1
00:00:00,000 --> 00:00:04,980


2
00:00:04,980 --> 00:00:08,010
So why don't we start out
just by introducing yourself.

3
00:00:08,010 --> 00:00:09,330
OK.

4
00:00:09,330 --> 00:00:10,920
So my name's Bill Swartout.

5
00:00:10,920 --> 00:00:17,190
And I'm now at USC, University
of Southern California.

6
00:00:17,190 --> 00:00:20,740
I got my undergraduate
degree at Stanford,

7
00:00:20,740 --> 00:00:22,980
which is where I actually
first got involved

8
00:00:22,980 --> 00:00:24,870
with artificial intelligence.

9
00:00:24,870 --> 00:00:34,200
This was-- I graduated in 1974
and worked with Cordell Green

10
00:00:34,200 --> 00:00:36,900
out there, who was
then on the faculty,

11
00:00:36,900 --> 00:00:40,770
and worked on the problem of--

12
00:00:40,770 --> 00:00:43,170
he was interested in automatic
programming, basically.

13
00:00:43,170 --> 00:00:45,070
Could you get it
so that, you know,

14
00:00:45,070 --> 00:00:48,720
computer from some
nonprogrammatic description

15
00:00:48,720 --> 00:00:50,940
like input-output pairs
or something like that

16
00:00:50,940 --> 00:00:54,870
could actually infer
what the program would

17
00:00:54,870 --> 00:00:58,950
be that would generate
that output from that input

18
00:00:58,950 --> 00:01:00,840
and generalize it.

19
00:01:00,840 --> 00:01:04,290
And so I was an undergraduate.

20
00:01:04,290 --> 00:01:08,280
And there was a program
that had been written

21
00:01:08,280 --> 00:01:15,990
by a graduate student that
actually did infer list

22
00:01:15,990 --> 00:01:19,050
programs from example
input-output pairs.

23
00:01:19,050 --> 00:01:22,580
And it worked on
exactly one case.

24
00:01:22,580 --> 00:01:25,140
And so my job--

25
00:01:25,140 --> 00:01:28,500
I took this course, and
you had to have a project.

26
00:01:28,500 --> 00:01:30,630
And, of course, my
job was basically

27
00:01:30,630 --> 00:01:33,030
to get it to work on
the other 15 examples

28
00:01:33,030 --> 00:01:38,380
that it was supposedly
originally designed to work on.

29
00:01:38,380 --> 00:01:42,000
And so this turned
out to be, actually,

30
00:01:42,000 --> 00:01:45,570
quite a intense
debugging effort.

31
00:01:45,570 --> 00:01:50,280
And I remember I finally
got it to work on about 12

32
00:01:50,280 --> 00:01:52,420
or 13 of the cases it--

33
00:01:52,420 --> 00:01:52,920
Impressive.

34
00:01:52,920 --> 00:01:54,450
--it was supposed to work on.

35
00:01:54,450 --> 00:01:57,400
And about 4 o'clock
in the morning

36
00:01:57,400 --> 00:02:00,750
I had an experience where
I realized, you know,

37
00:02:00,750 --> 00:02:03,840
how your memory can be
impaired as you get sleepier

38
00:02:03,840 --> 00:02:05,760
and sleepier and sleepier.

39
00:02:05,760 --> 00:02:09,160
And what was happening was I
was working on the final thing

40
00:02:09,160 --> 00:02:11,160
that I was going to do,
the final example that I

41
00:02:11,160 --> 00:02:12,180
was going to do.

42
00:02:12,180 --> 00:02:15,180
And there was some bug that
was causing it not to work.

43
00:02:15,180 --> 00:02:17,050
And I figured it out.

44
00:02:17,050 --> 00:02:19,260
But by the time I
had figured it out,

45
00:02:19,260 --> 00:02:20,850
I realized that my memory--

46
00:02:20,850 --> 00:02:23,430
I had lost all of
the memory traces

47
00:02:23,430 --> 00:02:25,680
back to where I was
starting from on this.

