Transcript Information Flows in Computational Approaches to the

Information Flows in
Computational Approaches to the
Evolution of Cooperation
Timothy A. Kohler
Department of Anthropology
Washington State University
Pullman, WA 99164-4910 USA
[email protected]
Plan
• What is information?
• How information functions in various
computational approaches proposed to
understand the problem of human
cooperation
– Simple, easy-to-understand systems
– Structure of such problems is—or can be recast to become—why people
forego immediate personal gain to pursue prosocial, longer-term goals that
might include achieving a sustainable but still high-payoff interaction with
the environment
Information (1)
• Intrinsic component of all physical systems?
• Flows, but cannot fill up an empty container
– Claude Shannon developed techniques for measuring
the information rate of a source and the capacity of a
channel
– Others argue that information itself has “zero
dimension” and that therefore, like contrast, symmetry,
correspondence, etc., cannot be “located” (Gregory
Bateson)
Information (2)
• Just as mass is a reflection of a system containing
matter, and heat is a reflection of a system
containing energy, organization is the physical
expression of a system containing information.
• Just as energy has as one of its fundamental
attributes the capacity to perform work,
information has as one of its fundamental
attributes the capacity to organize things.
Information (3)
• Information is the raw material which, when
information-processed, may yield a message.
Upon receipt of a message, the message must once
more be information-processed by the recipient for
the message to acquire meaning.
• There are many different kinds of informationprocessing systems (ISPs); for example, the
thermostat is a mechanical ISP; computers are
electronic ISPs; brains are neurological ISPs.
"Lineages" or "flow structures" of
information occur in three domains that
form a nested hierarchy (Goonatilake
1991):
• the genetic;
• the neural-cultural (including individual
memory and cognitive process, which I
consider useful to separate from culture and
its transmission);
• and the "extrasomatic" (artifacts, especially
information storage devices).
Cooperation & Sharing
• Two large classes of contemporary
approaches to understanding altruism:
– unselfish behaviors are really disguised selfish
behaviors
– truly unselfish behaviors have been able to evolve
through genetic or cultural group selection
• A behavior is altruistic when it increases the
fitness of others but decreases the fitness of
the actor.
Kin Selection
• Genes are the fundamental units of reproduction,
and we are only vessels for our genes
• We therefore ought to be molded by evolution to
assist our closest relatives preferentially
• Philip Morin et al. (1994) have demonstrated that
such cooperative behaviors among chimp males as
defense of territory and political support to
achieve alpha male status seem explicable through
principles of kin selection
• Unarguable, but incomplete
(Direct, Pairwise) Reciprocal
Altruism
• Classic IPD: role of information
extremely constrained
• In “memory-1” strategies such as TFT, the
only information one player has about the
other is whether that player cooperated or
defected in the past move.
Prisoner’s Dilemma
Row
Player
Cooperate
Defect
Column Player
Cooperate
Defect
R=3, R=3
S=0, T=5
T=5, S=0
P=1, P=1
Notes:
Payoffs to row player shown first in each ce ll
R=Reward for mutual cooperation
S=Sucker’s payoff
T=Temptation to defect
P=Punishment for mutual defection
Common values for these payoffs are shown, but others are possible if
they satisfy:
T > R > P > S, and
R > (S + T)/2
Lindgren & Nordahl
• possibility of the initially neutral evolution
of strategies with longer memory
• for example, a memory 2 strategy would
remember both an opponent's and self's
prior moves, allowing for the appearance of
more complicated strategies
• longer-memory strategies have an
advantage
Characteristics of Pairwise
Reciprocity
• In 2-person games cooperation may emerge
through selection in evolutionary games if pairs
of individuals interact for a sufficient number of
times, often leading to something generically
similar to a Tit-for-Tat strategy
• Populations composed of unconditional
defectors can resist invasion by reciprocators
unless there is some degree of assortative
pairwise interaction
Problems with Pairwise
Reciprocity
• When groups are larger than 2 are formed at
random, however, and the play is structured as
an n-way prisoner's dilemma, reciprocating
strategies become increasingly less likely
• Groups of 32 expecting to have about 1,000
interactions must have initial frequencies of
some 70% reciprocators for cooperative
strategies to increase (Boyd and Richerson
1988)
Indirect Reciprocity
• Pairs of players who encounter each
other infrequently
• Changes information that players
have about each other in an important
fashion
• One well-known formalization due to
Nowak and Sigmund (1998)
Nowak and Sigmund (1)
• Random pairs of players are drawn from a
population, one of whom is a potential donor
and the other a potential recipient
• Donor can cooperate and help the recipient
at a cost c to himself, in which case the
recipient receives a benefit of value b (b>c)
• If the donor decides not to help, both
individuals receive zero payoff
Nowak and Sigmund (2)
• Each player has an image score, s, which
in some games is visible to all players
• If a player chosen as a donor chooses to
help, her image score increases by 1; if
she chooses not to help, her image score is
decreased by one unit (the image score of
a recipient does not change in either case)
Nowak and Sigmund (3)
• Players have various strategies for “helping” and
decide to help based on a threshold k: if the
recipients image score (s) is greater than k they
help, if not, they don’t
• Players with the highest scores (the payoffs from
playing, not the image scores) at the end of a
round produce offspring in proportion to their
scores
Nowak and Sigmund (4)
• When everyone's image can be seen by each
player, and when there is no mutation, the
population is quickly dominated by players with
a k=0 strategy (they will cooperate with anyone
who has an image score of 0 or better)
– the most discriminating of all the cooperative strategies
• Under mutation and selection, strategies cycle
endlessly
Indirect Reciprocity (concluded)
• If players have information about the degree of other
players' propensities to cooperate, and make judgments
about whether to cooperate based on that "reputation,"
cooperation is easier to sustain, even in large groups, so
long as these "reputations" are widely known
• Some recent empirical work (Wedekind and Milinski
2000) suggests that people do not necessarily focus
their altruistic acts on those who have been kind to
others, calling into question the mechanism proposed
here
• Cognitive burdens?
