3 ways to make better decisions -- by thinking like a computer | Tom Griffiths

939,847 views

2018-10-05 ・ TED


New videos

3 ways to make better decisions -- by thinking like a computer | Tom Griffiths

939,847 views ・ 2018-10-05

TED


Please double-click on the English subtitles below to play the video.

00:13
If there's one city in the world
0
13407
1581
00:15
where it's hard to find a place to buy or rent,
1
15012
2333
00:17
it's Sydney.
2
17369
1150
00:19
And if you've tried to find a home here recently,
3
19043
2367
00:21
you're familiar with the problem.
4
21434
1840
00:23
Every time you walk into an open house,
5
23298
2014
00:25
you get some information about what's out there
6
25336
2191
00:27
and what's on the market,
7
27551
1393
00:28
but every time you walk out,
8
28968
1430
00:30
you're running the risk of the very best place passing you by.
9
30422
3214
00:33
So how do you know when to switch from looking
10
33660
2820
00:36
to being ready to make an offer?
11
36504
1562
00:39
This is such a cruel and familiar problem
12
39663
2401
00:42
that it might come as a surprise that it has a simple solution.
13
42088
3074
00:45
37 percent.
14
45588
1203
00:46
(Laughter)
15
46815
1828
00:48
If you want to maximize the probability that you find the very best place,
16
48667
3936
00:52
you should look at 37 percent of what's on the market,
17
52627
2825
00:55
and then make an offer on the next place you see,
18
55476
2310
00:57
which is better than anything that you've seen so far.
19
57810
2524
01:00
Or if you're looking for a month, take 37 percent of that time --
20
60358
3805
01:04
11 days, to set a standard --
21
64187
2915
01:07
and then you're ready to act.
22
67126
1575
01:09
We know this because trying to find a place to live
23
69829
2709
01:12
is an example of an optimal stopping problem.
24
72562
2325
01:14
A class of problems that has been studied extensively
25
74911
2483
01:17
by mathematicians and computer scientists.
26
77418
2105
01:21
I'm a computational cognitive scientist.
27
81502
2519
01:24
I spend my time trying to understand
28
84045
1960
01:26
how it is that human minds work,
29
86029
1798
01:27
from our amazing successes to our dismal failures.
30
87851
3671
01:32
To do that, I think about the computational structure
31
92552
2493
01:35
of the problems that arise in everyday life,
32
95069
2659
01:37
and compare the ideal solutions to those problems
33
97752
2375
01:40
to the way that we actually behave.
34
100151
1747
01:42
As a side effect,
35
102725
1185
01:43
I get to see how applying a little bit of computer science
36
103934
2715
01:46
can make human decision-making easier.
37
106673
1860
01:49
I have a personal motivation for this.
38
109828
1847
01:52
Growing up in Perth as an overly cerebral kid ...
39
112153
3166
01:55
(Laughter)
40
115343
4731
02:00
I would always try and act in the way that I thought was rational,
41
120098
3147
02:03
reasoning through every decision,
42
123269
1597
02:04
trying to figure out the very best action to take.
43
124890
2403
02:07
But this is an approach that doesn't scale up
44
127972
2190
02:10
when you start to run into the sorts of problems
45
130186
2242
02:12
that arise in adult life.
46
132452
1500
02:13
At one point, I even tried to break up with my girlfriend
47
133976
2728
02:16
because trying to take into account her preferences as well as my own
48
136728
3350
02:20
and then find perfect solutions --
49
140102
1683
02:21
(Laughter)
50
141809
2259
02:24
was just leaving me exhausted.
51
144092
1752
02:25
(Laughter)
52
145868
2533
02:28
She pointed out that I was taking the wrong approach
53
148425
2429
02:30
to solving this problem --
54
150878
1609
02:32
and she later became my wife.
55
152511
1382
02:33
(Laughter)
56
153917
2062
02:36
(Applause)
57
156003
4971
02:40
Whether it's as basic as trying to decide what restaurant to go to
58
160998
3461
02:44
or as important as trying to decide who to spend the rest of your life with,
59
164483
4055
02:48
human lives are filled with computational problems
60
168562
2375
02:50
that are just too hard to solve by applying sheer effort.
61
170961
4196
02:55
For those problems,
62
175650
1166
02:56
it's worth consulting the experts:
63
176840
1961
02:58
computer scientists.
64
178825
1151
03:00
(Laughter)
65
180000
1784
03:01
When you're looking for life advice,
66
181808
1819
03:03
computer scientists probably aren't the first people you think to talk to.
67
183651
3640
03:07
Living life like a computer --