48
00:02:25,680 --> 00:02:27,990
So I said, OK,
time to go to bed.

49
00:02:27,990 --> 00:02:32,040
It'll be fine if it
works on 13 cases.

50
00:02:32,040 --> 00:02:35,460
So that was my sort
of introduction to,

51
00:02:35,460 --> 00:02:39,060
you know, working in
a small way, at least,

52
00:02:39,060 --> 00:02:41,160
to working on AI.

53
00:02:41,160 --> 00:02:46,620
And later Cordell and I
and a graduate student

54
00:02:46,620 --> 00:02:49,930
who was involved published a
paper, my first paper in an AI

55
00:02:49,930 --> 00:02:51,738
conference on that work.

56
00:02:51,738 --> 00:02:52,780
And that was kind of fun.

57
00:02:52,780 --> 00:02:53,230
As an undergrad.

58
00:02:53,230 --> 00:02:53,730
Not bad.

59
00:02:53,730 --> 00:02:54,810
Yeah, It was kind of fun.

60
00:02:54,810 --> 00:03:00,100
And then I went
to MIT and worked

61
00:03:00,100 --> 00:03:03,100
with Peter Solovich in the--

62
00:03:03,100 --> 00:03:06,040
well, initially I
worked with Bill Martin

63
00:03:06,040 --> 00:03:08,590
and then later
with Peter Solovich

64
00:03:08,590 --> 00:03:11,410
on the clinical
decision-making group there,

65
00:03:11,410 --> 00:03:18,220
which was basically concerned
with trying to help physicians

66
00:03:18,220 --> 00:03:20,350
by writing programs
that could give

67
00:03:20,350 --> 00:03:25,660
advice about diagnoses and
therapy recommendations

68
00:03:25,660 --> 00:03:26,410
and all of that.

69
00:03:26,410 --> 00:03:27,700
He was [INAUDIBLE].

70
00:03:27,700 --> 00:03:28,320
Oh really?

71
00:03:28,320 --> 00:03:28,900
OK.

72
00:03:28,900 --> 00:03:33,370
So the program that I worked
on was the Digitalis Therapy

73
00:03:33,370 --> 00:03:35,290
Advisor, which was--

74
00:03:35,290 --> 00:03:41,830
Digitalis is a drug that's
actually still given today.

75
00:03:41,830 --> 00:03:47,580
It was first
prescribed or described

76
00:03:47,580 --> 00:03:50,790
in the medical literature
in like the 1700s.

77
00:03:50,790 --> 00:03:52,230
So it goes way back.

78
00:03:52,230 --> 00:03:55,200
It's derived from
foxglove plant.

79
00:03:55,200 --> 00:03:57,510
And it's a drug though
that's very difficult

80
00:03:57,510 --> 00:04:00,300
to give properly
because the difference

81
00:04:00,300 --> 00:04:02,952
between a toxic dose--

82
00:04:02,952 --> 00:04:04,410
one that's going
to make you sick--

83
00:04:04,410 --> 00:04:05,993
and one that's going
to be therapeutic

84
00:04:05,993 --> 00:04:08,970
is only about a factor of two
compared with, say, something

85
00:04:08,970 --> 00:04:13,830
like aspirin where the error
range is a factor of about 100.

86
00:04:13,830 --> 00:04:19,529
And in fact, the variability
on sensitivity between patients

87
00:04:19,529 --> 00:04:23,140
is bigger than the
therapeutic window.

88
00:04:23,140 --> 00:04:26,040
So you have to, essentially,
customize the dose

89
00:04:26,040 --> 00:04:28,560
for each particular
patient that you're

90
00:04:28,560 --> 00:04:30,600
going to give this drug to.

91
00:04:30,600 --> 00:04:35,490
So Howie Silverman, who
was a master's student

92
00:04:35,490 --> 00:04:37,260
at MIT, who was a
year ahead of me,

93
00:04:37,260 --> 00:04:41,250
had written an initial
version of this program.

