore 64
which you may or may not know had 64
kilobytes of memory so not gigabytes not
megabytes kilobytes of memory 64 000
characters fits in the memory of of this machine
machine
these days you can easily buy a computer
that has a million times as much memory
which is pretty insane you think about it
it
how can this possibly fit how can you
fit 4096 star systems in 64 kilobytes it
seems impossible well
well
the way it works is procedural
generation every time
basically you don't save the star system
you save a number and every time you get
to the star system and you can look at
that particular number use it as a seed
for a number of random number generators which
which
use a number of algorithms that builds
the star system up every time you get there
there
rogue is another one brook was created
in 1980 by
michael toy and glenn wickman at the
university of california santa cruz they
wanted to play dungeons and dragons on
their computer
um however they didn't want to create
the adventures themselves because that
would take a lot of time and they wanted
to and they didn't want and they wanted
to be surprised
they also didn't have the disk space to
save the adventures so they created this
game where with the state-of-the-art
graphics here like this thing here this
is this is you um this is smiley um you
move around and explore these rooms
and you'd encounter dragons and
different monsters you can get fine
treasures you can find the various
weapons and things with status effects
magic potions and stuff like this
in every time the thing is generated
every time you start the game so no two
players are the same this of course
started a whole genre of roguelikes
which um
are ever popular and more popular than
ever and i spent a lot of the spring and
summer playing hades for example which
is sort of a light to work like a rogue
light but this idea of like
the world is regenerated every time you
play it
diablo 3 is another example of roguelike
spelunky roguelike platformer
made a huge sort of um effect on the
indie gaming community back in 2008.
civilization games they also like in a
sense roguelikes
you need
in according to the whole game series
that you need to um
um you need to have a new game
a new sort of world to explore every
time you start playing
no man's sky which some of you may have played
played
some of you may have heard of probably
everyone is hugely divisive it is a
fantastic art piece um maybe not a
fantastic game but it is basically a
lead for the 21st century with um
every every planet has its own flora and
fauma and
and things like this it sort of lacks in
overall narrative structure but it has
um it's an amazing world generator in
there and it has more planets than that
than you could ever um then you could
visit in a lifetime
um the same idea everything is
you serve you store certain random seeds
and when you actually encounter that
sort of planet the planet is recreated
from those random seas so
so
i've been concerned a lot with how could
you bring procedural content generation
further could you um
cut game development time
oh i see someone in chat saying you've
seen a redemption now um uh in no man's
sky yes um
uh i i i want to actually restart it um
that's uh that's something i want to do
and sort of you know see what it's like
now in 2021
anyway so procedural gun generation
could you sort of help help creating
games making it possible to create games
with smaller um
um
smaller sort of team sizes because you
can just wave your hands and equate
things can you create games that adapt
the game worlds to the preference of the
player you can create endless games that
you never want to stop playing and that
actually meaningfully sort of you know
keep keeping different
not like the infinite words of minecraft
or something
um could you even circumvent or augment
the limits of for human creativity and
create new types of games
and another one for people who study
games is the very relevant one could you
understand game design better through
formalizing the design process so if
you know that you won't don't really
understand sorting until you've written
a sorting algorithm
so do you really understand game design
before you've written a game design algorithm
algorithm
um this is a controversial algorithm
once i repeat said this argument in a
little bit longer form in front of like
a few hundred
game studies
kind of people and
they were angry at me i think it's
fascinating um but
but but you know enlightening so
so
if we look at super mario bros that
we've used before and we looked at it before
before
and i see i need to speed up because i
need to be done in 10 minutes at the most
most um
um
um you could
generate levels in lots of different
ways one thing i've worked a lot with is
this thing called evolutionary computation
computation
where you basically mimic
artificial evolution in an algorithm so
you basically have lots of different
solutions in this case lots of different
pseudomore reverse levels and you make
mutations and crossover of these and
er
and keep the good ones discard the bad
ones in the end you get better and
better you climb the fitness landscape
so here is
some work i did watch my piece with my
my phd student back in malmo steve
dodskok miriam was actually on his phd
committee um we divided up the um um
the sort of
levels into number of different slices
or columns as we call them here here is
like slice one slice two sli no slice
one slice one again slice two slice one
again and here's slice three so you sort of
of
can can turn a level into a string
and then we evaluated this the fitness
function was how many different
design patterns are in these levels
and then you can
pretty quickly learn to generate a whole
bunch of different levels
levels
this way and you get like this infinite
level generator is really fast and it has
has
a lot of local patterns maybe not so
much in terms of global patterns but you
can control that as well
this is this idea has also been used by
others so um
this is some work by cameron brown
who is at maastricht now he was at
university queensland before and
and he
he
he generated completely new board games
so hit his population including the
rules of different board games it's like
um things like go gomoku
shakur's chinese chess
four in a row and so on we select
parents from these
