Definizione del

1

• Definizione del
• comportamento pro-sociale

2
Overview

• Definizione del comportamento pro-sociale
• L'altruismo come dilemma sociale
• Un esempio etologico i pipistrelli vampiro
• l'utilità di un'analisi cross-metodologica
• Cooperazione e gruppi
• Un modello multi agente per l'evoluzione della
  cooperazione
• I mediatori cognitivi del comportamento
  pro-sociale
• implementazione di un sistema (quasi)cognitivo
  per lo studio dell'altruismo

3
Il comportamento altruistico

• Unazione altruistica
• comporta un costo (c) per lindividuo che la pone
  in atto
• produce un beneficio (b) per un altro individuo
• b gt c gt 0
• Viene eseguita e/o replicata in ragione di tale
  beneficio (effettivo o rappresentato)

4
Meccanismi causali

• Funzione adattativa

5
Meccanismi causali

• Funzione adattativa
• Rappresentazioni
• mentali

6
Interazione strategica

• Dilemma del Prigioniero
• T gt R gt P gt S
• 2R gt T S

7
Interazione strategica

• Dilemma del Prigioniero
• T gt R gt P gt S gt b gt (b c) gt 0 gt c
• 2R gt T S gt 2(b c) gt (b c)
• Applicato al caso del comportamento altruistico
• T b R (b c) P 0 S c.

8
L'evoluzione della cooperazione

• Quando si considera il comportamento altruistico
  come il prodotto di una funzione adattativa, per
  giustificare l'evoluzione della cooperazione
  vengono solitamente utilizzati tre ordini di
  fattori
• 1. Parentela (kin selection, Hamilton 1964) la
  stabilità evolutiva di un comportamento
  altruistico è funzione del grado di parentela
  degli organismi considerati.
• r gt b/c
• r 1/2
• r b/2 .

9
L'evoluzione della cooperazione

• Quando si considera il comportamento altruistico
  come il prodotto di una funzione adattativa, per
  giustificare l'evoluzione della cooperazione
  vengono solitamente utilizzati tre ordini di
  fattori
• 2. Reciprocazione (reciprocal altruism, Trivers
  1971) la teoria dell'altruismo reciproco
  condiziona l'evoluzione della cooperazione alla
  possibilità che un atto altruistico venga
  restituito (probabilità di due altruisti di
  reincontrarsi).
• Tit-For-Tat metafora evoluzionistica
  dell'altruismo reciproco.

10
L'evoluzione della cooperazione

• Quando si considera il comportamento altruistico
  come il prodotto di una funzione adattativa, per
  giustificare l'evoluzione della cooperazione
  vengono solitamente utilizzati tre ordini di
  fattori
• 3. Gruppi (group selection, Wilson Sober 1994)
  in uno scenario in cui siano presenti più
  gruppi, ognuno contenente percentuali differenti
  di altruisti, un processo di selezione
  inter-gruppo tenderebbe a preferire quelli nei
  quali il gene altruista è presente in percentuale
  maggiore.

11

• Vampire Bats

12
Altruism In Nature

• In nature, examples of altruism abound (cfr.
  Brembs, 1996).
• Inter-specific mutualism (Leimar and Axén, 1993).
  
• Predator inspector in shoaling fish (different
  individuals swim towards the predator, gathering
  information about his location and his current
  motivational state (Milinski et al., 1990).
• Among mammals,
• blood-sharing in vampire bats (Wilkinson, 1984)
• many controversial examples among primates.

13
Ethological Data about Bats

• The species of vampires studied by Wilkinson
  lives in Central America, in small groups sharing
  cavity of trees (roost), where animals reproduce
  and perform social activities (nursing, grooming
  and sharing food)
• Each night vampires go out hunting (finding
  herbivores to suck blood from)
• about 7 hunters find no prey,
• survive thanks to luckier fellows regurgitating a
  portion of food ingested
• This behavior "depends equally and independently
  on degree of relatedness and an index of
  opportunity for reciprocation (Wilkinson, 1984).

14
The economy of food-sharing

• Un pasto garantisce ai pipistrelli 60h di
  autonomia.
• Il cibo donato, quantificato in base alla
  percentuale di peso corporeo perso dall'animale,
  avrà un'utilità marginale differente (h di
  autonomia) per chi lo riceve.
• Un atto di food-sharing avrà l'effetto di
  conferire ad un conspecifico le energie
  necessarie per tornare a cacciare.

15
Vantaggio evolutivo del food-sharing

• Numero fallimenti consecutivi nel procacciarsi il
  cibo in un anno 1,63.
• Vantaggio evolutivo del food-sharing tasso di
  mortalità annua ridotto dal 82 al 24.

