collec&ve(intelligence( as(an(emergentproperty€¦ · aristote((poli6que(iii xi) il peut...
TRANSCRIPT
Collec&ve intelligence as an emergent property
Daniel Andler h7p://andler.dec.ens.fr/
UFR de philosophie et sociologie, Université Paris-‐Sorbonne E.A. « Ra&onalités contemporaines »
UPMC -‐ ASHiC Colloque
« Réduc-onnisme et propriétés émergentes dans les sciences » 16 novembre 2012
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 2
Example 1 : The USS Scorpion
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 3
Example 2 : Galton’s ox
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 4
Example 3 : Elec&on results
Predic&on markets have been giving nearly perfect predic&ons for presiden&al (and other) elec&ons: average precision (up to 2 years ago) 1.37% for US elec&ons; 2% for non-‐US elec&ons. Par&cularly good predic&ons for Obama-‐Romney, overall and state-‐by-‐state.
Ques&ons for today
1. Is there anything worth the name “collec&ve intelligence”?
2. If so, is it anything we are not already familiar with?
3. If so, where does the difference reside?
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 5
Answers (tenta&ve) 1. Yes, there is something deserving of the name
“collec&ve intelligence”. 2. No, it’s not (necessarily) the same as any of the usual
suspects. 3. It is dis&nguished by its ‘magical’ character, i.e. by the
fact that it is emergent. → �If I’m right, it is an addi6onal example of an emergent phenomenon, not one which is usually considered. → �It may be an interes6ng comparison point for thinking of (individual) intelligence as emerging from non-‐intelligent components (neurons, Boolean gates...).
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 6
1. Is collec&ve intelligence anything new?
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 8
Three (+1) meanings of "collec&ve intelligence "
• The loose sense : a group’s disposi&on to behave so as to further its interests as a group ➢ core meaning with 2 components, intelligence and group, which can be construed in various ways.
• The pseudo-‐techical sense, in ‘management science’ : a ‘theory’ of management (example : Vincent Lehnardt & Philippe Bernard, L’intelligence collec6ve en ac6on, 2e éd. 2009) ➢ a method supposed to improve crea&vity and decision-‐making in corpora&ons and administra&ons [e.g.: aligning inten&ons, commitments, behaviors]
• The technical sense : formalized set-‐ups devised to tap and pool together the intellectual resources of the group’s members Ø In a restricted sense: internet-‐based (and more generally, ICT-‐based) set-‐ups.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 9
Did I hear you say Intelligence ?
• There are some perfectly decent working defini&ons of intelligence. ¤
• Intelligence can be studied empirically, without falling prey to the well-‐known difficul&es of IQ. ¤
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 10
A working defini&on of intelligence
• The ability to understand the world and to put this understanding to use in order to find in due &me acceptable solu&ons to an unlimited variety of pressing problems, including those stemming from the need to improve our understanding (to describe, explain, predict) the world.
• Intelligence is a maFer of degree: it varies both with the individual and with the task domain; and it is dependent on the cultural context. A given individual is more or less intelligent according to the problem (s)he faces, and is all the more intelligent as (s)he exhibits intelligence in more domains and can move from one to the other in a fluid way.
• ¤
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 11
¤
What about ‘collec&ve’?
Collec&ve intelligence is a feature of any kind of group: – whether it is cons&tuted before the process begins • and quite independently of the problem at hand • or precisely with the aim of solving the problem
– or it is co-‐cons&tuted by the problem and the solu&on-‐seeking process
– or it is self-‐organized among the people concerned
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 12
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 13
How novel is it?
Aristote (Poli6que III XI) Il peut arriver que la majorité, dont chaque membre n'est
pas vertueux, réunie toute ensemble soit meilleure que l'élite, non pas séparément mais collectivement[...]. Etant donné qu'ils sont nombreux, chacun détient une part de vertu et de sagesse, et, de cette réunion, la masse devient comme un seul homme, à plusieurs pieds et plusieurs mains, et pourvu de plusieurs sensations, et il en va de même pour son caractère et son intelligence.
Analysis + synthesis • Top-‐down schema:
The goal is known (i.e. iden&fied) ➢� tasks are assigned around (ac&ve ‘analysis’)
and the par&al results are mechanically assembled (passive ‘synthesis’) ➢� The goal is achieved.
• BoFom-‐up schema:
The goal is unknown ➢� the ‘tasks’ (epistemic resources) are distributed by some natural
process (passive ‘analysis’) and the outcomes are dialec&cally assembled (ac&ve ‘synthesis’)
➢� The goal is known (i.e. iden&fied)
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 14
Division of labor and delibera&on
• Top-‐down schema = technical division of labor ¤
• Bo7om-‐up schema = classical delibera&ve processes, giving rise (unsurprisingly) to a familiar kind of collec&ve intelligence: reflec6ve collec6ve intelligence CIr. ¤
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 15
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence
¤
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 17
¤
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 18
Specific collec&ve intelligence
• It’s the bo7om-‐up schema in which the dialec&cal (ac&ve) synthesis is replaced by a mechanical (passive) synthesis.