94
00:04:41,250 --> 00:04:49,320
And the job Pete gave me was to
basically, allow this program

95
00:04:49,320 --> 00:04:52,320
to explain its reasoning
to a physician,

96
00:04:52,320 --> 00:04:57,510
the idea being that if you
give a physician some advice

97
00:04:57,510 --> 00:05:00,930
but you don't also explain why
this is the right thing to do,

98
00:05:00,930 --> 00:05:02,970
they won't accept the advice.

99
00:05:02,970 --> 00:05:05,510
And so, basically,
I had to go back in

100
00:05:05,510 --> 00:05:09,390
and reprogram, develop,
essentially, a new approach

101
00:05:09,390 --> 00:05:12,930
to programming that
allowed the computer

102
00:05:12,930 --> 00:05:18,150
to take the routines and
procedures that it had

103
00:05:18,150 --> 00:05:21,060
and paraphrase
them into a form--

104
00:05:21,060 --> 00:05:26,560
into English in a reasonably
understandable sort of way,

105
00:05:26,560 --> 00:05:30,960
even for somebody who
isn't a computer genius.

106
00:05:30,960 --> 00:05:35,730
And so that became my
master's thesis, essentially.

107
00:05:35,730 --> 00:05:41,370
And I then went on on my PhD to
look at the issue of not only

108
00:05:41,370 --> 00:05:44,620
how can you get a program
to explain what it's doing,

109
00:05:44,620 --> 00:05:46,950
which is what looking
at the procedures

110
00:05:46,950 --> 00:05:52,590
does, but also look at how can
you get the program to explain

111
00:05:52,590 --> 00:05:54,610
why it's doing what it's doing?

112
00:05:54,610 --> 00:05:56,970
In other words, give
you the rationale

113
00:05:56,970 --> 00:06:00,330
behind the
recommendations that it's

114
00:06:00,330 --> 00:06:03,305
making that was sort of
the causal reasoning.

115
00:06:03,305 --> 00:06:04,680
The problem with
that, the reason

116
00:06:04,680 --> 00:06:06,660
why that was
particularly difficult

117
00:06:06,660 --> 00:06:09,900
is that that
information, the why,

118
00:06:09,900 --> 00:06:11,820
the why the program is
doing what it's doing,

119
00:06:11,820 --> 00:06:13,890
doesn't have to be
in the program for it

120
00:06:13,890 --> 00:06:15,060
to work correctly.

121
00:06:15,060 --> 00:06:17,520
You know, you can
get a rule that says,

122
00:06:17,520 --> 00:06:20,700
if you see this and
that phenomenon,

123
00:06:20,700 --> 00:06:23,117
you know, just reduce
the dose by 20%.

124
00:06:23,117 --> 00:06:25,200
But it doesn't tell you--
it does the right thing,

125
00:06:25,200 --> 00:06:27,880
but it doesn't tell you why
it's doing what it's doing.

126
00:06:27,880 --> 00:06:32,070
And so I basically looked into
developing some techniques

127
00:06:32,070 --> 00:06:38,670
for getting it so that we could
build a automatic programmer

128
00:06:38,670 --> 00:06:42,450
that would start from much
more abstract knowledge

129
00:06:42,450 --> 00:06:45,930
of the domain and
problem-solving principles

130
00:06:45,930 --> 00:06:48,840
and then from that derive--

131
00:06:48,840 --> 00:06:53,790
infer the working system,
keeping behind, essentially,

132
00:06:53,790 --> 00:06:57,210
the mental breadcrumbs of
where this procedure came

133
00:06:57,210 --> 00:06:59,580
from in this causal
reasoning chain

134
00:06:59,580 --> 00:07:03,810
so that then, you know, you were
running the program and said,

135
00:07:03,810 --> 00:07:05,500
OK, you know, why
is it doing this?