16
Why Bother with Vampires?

• We are not interested in sociobiological debate
  per se.
• Aims Model altruism and reciprocity at an
  abstract level.
• Using real data for setting parameters values in
  a non arbitrary way.
• Vampire bats offer real data about altruism as
  strongly interdependent with survival

17

• Cooperazione Gruppi

18
The Problem in Nature

• Vampire bats altruism (Wilkinson, 1984)
• If successful, hunters regurgitate in favour of
  unlucky fellows
• Survival rate
• rises up to 80 of the initial population,
• as opposed to 20 without altruism.
• Why?
• Kin selection? Not quite
• Low average rate of relatedness among individuals
  living in the same roosts (around 6).
• Reciprocal altruism? Perhaps. But
• Simulations supporting group selection theory
  have recently been run (see Paolucci et al.,
  2003).

19
Reciprocal Altruism Theory

• Direct version donor receives help from current
  recipient.
• Indirect version donor receives help from
  someone else reciprocity circulates in the
  group, increasing donors' fitness.
• Circularity makes reciprocity fragile (cheaters
  and noise).
• Why self-interested agents participate in
  indirect reciprocity? Altruists do not aim at
  reciprocity, nor calculate its probability.
• If altruism spreads, donors (offspring) are
  reciprocated.
• Altruistic acts increase donors' fitness.
• Reciprocity emergent effect of altruism.

20
Group Selection Theory

• Groups are units of selection and reproduction
  (Williams, 1971 Sober and Wilson, 1999), which
  compete on the same evolutionary stage
• groups with adaptive habits produce new offspring
  groups
• groups missing adaptive habits decline till
  desegregation or extinction.
• Haystack models between-group advantage of
  cooperation Vs within-group advantage of
  defection.
• asexual reproduction with inheritance.
• new groups are formed either randomly or
  nonrandomly
• Altruism evolves (Maynard Smith, 1964 Cohen and
  Eshel, 1976) with nonrandom matching processes
  (altruists are likely to be matched with
  altruists)
• Harsch controversy on GST (Palmer, 2002 review of
  Field's 2001), due to collectivist reading.

21
Haystack Models

• Gli haystak models sono disegnati per studiare le
  circostanze in cui il vantaggio della defezione
  per gli individui può essere compensato (o
  superato) dal vantaggio della cooperazione in un
  processo di selezione fra gruppi.
• L'algoritmo di un haystack model è suddiviso in
  due fasi
• reproductive phase ogni individuo nella
  popolazione --divisa in gruppi-- si riproduce
  asessualmente, generando una copia esatta di se
  stesso.
• dispersal phase gli individui vengono isolati e
  riasseganti casualmente ad un nuovo gruppo.

22
Haystack Models

• I componenti di un gruppo possono essere
  strategicamente cooperativi o non cooperativi.
• La formazione di gruppi nella dispersal phase
  viene definita nonassortativa quando la
  distribuzione delle strategie all'interno del mio
  gruppo è indipendente dalla mia strategia.
• In un haystack model, se i raggruppamenti sono
  non assortativi ed il tasso di riproduzione è
  determinato secondo i payoff di un Prisoner's
  Dilemma, l'unico risultato possibile è una
  popolazione composta esclusivamente da strategie
  non cooperative.

23
The First Place Challenge

• Altruists must be more likely to join groups
  containing high numbers of altruists (Cooper and
  Wallace, 2001). How? Typical question posed by
  many authors e.g.
• So even though a tribe full of selfless people
  would prevail over a tribe full of selfish
  people, it is hard to see how a tribe would get
  full of selfless people in the first place. ...
  Even if you magically intervened and implanted
  sympathetic genes in 90 percent of the
  population, these would steadily lose out to
  their less ennobling rivals. (Wright 1994, page
  187)
• Selfless groups would be perpetually undermined
  by the selfishness of their individuals. (Ridley
  1997).
• Any reasonable answer?

24
Aims Of The Present Work

• Test
• RAT (donations enhance donors fitness) and
• GST (donations enhance group fitness)
• Wilikinson's analyzes all-cooperators vs
  all-defectors.
• What happens in intermediate conditions?
• Which is the minimal share of altruist?
• Does increase of survival rate correspond to
  increase donors' fitness, or is it redistributed
  over entire population? if so, are individual
  donors always refunded or do they sustain a share
  of the costs of redistribution?
• Latter question crucial if donors are not
  reciprocated in person or along their future
  generations, reason to question reciprocal
  altruism interpretation, and look for concurrent
  explanation.

