Real NPCs
Philip Jones
I've been thinking about the problem of putting non-player characters into
games for quite a while now and I've realized that we have a problem that
can't be solved by AI, super-clever or detailed programming tricks etc.
Players just don't treat NPCs like real people. How often do you go up to
a stranger and pester them with questions? (Unless you work in market research
or something.) Yet this is the only way to get on in a piece of IF. But there
is no natural human interaction analogous to this. We mostly meet new people ,
by being introduced, or in a very structured environment (at work, parties,
college etc.) I've just been on the welcoming committee for new doctorate
students at my college and have been trying to list the conversations I've
been having. They tend to go,
Me: Hi, are you a new DPhil student.
Them : No, I'm an MSc, Undergrad, lecturer who's been here 20 years ... [or]
Yes, I'm working with X [or] on Y.
Me : (Desperately improvising) oh is that on Y [or] with X etc.
Them : yes
Me : where were you before, that's interesting, etc. etc.
Now what I notice is that, not only is the conversation structure fairly
cliche'd, it is all about the person's history, interests and stuff. ie. lots
of depth questions about them.
What we don't tend to say is.
Me : Hi etc.
Them : Yeah, I'm X. Can you tell me the best way to find the old haunted
mansion / the pirate's treasure chest ...
Sometimes, you *can* meet someone by answering technical questions
Newnbie : excuse me, can you tell me how to get e-mail on this system?
Me : sure, uh, if you type elm here ...
A welcoming stranger is going to ask you questions, I have yet to see an NPC
do that. (Except rhetorical ones.) If they did we wouldn't be able to poke
them and expose the machinery so much.
philip
Greg Ewing
In article <38bv5h$s...@infa.central.susx.ac.uk>, phi...@cogs.susx.ac.uk (Philip Jones) writes:
|>
|> Me: Hi, are you a new DPhil student.
|> Them : No, I'm an MSc, Undergrad, lecturer who's been here 20 years ... [or]
How about this:
A rotund gentleman with a bushy beard and friendly face
and a large pack on his back hikes up to you and extends
his hand. "Hello, I'm Rufus Rucksack. Welcome to the valley!
What's your name, by the way?"
> fred fencepost
"Glad to meet you, Fred. You look like the outdoors sort -
here for a spot of adventuring, are you?"
> yes
"How are you finding it so far? Enjoying yourself? Got
stuck on anything yet?"
> there's a rock blocking the path
"Damn, that must have come down in that storm a couple
days back. I always bring a crowbar with me - often
comes in handy for things like that. Got anything like
that by any chance?"
> no
"Hmmm... There's an old tool shed half way up the
hill behind the town, there..." He thumbs over his
shoulder. "I haven't been up there for yonks, but
there might be something there you could use.
Well, I can't stand here flapping my jaw all day.
If you're gonna be around these parts a while
we'll probably bump into each other again...
Meantime, best of luck..."
He waves goodbye and hikes away into the distance.
No fancy AI required - just a lot of canned monologue and
recognising some keywords in the responses.
Greg Ewing, Computer Science Dept, +--------------------------------------+
University of Canterbury, | A citizen of NewZealandCorp, a |
Christchurch, New Zealand | wholly-owned subsidiary of Japan Inc.|
gr...@cosc.canterbury.ac.nz +--------------------------------------+
David Baggett
Greg Ewing <gr...@cosc.canterbury.ac.nz> wrote:
> "How are you finding it so far? Enjoying yourself? Got
> stuck on anything yet?"
>
> > there's a rock blocking the path
>No fancy AI required - just a lot of canned monologue and
>recognising some keywords in the responses.
But what if the player types
>there's this rock, and it just so happens that I can't
get past the dang thing.
If you just look for keywords, you'll get into trouble:
>"Don't rock the boat," I always say. No trouble yet!
is going to match "rock" and completely confuse the game.
Without syntactic parsing, this is really unavoidable. And with it, you
need some decent knowledge representation; otherwise you might as well just
have the kind of parsing we have in IF games now.
Dave Baggett
__
d...@ai.mit.edu MIT AI Lab He who has the highest Kibo # when he dies wins.
ADVENTIONS: We make Kuul text adventures! Email for a catalog of releases.