136
00:07:05,500 --> 00:07:09,000
It would go back and say, well,
there's this causal interaction

137
00:07:09,000 --> 00:07:12,007
that happens between the drug
and what you're-- you know,

138
00:07:12,007 --> 00:07:12,840
and this phenomenon.

139
00:07:12,840 --> 00:07:15,570
And if you don't do this, you
know, bad things will happen.

140
00:07:15,570 --> 00:07:17,350
And here's what they are.

141
00:07:17,350 --> 00:07:19,170
And so that was pretty cool.

142
00:07:19,170 --> 00:07:20,180
That was pretty fun.

143
00:07:20,180 --> 00:07:24,990
Could those kind of systems
work to infer what--

144
00:07:24,990 --> 00:07:29,100
I don't know-- why a
certain diseases is caused?

145
00:07:29,100 --> 00:07:32,700
They would have information
like that in the program.

146
00:07:32,700 --> 00:07:36,240
And the program--
the system that's

147
00:07:36,240 --> 00:07:39,280
either diagnosing-- building
the diagnostic system

148
00:07:39,280 --> 00:07:43,350
or the therapeutic system would
use that causal information

149
00:07:43,350 --> 00:07:47,070
together with some strategies
for how to reason in the domain

150
00:07:47,070 --> 00:07:50,740
to produce, then, a system
that would do the diagnosis.

151
00:07:50,740 --> 00:07:52,880
So yes, basically.

152
00:07:52,880 --> 00:07:53,580
Cool.

153
00:07:53,580 --> 00:07:57,000
And what do you think the
most exciting potentials in AI

154
00:07:57,000 --> 00:07:57,870
are right now?

155
00:07:57,870 --> 00:07:58,650
Right now?

156
00:07:58,650 --> 00:08:02,940
Well, the area I'm working
on is kind of actually--

157
00:08:02,940 --> 00:08:04,813
if you look at the
Dartmouth conference

158
00:08:04,813 --> 00:08:06,480
and the goals for
that, which was really

159
00:08:06,480 --> 00:08:10,920
to produce a integrated
artificial intelligence that

160
00:08:10,920 --> 00:08:17,430
really exhibited the whole
gamut of human behavior, one

161
00:08:17,430 --> 00:08:20,910
of the things I'm working on
now is virtual humans, which

162
00:08:20,910 --> 00:08:24,780
are computer-generated
characters that appear

163
00:08:24,780 --> 00:08:28,110
in a simulation or a game.

164
00:08:28,110 --> 00:08:31,860
And the goal is
basically to have

165
00:08:31,860 --> 00:08:35,250
them behave much like people.

166
00:08:35,250 --> 00:08:37,580
So you talk to them
in natural language.

167
00:08:37,580 --> 00:08:39,081
They talk back to you.

168
00:08:39,081 --> 00:08:40,289
But they're not just talking.

169
00:08:40,289 --> 00:08:41,747
It's not just verbal
communication.

170
00:08:41,747 --> 00:08:43,919
They're also using
gestures where appropriate.

171
00:08:43,919 --> 00:08:47,380
They're using eye gaze
as appropriate, so

172
00:08:47,380 --> 00:08:50,170
engaging the listener.

173
00:08:50,170 --> 00:08:51,790
And they're exhibiting emotions.

174
00:08:51,790 --> 00:08:55,210
So they're-- which is
actually, interestingly,

175
00:08:55,210 --> 00:09:00,160
an area of research that
AI really hasn't paid much

176
00:09:00,160 --> 00:09:03,400
attention to, I would argue,
over the last 50 years.

177
00:09:03,400 --> 00:09:05,890
It's just starting
to become something

178
00:09:05,890 --> 00:09:07,370
that we're looking at.

179
00:09:07,370 --> 00:09:10,000
And one of the things
that I think is--

180
00:09:10,000 --> 00:09:14,347
that we found in our
work is that emotion--

181
00:09:14,347 --> 00:09:16,930
we initially went into this we
thought emotion would just sort

182
00:09:16,930 --> 00:09:18,325
of be an overlay,
that, you know,

183
00:09:18,325 --> 00:09:20,590
wit would kind of be a
modifier of behavior.