68
187315
1875
03:09
stereotypically deterministic, exhaustive and exact --
69
189214
2578
03:11
doesn't sound like a lot of fun.
70
191816
1552
03:14
But thinking about the computer science of human decisions
71
194153
2927
03:17
reveals that in fact, we've got this backwards.
72
197104
2408
03:19
When applied to the sorts of difficult problems
73
199536
2198
03:21
that arise in human lives,
74
201758
1239
03:23
the way that computers actually solve those problems
75
203021
2727
03:25
looks a lot more like the way that people really act.
76
205772
2896
03:29
Take the example of trying to decide what restaurant to go to.
77
209267
2922
03:33
This is a problem that has a particular computational structure.
78
213090
3088
03:36
You've got a set of options,
79
216202
1524
03:37
you're going to choose one of those options,
80
217750
2049
03:39
and you're going to face exactly the same decision tomorrow.
81
219823
2811
03:42
In that situation,
82
222658
1151
03:43
you run up against what computer scientists call
83
223833
2295
03:46
the "explore-exploit trade-off."
84
226152
2372
03:49
You have to make a decision
85
229115
1334
03:50
about whether you're going to try something new --
86
230473
2334
03:52
exploring, gathering some information
87
232831
2183
03:55
that you might be able to use in the future --
88
235038
2285
03:57
or whether you're going to go to a place that you already know is pretty good --
89
237347
3793
04:01
exploiting the information that you've already gathered so far.
90
241164
3003
04:05
The explore/exploit trade-off shows up any time you have to choose
91
245488
3096
04:08
between trying something new
92
248608
1350
04:09
and going with something that you already know is pretty good,
93
249982
2980
04:12
whether it's listening to music
94
252986
1550
04:14
or trying to decide who you're going to spend time with.
95
254560
2634
04:17
It's also the problem that technology companies face
96
257218
2494
04:19
when they're trying to do something like decide what ad to show on a web page.
97
259736
3667
04:23
Should they show a new ad and learn something about it,
98
263417
2572
04:26
or should they show you an ad
99
266003
1383
04:27
that they already know there's a good chance you're going to click on?
100
267420
3306
04:30
Over the last 60 years,
101
270750
1151
04:31
computer scientists have made a lot of progress understanding
102
271925
2892
04:34
the explore/exploit trade-off,
103
274841
1452
04:36
and their results offer some surprising insights.
104
276317
2398
04:39
When you're trying to decide what restaurant to go to,
105
279291
2598
04:41
the first question you should ask yourself
106
281913
2000
04:43
is how much longer you're going to be in town.
107
283937
2179
04:46
If you're just going to be there for a short time,
108
286505
2342
04:48
then you should exploit.
109
288871
1515
04:50
There's no point gathering information.
110
290410
1905
04:52
Just go to a place you already know is good.
111
292339
2048
04:54
But if you're going to be there for a longer time, explore.
112
294411
2929
04:57
Try something new, because the information you get
113
297364
2387
04:59
is something that can improve your choices in the future.
114
299775
2974
05:02
The value of information increases
115
302773
1979
05:04
the more opportunities you're going to have to use it.
116
304776
2600
05:08
This principle can give us insight
117
308193
1754
05:09
into the structure of a human life as well.
118
309971
2070
05:13
Babies don't have a reputation for being particularly rational.
119
313104
3053
05:17
They're always trying new things,
120
317098
1825
05:18
and you know, trying to stick them in their mouths.
121
318947
2836
05:22
But in fact, this is exactly what they should be doing.
122
322621
2952
05:25
They're in the explore phase of their lives,
123
325597
2430
05:28
and some of those things could turn out to be delicious.
124
328051
2621
05:32
At the other end of the spectrum,
125
332040
1572
05:33
the old guy who always goes to the same restaurant
126
333636
2506
05:36
and always eats the same thing
127
336166
1635
05:37
isn't boring --
128
337825
1328
05:39
he's optimal.
129
339177
1509
05:40
(Laughter)
130
340710
3830
05:44
He's exploiting the knowledge that he's earned
131
344564
2228
05:46
through a lifetime's experience.
132
346816
1767
05:50
More generally,
133
350406
1151
05:51
knowing about the explore/exploit trade-off
134
351581
2031
05:53
can make it a little easier for you to sort of relax and go easier on yourself