Greg Ewing
In article <38ubv4...@life.ai.mit.edu>, d...@case.ai.mit.edu (David Baggett) writes:
|>
|> >"Don't rock the boat," I always say. No trouble yet!
|>
|> is going to match "rock" and completely confuse the game.
Yes, you're probably right - it would be safer to stick
to the traditional ASK RUFUS ABOUT ROCK, etc.
The main point I was trying to make - I think - is that
if you can keep the player busy responding to what the
NPC is talking about, he won't get a chance to probe
too deeply and expose the shallowness of the NPC's
personality.
|> Dave Baggett
|> d...@ai.mit.edu MIT AI Lab He who has the highest Kibo # when he dies wins.
Greg Ewing, Computer Science Dept, +--------------------------------------+
David Baggett
Greg Ewing <gr...@cosc.canterbury.ac.nz> wrote:
>The main point I was trying to make - I think - is that if you can keep the
>player busy responding to what the NPC is talking about, he won't get a
>chance to probe too deeply and expose the shallowness of the NPC's
>personality.
I know I must sound incredibly surly and negative to readers of this group,
but these kinds of claims (which are certainly interesting and tempting)
vastly underestimate human intelligence. The reader will figure out that
your NPC is made of cardboard in *minutes* no matter how much smoke you put
between him and the NPC. The Ultima solution (canned player responses)
just makes the little game (which has little to do with person-to-person
communication) slightly more interesting. It's still painfully obvious
that you're talking to a computer.
It's a pity that so many of these problems are currently intractable, and
don't appear to be in solving range any time soon (many decades,
centuries). But that being the case, we really need to exploit what IF can
be good at, not try to hide its shortcomings in vain.
Eliza is a good example -- people see through this cleverly designed
program (with cleverly chosen domain -- the already nebulous scope of
psychiatric advice) after only a few responses.
Maybe there *are* some simple things you can do to fool the reader into
thinking your characters are real for a good long time. But I doubt it.
:)
Dave Baggett
__
d...@ai.mit.edu MIT AI Lab He who has the highest Kibo # when he dies wins.
Gareth Rees
> Eliza is a good example -- people see through this cleverly designed
> program (with cleverly chosen domain -- the already nebulous scope of
> psychiatric advice) after only a few responses.
That's not true for everyone. We - that is, people with backgrounds in
computer science, who have probably studied AI at some point - are wise
to ELIZA and similar programs now. That's hasn't always been true for
everyone. Weizenbaum writes (in "Computer Power and Human Reason"):
"People who knew very well that they were conversing with a machine
soon forgot that fact ... This illusion was especially strong and
most tenaciously clung to among people who knew little or nothing
about computers. They would often demand to be permitted to
converse with the system in private, and would, after conversing
with it for a time, insist, in spite of my explanations, that the
machine really understood them."
Weizenbaum compares conversing with ELIZA to attending a theatre play
and entering into a suspension of disbelief. Exactly the same thing
happens, or ought to happen, to us when we play adventure games, and the
task of the game designer is to make this suspension easy.
In the case of NPCs, this amounts to a set of tricks by which the game
designer suggests there's more to the NPC than meets the eye:
1 Make the NPC get bored and wander away after a few interactions, so
that the player doesn't have much time to discover his limitations.
2 Lower the player's expectations - make the NPC a dog or a robot
(Floyd in Planetfall), or deaf, or not speak English, or stupid
(Trent/Tiffany in LGOP).
3 Extensively playtest to discover the kinds of things players type at
the NPC and code up special responses to these.
4 Give the NPC a strong motivation that accounts for his not
responding very much to the player.
5 Give the NPC several roles in the story so that the player can't
dismiss him as e.g. "the guard who blocks this door here".
Yes, the player will see through these tricks very quickly (unless you
work very hard on #3 and are very lucky!), but even after they've been
"seen through" the tricks will help him or her to suspend disbelief.
N.B. ELIZA had a different set of tricks (pretending to be a
psychologist, reflecting user's input, always asking questions etc) that
won't transfer easily into interactive fiction. I don't the ELIZA model
for repsonding to player input will work effectively enough to be worth
implenting in IF. (Prove me wrong!)