Costly Signaling (1)
• Some similarities to previous approach: signals that
are costly to the emitter constitute information used
by a recipient in choosing an action, such as
whether or not to cooperate with the signaler
• But here, signal is connected to some underlying but
poorly observable quality of the signaler which
nevertheless is of importance to the receiver
• Also, signaler is providing a public benefit through
his or her signal—for example, turtle hunting
among the Merriam (Smith and Bliege Bird 2000)
Costly Signaling (2)
• Prosocial—or any other type of—
signaling is a Nash equilibrium if—
– low-quality types pay greater marginal costs for
signaling than do high-quality types
– other group members benefit more from
interacting with high-quality than with lowquality types
– other group members benefit more from
interacting with high-quality than with lowquality types (Smith, Bowles, and Gintis 2000)
Costly Signaling (3)
• If the reason that indicator traits of
underlying qualities are often "prosocial"
(enhancing the well-being of members of
Ego's social group beyond his or her
immediate kin) is that pro-social traits are
valued in and of themselves (since they
signal the signaler's value as a potential ally)
then the line between CST and indirect
reciprocity based on image scoring is
extremely fine
“Apparent” Altruism
(Concluded)
• Except perhaps for indirect reciprocity
through image scoring, any of the
pathways to cooperation considered above
employ mechanisms that would be as
available to other animals as to humans
• Seminal contributions to reciprocal
altruism and costly signaling were by
biologists (Trivers, and Zahavi and
Grafen, respectively)
Strong Reciprocity and
Multilevel Selection (1)
• Relies on a cultural information channel
• A form of altruism that benefits group
members at a cost to the strong reciprocators,
and involves a predisposition to follow a
social norm to cooperate with others and
punish non-cooperators
– Defined to contrast with reciprocal altruism which is
considered "weak reciprocity" because it is so dependent
on high probabilities for future interaction
– Result is truly unselfish behavior, not disguised
selfish behavior
Necessary Requirements for Group Selection:
1. There must be more than one group (there must be a population of groups);
2. Groups must vary in their proportion of altruistic types;
3. There must be a direct relationship between the proportion of altruists in
the groups and the groups’ fitnesses (groups with more altruists must produce
more offspring);
4. Groups must be isolated for at least a portion of their life cycles; however,
their progeny must be able to mix or compete in the formation of new groups.
Given these conditions, group selection can be effective if:
6. The differential fitness of groups (the force favoring the altruists) must be
strong enough to overcome the differential fitness of individuals within
groups (the force favoring the selfish types).
—Sober & Wilson 1998:26
Strong Reciprocity and
Multilevel Selection (2)
• Imagine a group of foragers, largely
unrelated, in which agents can either work
alone, or work cooperatively in a group,
which in general is more rewarding.
• Output is shared equally by all agents.
• Agents may shirk, which reduces the
output to be shared
– it is advantageous to the shirker
– if there were no policing of free riders, even
complete shirking would promote a member's fitness.
Strong Reciprocity and
Multilevel Selection (3)
• Group is small enough that members can
be monitored, and if detected shirking,
may be punished, at some cost to the
punisher.
• Punishment consists of a shirker being
ostracized from the group.
• Ostracized agents work alone for a period
before being readmitted to a different
group.
Strong Reciprocity and
Multilevel Selection (4)
• Groups consist of two types of people:
– reciprocators who work and always punish
shirkers when they see them, even though there
is a cost to doing so;
– self-interested individuals who maximize their
fitness and therefore never punish, and work
only to the extent that the expected cost of doing
so is less than the expected cost of being
punished.
• Punishment consists of a shirker being
ostracized from the group.
Strong Reciprocity and
Multilevel Selection (5)
• Analysis of fitness consequences of reasonable
parameter settings for the costs and benefits
shows that a stationary equilibrium can be
expected composed of
– 70% of the population in groups (and therefore 30%
not in groups);
– these groups composed disproportionately of
reciprocators.