135
353636
3667
05:57
when you're trying to make a decision.
136
357327
1810
05:59
You don't have to go to the best restaurant every night.
137
359161
2728
06:01
Take a chance, try something new, explore.
138
361913
2979
06:04
You might learn something.
139
364916
1627
06:06
And the information that you gain
140
366567
1589
06:08
is going to be worth more than one pretty good dinner.
141
368180
2543
06:12
Computer science can also help to make it easier on us
142
372178
2699
06:14
in other places at home and in the office.
143
374901
2152
06:17
If you've ever had to tidy up your wardrobe,
144
377860
2453
06:20
you've run into a particularly agonizing decision:
145
380337
2671
06:23
you have to decide what things you're going to keep
146
383032
2382
06:25
and what things you're going to give away.
147
385438
2009
06:27
Martha Stewart turns out to have thought very hard about this --
148
387974
3000
06:30
(Laughter)
149
390998
1205
06:32
and she has some good advice.
150
392227
1678
06:33
She says, "Ask yourself four questions:
151
393929
2295
06:36
How long have I had it?
152
396248
1526
06:37
Does it still function?
153
397798
1450
06:39
Is it a duplicate of something that I already own?
154
399272
3199
06:42
And when was the last time I wore it or used it?"
155
402495
2411
06:46
But there's another group of experts
156
406648
1715
06:48
who perhaps thought even harder about this problem,
157
408387
3130
06:51
and they would say one of these questions is more important than the others.
158
411541
3660
06:55
Those experts?
159
415819
1150
06:57
The people who design the memory systems of computers.
160
417281
2848
07:00
Most computers have two kinds of memory systems:
161
420153
2254
07:02
a fast memory system,
162
422431
1387
07:03
like a set of memory chips that has limited capacity,
163
423842
3169
07:07
because those chips are expensive,
164
427035
2052
07:09
and a slow memory system, which is much larger.
165
429111
3260
07:13
In order for the computer to operate as efficiently as possible,
166
433239
3000
07:16
you want to make sure
167
436263
1151
07:17
that the pieces of information you want to access
168
437438
2352
07:19
are in the fast memory system,
169
439814
1477
07:21
so that you can get to them quickly.
170
441315
1754
07:23
Each time you access a piece of information,
171
443093
2061
07:25
it's loaded into the fast memory
172
445178
1525
07:26
and the computer has to decide which item it has to remove from that memory,
173
446727
3877
07:30
because it has limited capacity.
174
450628
1741
07:33
Over the years,
175
453794
1151
07:34
computer scientists have tried a few different strategies
176
454969
2739
07:37
for deciding what to remove from the fast memory.
177
457732
2292
07:40
They've tried things like choosing something at random
178
460048
3176
07:43
or applying what's called the "first-in, first-out principle,"
179
463248
2906
07:46
which means removing the item
180
466178
1399
07:47
which has been in the memory for the longest.
181
467601
2176
07:50
But the strategy that's most effective
182
470091
2713
07:52
focuses on the items which have been least recently used.
183
472828
3229
07:56
This says if you're going to decide to remove something from memory,
184
476972
3191
08:00
you should take out the thing which was last accessed the furthest in the past.
185
480187
4968
08:05
And there's a certain kind of logic to this.
186
485179
2159
08:07
If it's been a long time since you last accessed that piece of information,
187
487362
3598
08:10
it's probably going to be a long time
188
490984
1762
08:12
before you're going to need to access it again.
189
492770
2192
08:15
Your wardrobe is just like the computer's memory.
190
495787
2774
08:18
You have limited capacity,
191
498585
2030
08:20
and you need to try and get in there the things that you're most likely to need
192
500639
5327
08:25
so that you can get to them as quickly as possible.
193
505990
2419
08:29
Recognizing that,
194
509016
1151
08:30
maybe it's worth applying the least recently used principle
195
510191
2813
08:33
to organizing your wardrobe as well.
196
513028
1949
08:35
So if we go back to Martha's four questions,
197
515001
2082
08:37
the computer scientists would say that of these,
198
517107
2861
08:39
the last one is the most important.
199
519992
1941
08:43
This idea of organizing things
200
523473
2055
08:45
so that the things you are most likely to need are most accessible
201
525552