• This gives rise to seemingly unintelligent collec&ve processes, involving no delibera&on, yet producing (surprisingly) a specific collec6ve intelligence collec6ve CIs, a.k.a. wisdom of crowds WoC.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 19
Principles of CIs
The basic idea
(i) Individuals hold local, unsystema&c bits of knowledge, each of which is (or may be) of crucial importance, but which cannot be deliberately synthesized without loss or skewing.
(ii) The required synthesis can only be achieved by means of a blind, market-‐type process.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 21
The market place of ideas • The original insight: Markets (and markets only) have the
werewithal to determine the correct prices. Hayek, The use of knowledge in society, American Economic Review (1945).
• CIs generalizes this idea: Markets (and
only markets) have the werewithal to determine the correct solu&on to certain problems.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 22
What’s transgressive about CIs • Ra&onality postulate: only individuals are capable of ra&onality.
• Empirical fact: ra&onality in the real world relies on social processes.
• Concilia&on: The principle of individual recapitula&on–founda&on of CIr. CIs rests on the rejec6on of the principle of
individual recapitula6on.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 23
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 24
In what kinds of situa&ons can CIs be deployed?
• Pure formes: – Es&ma&on – Predic&on
• Hybrid forms [CIh]: – Problem solving – Coordina&on – Crea&on
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 25
CIs for es&ma&on
Scorpion Galton’s ox and Treynor’s jelly beans Goldcorp’s Red Lake gold mine d'or Challenger
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 26
CIs for predic&on • Predic&on or predic&ve markets, a.k.a informa&on markets, decision
markets, idea futures, event deriva&ves: rhe purest, most elaborate and studied applica&on of CIs.
• The oldest: IEM (Iowa Electronic Market) – launched in 1988, in recognizable form (‘winner take all’) 1992
• Fields of applica&on : forecast – elec&ons – poli&cal event – sports – evolu&on of a market, of an epidemic – chances of new product, &me of achievement of complex task... – just about any complex evolu&on with an uncertain outcome
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 27
Predic&on markets: the core ideas Stock exchange // Idea futures An agent’s judgment concerning
the value of a business // some given event gives rise to a
monetary// monetary or pseudo-‐monetary bet that is propor&onal to the be7or’s confidence
(and thus in normal cases, on average, to her competence) and secondarily to the difference between her belief and the average es&mate.
The bets are taken by a self-‐appointed popula&on of agents. The market, which serves to broadcast the bets on offer, provides both
a balancing procedure between supply and demand and an aggrega&on procedure des croyances des parieurs.
Les parieurs exercent leur intelligence individuelle. Le mécanisme d'agréga&on est automa&que, à la fois impar&al et "aveugle"
et fournit un résultat d'ICs.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 28
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 29
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 30
CIs for coordina&on, problem solving, crea&on
...
Theories of ICs
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 31
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 32
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 33
Counterintui&ve theorems regarding CIs
Under certain reasonable assump&ons, 1. In predic&on tasks, diversity is as important as
indiviual accuracy. A crowd always does be7er than the average, and can do be7er than the best of its members.
2. In problem solving, individual excellence is less important than diversity: a diverse group does be7er than a homogeneous one, however elite.
Sco7 E Page, Lu Hong
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 34
¤
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 35
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 36
What makes CIs work? 1. Two basic mechanisms
• Aggrega&on of local, par&al informa&on • Automa&on of aggrega&on
2. General success condi&ons • Diversity of the individuals’ models of the situa&on • Mutual independence of the agents • Non-‐centralized aggrega&on
3. Favorable factors • Aggrega&ng not opinions but par&al knowledge of as many
independent agents as possible • Weighing par&al informa&on by the agents’ (self-‐assessed)
but...
... but other factors might come into play
Other possible reasons (making CI ‘work’ in realis&c situa&ons):
– Sta&s&cal ‘magic’? The CI of Buffon’s needles... – Luck, coincidence (anecdotes) – Hidden variables: good old-‐fashioned ra&onality working behind the
scene
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 37
Is ICs emergent?
In some loose sense, yes
• Two levels: micro (individual agents), macro (output: es&mates, solu&ons, design, tools...)
• No account in terms of micro-‐proper&es + assembly [elementary structural proper&es]
• Surprise factor • No predictability from first principles, simula&on only
• Control by global proper&es (such as diversity, independence...)
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 39
A new species of emergence?
• No ontological novelty: end result does not differ in its nature or essence from anything that could not have been produced at the micro level.
• But produc&ve or heuris&c novelty: within the set of like results, the selec&on or iden&fica&on of this par&cular one is emergent.
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 40
Concluding ques&ons
• Does CIs count as a genuinely emergent phenomenon?
• Does it extend in an interes&ng way the repertory of examples?
• Could it benefit from the analy&cal tools developed for the general theory of emergence?
Réduc&onnisme & émergence D. Andler. Collec&ve intelligence 41