184
00:09:20,590 --> 00:09:23,950
So if you're speaking and
you're angry, you kind of frown.

185
00:09:23,950 --> 00:09:26,260
But actually, what
we're finding is

186
00:09:26,260 --> 00:09:31,360
that emotion cuts much
deeper than that then.

187
00:09:31,360 --> 00:09:34,510
There are aspects of
emotions that are pervasive.

188
00:09:34,510 --> 00:09:39,400
So, for example, it
can affect the way

189
00:09:39,400 --> 00:09:40,820
that the system reasons.

190
00:09:40,820 --> 00:09:44,860
So if you're angry, if
we as humans are angry,

191
00:09:44,860 --> 00:09:51,250
you know, we will tend to reach
conclusions much more rapidly.

192
00:09:51,250 --> 00:09:54,220
If we're depressed, we'll
be more deliberative.

193
00:09:54,220 --> 00:09:56,110
We'll think things through more.

194
00:09:56,110 --> 00:09:59,770
And it's clear that
emotions are functioning

195
00:09:59,770 --> 00:10:02,770
as a kind of metacontrol
on some of our reasoning

196
00:10:02,770 --> 00:10:04,330
processes that we use.

197
00:10:04,330 --> 00:10:07,990
Seems like, maybe,
AI systems could

198
00:10:07,990 --> 00:10:11,530
benefit from some of that.

199
00:10:11,530 --> 00:10:17,890
It also seems like emotions
provide another window,

200
00:10:17,890 --> 00:10:22,390
essentially, into the
status of what's going on--

201
00:10:22,390 --> 00:10:30,210
could provide a status of what's
going on with these characters.

202
00:10:30,210 --> 00:10:34,570
Very often now if you
talk in the military,

203
00:10:34,570 --> 00:10:37,670
if you talk to generals,
they say, you know,

204
00:10:37,670 --> 00:10:39,580
we're getting reports
from the field.

205
00:10:39,580 --> 00:10:42,550
We would much rather get
a report from the field

206
00:10:42,550 --> 00:10:47,380
by radio rather than by
some sort of email message.

207
00:10:47,380 --> 00:10:49,300
The reason is because
they could hear

208
00:10:49,300 --> 00:10:53,380
in the tone of the voice of
the soldier, you know, how good

209
00:10:53,380 --> 00:10:56,000
or how bad things are going.

210
00:10:56,000 --> 00:11:00,160
And so, it's clear that
we, as people, use emotions

211
00:11:00,160 --> 00:11:02,560
as a kind of shorthand
for a lot of things.

212
00:11:02,560 --> 00:11:04,990
And it's part of
the communication.

213
00:11:04,990 --> 00:11:07,540
And it seems that
if we're really

214
00:11:07,540 --> 00:11:09,730
going to build AI
systems, you know,

215
00:11:09,730 --> 00:11:11,800
they have to have
that capability

216
00:11:11,800 --> 00:11:14,018
for a lot of different reasons.

217
00:11:14,018 --> 00:11:15,310
What are some of those reasons?

218
00:11:15,310 --> 00:11:19,240
Well, one is there's this
communication aspect.

219
00:11:19,240 --> 00:11:25,540
The other is that,
actually, people

220
00:11:25,540 --> 00:11:27,880
expect if they're
looking at something that

221
00:11:27,880 --> 00:11:29,920
looks like a person
up on the screen,

222
00:11:29,920 --> 00:11:35,350
they, in fact, will ascribe
emotions to that character

223
00:11:35,350 --> 00:11:36,850
whether they've
been modeled or not.

224
00:11:36,850 --> 00:11:38,892
It's just because we're
terribly anthropomorphic.

225
00:11:38,892 --> 00:11:41,080
You know, we think our
pets have emotions.

226
00:11:41,080 --> 00:11:43,670
All sorts of things
have emotions.