3129
08:48
can also be applied in your office.
202
528705
1676
08:51
The Japanese economist Yukio Noguchi
203
531091
1930
08:53
actually invented a filing system that has exactly this property.
204
533045
3055
08:57
He started with a cardboard box,
205
537161
1609
08:58
and he put his documents into the box from the left-hand side.
206
538794
2940
09:02
Each time he'd add a document,
207
542067
1434
09:03
he'd move what was in there along
208
543525
1621
09:05
and he'd add that document to the left-hand side of the box.
209
545170
2810
09:08
And each time he accessed a document, he'd take it out,
210
548004
2667
09:10
consult it and put it back in on the left-hand side.
211
550695
2608
09:13
As a result, the documents would be ordered from left to right
212
553327
3281
09:16
by how recently they had been used.
213
556632
1761
09:18
And he found he could quickly find what he was looking for
214
558417
2774
09:21
by starting at the left-hand side of the box
215
561215
2054
09:23
and working his way to the right.
216
563293
1572
09:25
Before you dash home and implement this filing system --
217
565215
2621
09:27
(Laughter)
218
567860
1731
09:29
it's worth recognizing that you probably already have.
219
569615
2638
09:32
(Laughter)
220
572891
3310
09:36
That pile of papers on your desk ...
221
576225
2866
09:39
typically maligned as messy and disorganized,
222
579115
2474
09:41
a pile of papers is, in fact, perfectly organized --
223
581613
2565
09:44
(Laughter)
224
584202
1000
09:45
as long as you, when you take a paper out,
225
585226
2014
09:47
put it back on the top of the pile,
226
587264
2432
09:49
then those papers are going to be ordered from top to bottom
227
589720
2826
09:52
by how recently they were used,
228
592570
1552
09:54
and you can probably quickly find what you're looking for
229
594146
2701
09:56
by starting at the top of the pile.
230
596871
1700
09:59
Organizing your wardrobe or your desk
231
599988
1848
10:01
are probably not the most pressing problems in your life.
232
601860
2681
10:05
Sometimes the problems we have to solve are simply very, very hard.
233
605588
3602
10:09
But even in those cases,
234
609716
1150
10:10
computer science can offer some strategies
235
610890
2021
10:12
and perhaps some solace.
236
612935
1669
10:16
The best algorithms are about doing what makes the most sense
237
616048
3075
10:19
in the least amount of time.
238
619147
1502
10:22
When computers face hard problems,
239
622627
1958
10:24
they deal with them by making them into simpler problems --
240
624609
2762
10:27
by making use of randomness,
241
627395
1346
10:28
by removing constraints or by allowing approximations.
242
628765
3817
10:32
Solving those simpler problems
243
632606
1494
10:34
can give you insight into the harder problems,
244
634124
2920
10:37
and sometimes produces pretty good solutions in their own right.
245
637068
3031
10:41
Knowing all of this has helped me to relax when I have to make decisions.
246
641658
3569
10:45
You could take the 37 percent rule for finding a home as an example.
247
645251
3255
10:49
There's no way that you can consider all of the options,
248
649015
2702
10:51
so you have to take a chance.
249
651741
1580
10:53
And even if you follow the optimal strategy,
250
653888
2860
10:56
you're not guaranteed a perfect outcome.
251
656772
1931
10:59
If you follow the 37 percent rule,
252
659198
2154
11:01
the probability that you find the very best place is --
253
661376
3279
11:04
funnily enough ...
254
664679
1508
11:06
(Laughter)
255
666211
1014
11:07
37 percent.
256
667249
1420
11:09
You fail most of the time.
257
669966
1689
11:12
But that's the best that you can do.
258
672522
1744
11:14
Ultimately, computer science can help to make us more forgiving
259
674846
3052
11:17
of our own limitations.
260
677922
1909
11:20
You can't control outcomes, just processes.
261
680333
2539
11:22
And as long as you've used the best process,
262
682896
2119
11:25
you've done the best that you can.
263
685039
1763
11:26
Sometimes those best processes involve taking a chance --
264
686826
3688
11:30
not considering all of your options,
265
690538
2115
11:32
or being willing to settle for a pretty good solution.
266
692677
2729
11:35
These aren't the concessions that we make when we can't be rational --
267
695430
3350
11:38
they're what being rational means.
268
698804
1629
11:40
Thank you.
269
700833
1151
11:42
(Applause)
270
702008
4547
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7