cross over mutate check the rules if the
game is not well formed so for example
there's no win condition you throw it
away give it a name he had this like
name generator which was lord of rings
inspired check if it
is too slow to play if that you throw it
away otherwise you
optimize a game playing agent to play it
it's based on minimax but you evolve in
your networks if it always leaves a draw
you throw it away otherwise if it's
inbred too similar in house place to
other games it's um um you throw it away
otherwise you evaluate it in the
variation function there's a lot of
different things that are tested for
example how late in this game can you
tell who's going to win is one of the
and then it's put back in the population
and he ran this for a long time
and one of the things that came out was
his game jabalath with when you basically
basically
need to this is the description of the
avalanche you win if you have four in a
row but you lose if you have three in a
row so it's a complicated version of
four of four and a row essentially
um and this is what it looks like let me
see you can buy it boxed
cameron has actually made some money
and it says here that the avalanche is
not designed by ludi luda is the system
that not designed by cameron brown but
it's designed by cameron brown's ludi
ludi is a system he he built that
designed the game
so cameron brown did not design this
game cameron brown designed the software
that designed this game however
however
cameron is selling it cameron is getting
all the money and his software is not
discuss well
well
it's it's fun i think it's really fun
that you see um
this is a complete game that's
completely designed by an evolutionary algorithm
algorithm now
now
i'm almost done here just gonna show a
few more little things here um
there's a lot of work in designing
levels but you can also design levels
how can you collaborate with a
content generator
so this here is a game that was popular
maybe seven years ago or so on mobile
called cut the rope i guess it's still
out there you basically it's a
physics-based puzzle game where you have
to feed this little
frog monster candy by
pushing um the candy around cutting
ropes when you need to and so on
and here's the system that
we built called the raposa possum
where you which can generate new levels
for this game um
um
you press a button and it generates new
level it's based on grammatical
evolutions and evolutionary algorithm
that interacts with the grammar system
um you can also design it yourself um
you can sort of put things out here as
you would with any kind of wig
designer um here's cushions that the
candy will bounce off and then you can
put like an air cushion here and
and
a bubble that will catch the
the candy and so on and it's a rocket
and then you can check you can ask it
please system can you play it for me and
it tries to play it and it tells you you
see here as it plays around it has all
the different sort of um actions you
need to take up here
and it can help you analyze what you're
building by playing it as you sort of um
you see what's possible was not possible
it can tell you if it's
not possible you can view the actions of
what it takes this is basically what the
grammar based system produces um
and you then you can tell it that okay i
want to keep some of these things and i
want you the system to redesign rest so
here we
redecide the rest of the level it turns
out that this is a very boring level
because it basically you just cut all of
the ropes at the same time and you win
so you do another one um
and um
and you have it to optimize the design
and it gives you a pro sort of
surprising one how would you play this level
level
well it turns out that
if you um
um
if you load the rope then load the candy
onto the rocket and cut it it work at
the right time um the um
the candy will actually bounce off this
bouncy thing
so it basically the system
shows you on
um it shows you sort of you know
surprising solutions to the um
to the levels you're designing
among other
things um
last thing i want to tell you i have no
time to go into this there's also a lot
of ai work in modeling what players do
and what players feel so this is some
work that was done for a long time ago
and there's been lots of follow-up to this
this
where we tried to
this is the same good old mario ai framework
framework
we tried to um
teach neural networks
we had people play this game play
different levels and see which of these
two levels was most fun which was the
most frustrating which was most
challenging and then we could model how
a particular playing style
would think about a particular level and
then you can optimize you can search the
neural network
for levels that are maximally
frustrating and minimally fun
no no the other way around
maximally fun minimum frustrating
i'm probably way over time
so here here's what
how i think it all relates video games
are perfect aspects for ai they're very
cheap better than robots that is time to
challenge the mind i didn't even get
into that argument why games are great
for challenging minds
general video game playing is important
playing not one game but a lot of times
um but also ai is the future of game
design so basically you can use
protesting games
generating content together with humans
or on his own generating tutorials um
and so on
so um you already i wrote this popular
science book which might be a fun read
for some people and i think miriam has
already told you about this artificial
intelligence games book which is nicely
available online we also run a summer
school every summer
which is connected to this book
thank you so much julian what an expose uh
uh
thank you so much and let's uh give a
big hand to [Applause]
[Applause]
julian i'm going to
i'm going to stop the recording so that
i can turn the camera around so that you
can see the audience and and and they
can ask you questions so
uh with that i'm now stopping the
recording here um and
so to you who are looking a later
recording goodbye for now
Click on any text or timestamp to jump to that moment in the video
Share:
Most transcripts ready in under 5 seconds
One-Click Copy125+ LanguagesSearch ContentJump to Timestamps
Paste YouTube URL
Enter any YouTube video link to get the full transcript
Transcript Extraction Form
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.