• Population as a whole should be composed of
65% reciprocators, but the groups should
stabilize with 70% reciprocators.
Strong Reciprocity and
Multilevel Selection (6)
• To the extent that the authors hazard a
historical reconstruction of how such
processes might have played out in the
Pleistocene, they favor the idea that "the
cognitive and affective traits required to
fashion, learn, detect violations of, and wish to
uphold social norms may be genetically
transmitted, while the content of the norms
(and in particular the linking of nonshirking
and punishing) may be culturally transmitted"
(Bowles and Gintis 2001:16).
Strong Reciprocity and
Multilevel Selection (7)
• Useful to draw on culture—our capacity
for which differentiates us from the rest of
the animal kingdom—to explain our
propensity to live in highly and flexibly
cooperative but largely unrelated
groups—a practice that likewise
distinguishes us from other animals, at
least in degree, and does much to explain
the success of the species.
How do these models map into
the prehistory of human societies?
Extremely Slow Growth in H.
erectus
• erectus populations in Asia by 2 mya
• linguistic capabilities of these hominids will
probably never be known in detail.
Evolutionary psychologist Robin Dunbar
(1996) suggests that language is used in
three different ways:
– formulaic usages with little semantic content;
– “gossip” for passing social information;
– and symbolic language of the sort employed in
a presentation like this.
Language uses (continued)
• Of these, the second by far dominates our
everyday uses, and greatly facilitates the
integration of social groups.
• Among primates, grooming is a key
mechanism for maintaining social
relationships, and grooming time appears to
increase linearly with group size among the
Old World monkeys and apes.
• Neocortex size relative to total brain size
likewise scales with mean group size.
Language uses (continued)
• “If we interpolate the predicted group size for
humans based on our neocortex size into the
relationship between group size and grooming
time, we find that humans would have to spend
something in the order of 40 per cent of the day
engaged in grooming in order to maintain …
cohesion” (Dunbar 1996:383).
– The earliest steps towards the development of language,
therefore, may have been as substitute verbal grooming,
or gossip, mitigating this time crisis
– Seems unlikely that language would have evolved
beyond that point in this period
Pre-sapiens (concluded)
• Of those mechanisms for cooperation
discussed here, which would have been
available to these populations?
• Certainly any but perhaps "strong
reciprocity," which may require greater
symbolic abilities to understand and
teach the social norms on which it
depends
Acceleration of population growth
with the appearance of anatomically
modern humans
• 120 to 35 ka for this process (Klein 1992)
• Leading components probably include
– longer juvenile dependence and lower juvenile
mortality
– increased provisioning of females by males
with the development of stronger male–female
bonds and more exclusive mating patterns
(Foley 1992); nuclear families?
Leading components in transition to
anatomically modern humans (cont.)
• Longer period of juvenile dependency
within the various families of the band
provided a more variable, and more often
one-to-one enculturation process that would
lead to more rapid innovation than the
strongly conservative many-to-one
enculturation inferred for earlier times.
Archaeology of the Upper Paleolithic
• Aggregation sites of the later Upper
Paleolithic suggest the existence of social
units that are too large for the participants to
be related genetically, or perhaps even to be
known to each other through regular face-toface contacts.
– Such aggregations, even if only periodic, would have
been promoted by (and may not be possible without) the
existence of cultural norms dictating appropriate behavior
(Chase 1994).
Archaeology of the Upper Paleolithic
(cont.)
• Many of the Upper Paleolithic symbolic
activities seem to be ritual in nature, and
ritual enforces, and reinforces, such norms.
• Language may have existed prior to such
symbolic activity—Dunbar’s model predicts
that “social language” (gossip) should
emerge well before this time—but these
symbolic activities must have been carried on
through fully symbolic language.
Archaeology of the Upper Paleolithic
(concluded)
• Shared, symbolic norms may have very
tangible consequences for individual and
group survival.
• Richerson and Boyd (1992:69–71) explain
the origin of culture as due to its success
(when coupled with individual learning and
genetic inheritance) in providing a selective
advantage in environments that are neither
too constant nor too variable from generation
to generation.
Upper Paleolithic Societies
• Success of these populations in part due
to the greater cooperative possibilities
for large groups made available by
strong reciprocity.
• If so, then all the mechanisms for
cooperation considered here would have
been available to human populations by
35 ka at the latest.
Neolithic Societies
• What then accounts for the next feature of interest
in population graph, the large increase in rate of
population growth associated with Neolithic
societies?
• Problems of organizing larger settlements and
societies remind us that Goonatilake's third lineage
of information—"extrasomatic" artifacts, especially
information storage devices—has not been
exploited in the accounts of cooperation presented
to this point.
External Information Storage
• Artifacts such as tokens that allowed storage of
economic information
• Institutions that provided a framework for the
organization of people in ways that cross-cut kin
networks
• eventually including stratified societies that radically
changed the egalitarian assumption of all the models
presented above
• All were of selective value in societies that had
grown beyond the size that could be maintained
solely by the mechanisms discussed earlier.