227
00:11:43,670 --> 00:11:47,590
And so if you're going to avoid
miscuing people, you know,

228
00:11:47,590 --> 00:11:49,380
you've really got
to get into the game

229
00:11:49,380 --> 00:11:54,760
of modeling these emotions.

230
00:11:54,760 --> 00:11:59,080
So that's kind of a very
quick summary of why some

231
00:11:59,080 --> 00:12:00,672
of these things are important.

232
00:12:00,672 --> 00:12:02,380
Has there been a lot
of cross-pollination

233
00:12:02,380 --> 00:12:06,130
between simulations and
artificial Intelligence

234
00:12:06,130 --> 00:12:08,140
and the video game world?

235
00:12:08,140 --> 00:12:14,110
That is a area that
is already quite warm

236
00:12:14,110 --> 00:12:16,120
and is really heating up.

237
00:12:16,120 --> 00:12:19,240
What's been happening with
video games, as you know,

238
00:12:19,240 --> 00:12:23,800
is over the last 10
years or so, the big--

239
00:12:23,800 --> 00:12:27,220
for a game company, the
big differentiator for them

240
00:12:27,220 --> 00:12:28,490
was the graphics.

241
00:12:28,490 --> 00:12:33,100
So if next Christmas your game
had the greatest graphics,

242
00:12:33,100 --> 00:12:34,110
you know, you had a hit.

243
00:12:34,110 --> 00:12:37,330


244
00:12:37,330 --> 00:12:40,630
What's happening
now with the degree

245
00:12:40,630 --> 00:12:44,350
of the advancement
in graphics cards

246
00:12:44,350 --> 00:12:47,050
and also the
underlying middleware

247
00:12:47,050 --> 00:12:50,380
for producing graphics
that our games have,

248
00:12:50,380 --> 00:12:52,933
it's getting to the point
where, first of all,

249
00:12:52,933 --> 00:12:54,850
the graphics are getting
pretty close to photo

250
00:12:54,850 --> 00:12:57,010
real in these environments.

251
00:12:57,010 --> 00:13:00,190
And pretty much all
the games are kind of

252
00:13:00,190 --> 00:13:02,710
at a similar level
in terms of graphics.

253
00:13:02,710 --> 00:13:05,698
Once you get to photo real, you
can't go any further, really.

254
00:13:05,698 --> 00:13:07,490
I mean, you're dealing
with the limitations

255
00:13:07,490 --> 00:13:09,370
now of the human eye.

256
00:13:09,370 --> 00:13:12,610
So that means
they're recognizing

257
00:13:12,610 --> 00:13:15,460
that graphics isn't going to be
the big differentiator for them

258
00:13:15,460 --> 00:13:16,320
anymore.

259
00:13:16,320 --> 00:13:20,110
And so what they're starting
to turn to is behavior,

260
00:13:20,110 --> 00:13:24,700
looking at how do these things,
these characters behave?

261
00:13:24,700 --> 00:13:27,460
The other thing that they're
looking to is how can they

262
00:13:27,460 --> 00:13:31,840
expand the market base beyond,
kind of, the first-person

263
00:13:31,840 --> 00:13:34,105
shooter--

264
00:13:34,105 --> 00:13:38,050
typical male, younger male--

265
00:13:38,050 --> 00:13:39,850
to a much broader audience?

266
00:13:39,850 --> 00:13:43,630
And what they're
finding is that games

267
00:13:43,630 --> 00:13:50,110
that emphasize relationships
tend to appeal more to--

268
00:13:50,110 --> 00:13:52,580
are more appealing to,
I should say, to women.

269
00:13:52,580 --> 00:13:54,970
And in fact, right
now, The Sims 2

270
00:13:54,970 --> 00:13:58,600
is, I believe, the
largest-selling computer

271
00:13:58,600 --> 00:14:00,890
game out there.