25
The Simulation Model (1)

• Roost is a social space containing any number of
  bats.
• In-roosts are allowed to perform sharing food and
  grooming (no other social activity has been
  modeled).
• In one simulation tick ( 24 hours), two stages
• daily grooming and food sharing
• nightly hunting
• as in nature (Wilkinson 1990), 93 agents find
  food
• remaining 7 starve, unless helped from fellows
• No accumulation short-term consumption
• Infrequent lethal food scarcity (1.65 double
  unsuccessful hunt p. animal p. year) .
• Average lifetime around 10 years

26
The Simulation Model (2)

• Each day, agents choose grooming partners from
  roost
• If starving, grooming partner is asked for help
• Non-cheater agents give help, if full (have had
  good hunt)
• Donor agents will give away blood for 6 hs of
  their autonomy, giving 18 hs to recipients.
• No direct retaliation victims of cheating die on
  the spot.

27
Reproductive Algorithm

• In nature
• Female bats give birth to one child per time
• Reproduce every 10 months.
• Newborns leave the roost when self-sufficient,
• but they are never found in isolation chances of
  survival for a lone individual are extremely
  scarce.

• In simulated experiments
• Individuals are identical at birth and sexless
• cloning every ten months, starting from the
  twentieth month,
• at each occurrence each agent has 50 probability
  to clone
• As opposed toclassic haystack models, roosts are
  unit of selection and reproduction, like
  individual organism
• Roosts are formed when a critical mass (20
  individuals) is reached.
• Rationale Reproductive success more the
  in-roosts, higher the number of new roosts.

28
Experimental Conditions

• Control
• One-roost world
• Food-sharing always allowed
• Reproduction is possible
• Variable percentage of cheaters (never giving
  help when asked, even if full unlike altruists,
  they sustain no costs)
• No retaliation
• Cheaters are expected to prosper, reducing the
  efficiency of the system as a whole

• Experimental
• All previous conditions apply in
• Multi-roost world the same number of agents
  distributed over a variable number of roosts.
• Population growth by altruistic roost formation
• Lineage is followed
• If fitness of donors is higher than average,
  then RAT proves more valid.
• Otherwise, GST is preferable.

29
Findings (1). Single Roosts

• Cheaters cause demographic catastrophes.
• No reciprocity emerges cheaters exploit others,
  incurring neither retaliation nor isolation.
• But cheating is self-defeating in the long run
• after exploiting altruistic in-roosts to death,
• cheaters are bound to die.
• Graph
• Single roost with 300 agents for 20000 ticks.
  Red population Dark Blue n. of roosts (x10)
  Light Blue n. of cheaters (x10)

30
Findings (2). Multi-Roost World

• If altruists survive cheaters, they will
  repopulate the roost and produce new ones.
• If no altruist survives cheaters, the roost will
  extinguish.
• After a critical period in which global fitness
  declines
• altruists take off
• number of roosts grows.
• Roosting matters!
• Roosts with cheaters disappear,
• But if any roost without cheaters appears, it
  repopulate the world.
• Graph
• 300 agents in 10 roosts for 8000 ticks. Red
  population Dark Blue n. of roosts (x10) Light
  Blue n. of cheaters (x10)
• Above 10 cheaters
• Below 20 cheaters

31
Findings (3). Donors Fitness

• Subtle measure
• A subset of population altruists dying old
  (agents which reach 10 years age are removed from
  simulation).
• By tracking their descendants, the correlation
  between number of helps and size of offspring can
  be checked.
• In a set of 100 simulations,
• with 10 roosts 30 agents each,
• 10 cheaters,
• 5000 ticks,
• agents that lived to the maximum age were
  extracted, keeping track of donations and number
  of agents in the lineage up to the simulation
  end.
• A very low factor (0.07) of positive correlation
  between these measures was found.
• No linear relation between number of donations in
  life and lineage size!

32
Discussion

• Our findings support Wilkinsons
• Altruism reduces mortality even one single
  cheater may lead the roost to extinction.
• Cheaters survive longer than altruists but are
  self-defeating in the long run.
• What about roaming cheaters? In nature, newcomers
  find hard to be accepted by in-roosts an
  entrance fee must have co-evolved to protect
  altruists.

33

• Mediatori Cognitivi del
• Comportamento Altruistico

34
The Problem

• If the rationale of altruism is reciprocity
• How tell that current recipients will be future
  donors?
• Which mental construct behind reciprocity?
• Aims of the present work
• Model motivations as one main aspect of
  cognition.
• Observe their impact on altruism.