--
Gareth Rees
Phil Goetz
Gerry Kevin Wilson <whiz...@uclink.berkeley.edu> wrote:
> Here's a non-text adventure related thing, for all you graphics
>gurus out there. I'm interested in certain aspects of this, and I was
>wondering, has anyone considered writing a program to simulate the human
>face? Just a small boxes' worth, in the middle of the screen, but a
>total facial simulation. Could be worth a thousand digitized scenes...
Lots of work has gone into this. Some Japanese corporation has
developed "Neurobaby", for one. I expect the Oz people at Carnegie
Mellon could tell you more. I have forgotten other work, but am
fairly confident that there are several high-budget teams working on
this.
Phil go...@cs.buffalo.edu
William S. Reilly
Neurobaby is a bit abstract as a face, but if you'd like to look into
it, check out the SIGGRAPH Visual Proceeding for 1993. The same proceedings
have info on "KA-O-RI" from Hitachi that looks interesting, but that I
don't know much about. Unfortunately, these proceedings do not have
references to other work, but they do contain contact information that
could be useful.
Other work to look at includes:
Katashi Nagao and Akikazu Takeuchi, "Social Interaction: Multimodal
Conversation with Social Agents", Proceedings of AAAI-94. This paper
is more about the social interaction than the face itself, but it
refernces work you might find more useful. This work is out of the
SONY Computer Science Lab.
Also, I don't have any references handy, but the work of Demetri
Terzopoulus at the University of Toronto is quite good.
As Phil suggested, there's plenty more, but this is what comes to mind
immediately. I hope it's of some use.
Scott
Gerry Kevin Wilson
Well Dave, perhaps the reason that IF NPCs are seen as dull, lifeless,
and cardboard-like, is simply because they are. But then, you knew
that. So, a couple of strategies to bring them to life? (Wot, more?)
Think of the novel character. What is the key difference, the
'generation gap' between it and and IF character. Interaction, of
course. Now, as you say, there's only so much smoke you can blow to
cover that up.
HOWEVER, that's not where most IF NPCs are lacking. Assuredly, it's
ONE AREA they're lacking in, but it's not the only one. There's a lot to
be said for ordinary old literature techniques. How does a novel bring a
character to life? Through reactions to the main character? Perhaps in
a poorly written novel, or mainstream mystery novel, but that often
doesn't cut it. The NPC's ACTIONS are just as important as its REACTIONS.
Splitting hairs yet? Perhaps. BUT, what does everyone remember
about Floyd? That damn hucka-bucka beanstalk, most likely. That was an
idiosyncracy(sp?) of Floyd's personality. He liked that game. Why is
this important? Because it was something that originated with Floyd, and
not the player. Not the best example, I know. Still, think back on the
few scraggly NPCs you've seen. How many have pasts, lives,
personalities, or even the most basic thing that had nothing to do with
the game's goal? Not a helluva lot, I'll bet. I myself thought Lloyd
the Insurance Robot, from Ditch Day Drifter, was pretty neat. He acted
appropriately in response to his environment. He wasn't much on
interaction, but he DID things. Hmm, well anyways, we'll see what comes
of this theory once Avalon is done. The proof is in the pudding, as they
say.
Here's a non-text adventure related thing, for all you graphics
gurus out there. I'm interested in certain aspects of this, and I was
wondering, has anyone considered writing a program to simulate the human
face? Just a small boxes' worth, in the middle of the screen, but a
total facial simulation. Could be worth a thousand digitized scenes...
--
<~~~~~~~S~W~~~~~~~~~~~~~~~AVALON~~~~~~~~~~DUE~EARLY~1995~~~~~~~~~~~~~~|~~~~~~~>
< ERT O A In the midst of the Vietnam War, one man dies, and is | ~~\ >
< V IGO F R charged with a quest from King Arthur. Live the quest! | /~\ | >
<_______T_E_____________...@uclink.berkeley.edu__|_\__/__>
Philip Jones
David Baggett <d...@ai.mit.edu> wrote:
>
>I know I must sound incredibly surly and negative to readers of this group,
>but these kinds of claims (which are certainly interesting and tempting)
>vastly underestimate human intelligence. The reader will figure out that
>your NPC is made of cardboard in *minutes* no matter how much smoke you put
>between him and the NPC.