272
00:14:00,890 --> 00:14:06,370
And that's a basically
nonviolent game

273
00:14:06,370 --> 00:14:08,800
that is really all
about the relationships

274
00:14:08,800 --> 00:14:10,850
between these
simulated characters.

275
00:14:10,850 --> 00:14:13,120
Now, if you look at
The Sims, you know,

276
00:14:13,120 --> 00:14:18,370
the characters by AI standards
are still pretty primitive.

277
00:14:18,370 --> 00:14:21,520
And so we're in this
really interesting position

278
00:14:21,520 --> 00:14:23,350
from the standpoint
of AI research

279
00:14:23,350 --> 00:14:25,600
that there's a lot
of technology that we

280
00:14:25,600 --> 00:14:27,400
have in the lab that
without, I think,

281
00:14:27,400 --> 00:14:29,770
without too much
trouble, we can start

282
00:14:29,770 --> 00:14:34,720
to move out into these areas
and really create some really

283
00:14:34,720 --> 00:14:39,130
interesting game experiences.

284
00:14:39,130 --> 00:14:44,970
I think there will be a real
revolution in video games.

285
00:14:44,970 --> 00:14:46,690
Are there any other
places where you see

286
00:14:46,690 --> 00:14:50,510
that sort of symbiosis growing?

287
00:14:50,510 --> 00:14:53,910
I think another area
where this kind of thing--

288
00:14:53,910 --> 00:14:57,250
I would say education is
one, education and training.

289
00:14:57,250 --> 00:15:01,127
Basically, being able
to construct simulations

290
00:15:01,127 --> 00:15:03,460
of various-- and one of the
things we've been working on

291
00:15:03,460 --> 00:15:10,810
is a simulator that uses virtual
humans for teaching people

292
00:15:10,810 --> 00:15:12,250
how to negotiate.

293
00:15:12,250 --> 00:15:16,210
So, basically, you're
talking to a virtual human.

294
00:15:16,210 --> 00:15:18,940
And you're trying
to strike a deal

295
00:15:18,940 --> 00:15:23,110
with this virtual human, who
has his own internal goals

296
00:15:23,110 --> 00:15:24,730
and beliefs and desires.

297
00:15:24,730 --> 00:15:26,860
And as you talk
to him, you know,

298
00:15:26,860 --> 00:15:30,730
he's appraising what you say
in terms of his own desires

299
00:15:30,730 --> 00:15:32,410
and then reacting
and responding.

300
00:15:32,410 --> 00:15:35,620
And basically, the
characters respond

301
00:15:35,620 --> 00:15:39,520
using psychologically-motivated
strategies.

302
00:15:39,520 --> 00:15:41,565
So they do, you know,
they have a variety

303
00:15:41,565 --> 00:15:42,940
of different
strategies that they

304
00:15:42,940 --> 00:15:45,200
use in responding to people.

305
00:15:45,200 --> 00:15:47,830
So this kind of thing--
you can use this,

306
00:15:47,830 --> 00:15:53,050
you can imagine using this
not only in educating people

307
00:15:53,050 --> 00:15:54,820
about how negotiation
takes place

308
00:15:54,820 --> 00:15:58,660
but also as a way of
doing sort of role-playing

309
00:15:58,660 --> 00:16:02,450
for a variety of
business situations,

310
00:16:02,450 --> 00:16:06,250
for example, where you might
want to give people practice

311
00:16:06,250 --> 00:16:09,100
in how to close a deal.

312
00:16:09,100 --> 00:16:12,490
You might also want to use it
as a way for human resources

313
00:16:12,490 --> 00:16:16,900
in how to deal with
a problem employee.

314
00:16:16,900 --> 00:16:19,600
These are things that all
managers hate to deal with.

315
00:16:19,600 --> 00:16:22,300
So giving them the
ability to rehearse

316
00:16:22,300 --> 00:16:26,210
this with a artificial character
might be really appealing.

317
00:16:26,210 --> 00:16:30,310
So have you encountered any
really amusing instances

318
00:16:30,310 --> 00:16:34,690
of versions of these programs
or even just in your research

319
00:16:34,690 --> 00:16:36,820
group over the years?