35
Why Bother With Motivations?

• Motivations are usually overlooked.
• Theories of intelligent or rational action
  emphasise the role of (limited) knowledge (Simon,
  Kahneman and Tversky, etc.).
• Evolutionary theory of mind stresses the
  representational side of cognition.
• In AI and agent systems, where autonomous systems
  are constructed for task execution, goals as
  sources and guide of action are acknowledged.
• Motivation is here meant to account for the
  strength of goals (their energetic component).

36
Two Altruistic Algorithms

• Simple, or sub-cognitive
• agents apply routines under given conditions.
• Smart, or quasi-cognitive
• agents execute actions to achieve their goals.

37
Simple and Smart Vampires

• Simple production rule
• Taking an external condition as input.
• Giving an action as output.
• Smart endowed with a social norm
• Taking an external condition as input.
• Giving a goal of variable force as output.
• Which is better, and when?

38
First Study

• Control condition
• If satiated, agents will give away blood for 6 hs
  of their autonomy, giving 18 hs to recipients.
• Otherwise help is denied.

• Exp. condition
• Agents have goals
• survival and
• normative (give help).
• with variable strength.
• Five patterns of goal relationships emerge
• NG is null cheaters.
• SG is null martyrs.
• Gs are equal fair.
• NG is stronger generous.
• SG is stronger prudent (simple-like).

39
Effetti della reciprocazione

• Dotando gli agenti della capacità di
  riconoscimento individuale, conferiamo loro la
  possibilità di non essere sfruttati dagli egoisti
• Questo meccanismo non è sufficiente per
  determinare un cambiamento nelle prestazioni
  della popolazione

40
Che cosaltro?

• Uneuristica per il cambiamento degli scopi
• La dinamica degli scopi è un aspetto fondamentale
  della cognizione (Conte, 2000) In base alle
  conoscenze possedute, gli scopi possono essere
  generati o abbandonati, per effetto di un
  cambiamento del loro valore

41
Dynamic Smart Vampires

• Goal-dynamics an essential aspect of cognition
  (Conte, 2000)
• Based on beliefs, goals are generated,
  interrupted, dropped etc. as an effect of value
  change.
• How does goal value change?
• In our simulations, NG
• Grows with donations received
• Decreases with unreciprocated donations
• Essentially, keeping constant for simplicity SG,
  the strategy played by an agent is caused by the
  dynamics of its NG.

42
Findings (1) Strategies Emerge

• With 10 roosts of 15 dynamic smart agents each
  (same conditions as in previous experiment),
  starting prudent and cheaters, 40 ys (4 gen.)
• Strategies differentiate
• Altruistic ones being fitter than cheating,
• Cheating is kept inoffensive (never grows even
  when cheaters are a majority)
• Difference increases with inheritance of
  goal-value.

43
Findings (2). Dominance of Martyrdom

• Martyrdom is dominant, because it is
  self-reinforcing
• if donations
• increase, NG grows ad libitum.
• decrease, NG cannot lower below a given
  threshold.
• with an upper limit,
• it behaves like other straegies.
• population varies periodically.

44
Findings (3). Initial Strategy Does Not Matter

• Findings are stable while initialising
  simulations with different strategies
• The most altruistic strategies are always
  dominant
• Graph
• Above all strategies present at the beginning of
  simulation
• Below martyrs cheaters at start

45
Findings (4). Retaliation Severity Matters

• With more severe retaliation (debtors are denied
  help until credits are extinguished)
• Population extinguishes
• Graph
• Above all strategies present at the beginning of
  simulation
• Below prudents cheaters at start

46
Findings (5). Controlling cheaters

• Smart agents can survive high percentages of
  cheaters
• Simple agents get usually extinct with 60
  cheaters,
• While a population starting with 40 martyrs and
  60 cheaters reaches 40 ys
• Graph
• Above simple agents, 60 cheaters
• Below martyrs cheaters at start

47
Advantages of Present Study

• Reduce arbitrariness of multi-parametered
  simulations.
• Model is anchored to real data.
• Increase plausibility of adaptive heuristics
• Unlike stupid adaptive agents, which
• Imitate fittest behaviour (how tell?)
• Imitate frequent behaviour (how tell global
  frequency?)
• Dynamic smart agents are modified by the effects
  of others actions on themselves.

48
What We Have Learned

• Trade-off between reciprocity and altruism
• retaliation reduces donations occurrence
• but, it encourages reciprocity (essential for
  altruism).
• Mechanisms ensuring reciprocity (retaliation)
  ought to be mild.
• Flexible intelligence contributes to altruism
• Does it mean vampires are smart? Probably not.
• But altruistic behaviour can be learned by
  goal-based systems in a rather plausible way.