Like I keep saying. The problem is that players don't treat NPCs with the due
respect they would treat other humans. How does a real human react to the kind
of questioning that we regularly dish out to NPCs? Once we know this, then at
least we'll have a better idea as to how NPCs should behave during those first
minutes.
>
>Eliza is a good example -- people see through this cleverly designed
>program (with cleverly chosen domain -- the already nebulous scope of
>psychiatric advice) after only a few responses.
But Eliza is a hell of a lot better than any NPC I've ever come across in IF.
I would really like to see what would happen if you put an Eliza in a game
context (may be with some information to get out if you got her to like /
trust you.)
philip
Phil Goetz
Philip Jones <phi...@cogs.susx.ac.uk> wrote:
>But Eliza is a hell of a lot better than any NPC I've ever come across in IF.
>I would really like to see what would happen if you put an Eliza in a game
>context (may be with some information to get out if you got her to like /
>trust you.)
I put a bunch of Elizas in my game _Inmate_, each with a different
conversation database, and some facts to impart. Plus, I let matches
to patterns invoke actions.
It is playable on a Unix system, using Rick Skrenta's Apple // emulator,
but a little slow, and compiling the emulator is as hard as compiling itf.
Well, maybe not, since I never was able to compile itf.
Phil go...@cs.buffalo.edu
David Baggett
>HOWEVER, [interaction is] not where most IF NPCs are lacking. Assuredly,
>it's ONE AREA they're lacking in, but it's not the only one. There's a lot
>to be said for ordinary old literature techniques.
Still, if you make your NPC a very good character it's going to build up
the player's expectations, so that when he interacts with the character and
discovers that he's actually dealing with a finite state machine he'll be
all the more put off. These two aspects of character in IF are
inseparable.
The examples you cited (Floyd, Lloyd) are interesting, because these
characters are in a sense very specially designed. Since they are robots,
the player does not expect them to respond like people. So their behavior
can be reasonably sophisticated without creating an expectation of
person-to-person interaction.
William S. Reilly
|> Gerry Kevin Wilson (whiz...@uclink.berkeley.edu) wrote:
|> : Here's a non-text adventure related thing, for all you graphics
|> : wondering, has anyone considered writing a program to simulate the human
|> : face? Just a small boxes' worth, in the middle of the screen, but a
|> : total facial simulation. Could be worth a thousand digitized scenes...
|>
|> whole issue deals with "simulated agents." Anyway, one of the articles
|> in there (near the back, I'm sorry I can't remember which one) detailed
|> a system someone came up with to reflect the state of a "helper" agent.
|>
|> The agent was represented by a box like you described, inside of which
|> was a cartoonish facial expression. The expression could be bored (when
|> you're not telling it to do anything), happy (when it comes up with a
|> solution to the task you give it), perplexed (when it requires more infor-
|> mation from you), as well as a few others.
That is the work of Pattie Maes at MIT. As Derek points out, that is a
cartoon face, but something along those lines could be effective. In fact,
it seems that people are so well tuned for facial recognition, that unless
a "realistic" face is pretty much perfect, it can be very disconcerting.
Obviously, cartoon faces do not have that problem. Another cartoon face
that you could look at is Kristinn Thorisson's work on J. Jr, also out of
the MIT Media Lab.
|> I don't have the source in front of me, but the entire issue is pretty
|> interesting. The articles don't focus on the IF potential of agency, but
|> it isn't hard to see where some of the ideas could be adapted to create
|> better NPCs.
One paper in CACM that does deal with interactive drama, but using examples
from animation, is Joe Bates's paper on believable agents. Much of the
discussion is likely to be of interest to IF developers who want to build
believable NPCs. The paper is called "The Role of Emotion in Believable
Agents" and is also available via ftp or the WWW.
FTP: ftp.cs.cmu.edu:/afs/cs.cmu.edu/project/oz/ftp/papers/ba-and-emotion.ps
URL: http://www.cs.cmu.edu:8001/afs/cs.cmu.edu/project/oz/web/papers.html
Scott