320
00:16:36,820 --> 00:16:37,750
Well, let's see.

321
00:16:37,750 --> 00:16:41,100


322
00:16:41,100 --> 00:16:46,530
So I believe-- I mean, give me
a second to think about that.

323
00:16:46,530 --> 00:16:51,930


324
00:16:51,930 --> 00:16:53,902
In some of the--

325
00:16:53,902 --> 00:16:55,860
well, I'll tell you one
thing that was actually

326
00:16:55,860 --> 00:16:58,890
quite moving to me.

327
00:16:58,890 --> 00:17:02,730
We've been working a
lot with the US Army.

328
00:17:02,730 --> 00:17:11,730
And we constructed a simulation
of a scene in Bosnia.

329
00:17:11,730 --> 00:17:13,710
And it was using virtual humans.

330
00:17:13,710 --> 00:17:17,958
And the idea was we
wanted to put a--

331
00:17:17,958 --> 00:17:19,647
it was just a prototype system.

332
00:17:19,647 --> 00:17:21,480
But the idea was we
wanted to be able to put

333
00:17:21,480 --> 00:17:25,260
somebody like a lieutenant
into a typically--

334
00:17:25,260 --> 00:17:27,524
a dilemma that might
occur, typically,

335
00:17:27,524 --> 00:17:33,120
in a operation but
the kind of thing

336
00:17:33,120 --> 00:17:37,540
that is not typically part of
the standard training manual.

337
00:17:37,540 --> 00:17:39,570
So this was a
situation where you're

338
00:17:39,570 --> 00:17:41,220
in Bosnia with your platoon.

339
00:17:41,220 --> 00:17:44,430
You're going to reinforce
another unit that's

340
00:17:44,430 --> 00:17:46,230
dealing with some civil unrest.

341
00:17:46,230 --> 00:17:52,290
And on the way, your
group, your platoon

342
00:17:52,290 --> 00:17:55,350
gets involved with
a traffic accident.

343
00:17:55,350 --> 00:17:57,920
And there's a small child on
the ground seriously injured.

344
00:17:57,920 --> 00:18:00,060
And a crowd starts to form.

345
00:18:00,060 --> 00:18:03,630
TV camera crew shows up
and starts to film things.

346
00:18:03,630 --> 00:18:05,100
This is all done virtually.

347
00:18:05,100 --> 00:18:07,020
And you know, now you,
as the lieutenant,

348
00:18:07,020 --> 00:18:08,640
have to figure out what to do.

349
00:18:08,640 --> 00:18:11,700
Well, we showed this to
a bunch of army people.

350
00:18:11,700 --> 00:18:14,250
And actually, some
of them were really

351
00:18:14,250 --> 00:18:17,850
moved by this because it was--

352
00:18:17,850 --> 00:18:20,560
while the graphics weren't
perfect or anything like that,

353
00:18:20,560 --> 00:18:21,870
it was evocative enough.

354
00:18:21,870 --> 00:18:23,970
They'd actually been
on duty in Bosnia

355
00:18:23,970 --> 00:18:25,800
and it reminded them of things.

356
00:18:25,800 --> 00:18:29,160
And it really made me feel
like this is something

357
00:18:29,160 --> 00:18:32,620
that I can see how this
might actually work out.

358
00:18:32,620 --> 00:18:35,290
So it was a nice experience.

359
00:18:35,290 --> 00:18:36,150
Well, that's cool.

360
00:18:36,150 --> 00:18:37,560
Well, thank you very much.

361
00:18:37,560 --> 00:18:39,390
Thank you.

362
00:18:39,390 --> 00:18:40,560
Great.

363
00:18:40,560 --> 00:18:42,720
I love to bring so many
different perspectives

364
00:18:42,720 --> 00:18:43,330
to this field.

365
00:18:43,330 --> 00:18:45,680
It's absolutely amazing.