Multi agent systems

1
Multi agentsystems

• Guido Boella
• Dipartimento di Informatica
• Università di Torino

2
Intentional stance

• According to Dennet, attitudes like belief and
  desire are folk psychology concepts that can be
  fruitfully used in explanations of rational human
  behavior. If you were asked to explain why
  someone is carrying an umbrella, you may reply
  that he believes it is going to rain and that he
  does not want to get wet. For an explanation of
  behavior it does not matter whether one actually
  possesses these mental attitudes we describe the
  behavior of an affectionate cat or an unwilling
  screw in terms of mental attitudes. Dennet calls
  treating a person o artifact as a rational agent
  the intentional stance.

3
(No Transcript)
4

• Here is how it works first you decide to treat
  the object whose behavior is to be predicted as a
  rational agent then you figure out what beliefs
  that agent ought to have, given its place in the
  world and its purpose. Then you figure out what
  desires it ought to have, on the same
  considerations, and finally you predict that this
  rational agent will act to further its goals in
  the light of its beliefs. A little practical
  reasoning from the chosen set of beliefs and
  desires will in most instances yield a decision
  about what the agent ought to do that is what
  you predict the agent will do.

5
HUMAN AGENTHOOD

• Il modello che i bdi agents vogliono imitare e
  il comportamento umano.
• In particolare il fatto che gli esseri umani
  siano guidati da tre attitudini

BELIEFS, DESIRES AND INTENTIONS CREDENZE, GOAL E
INTENZIONI
Ambiguo fra il significato di scopo e desiderio.
6
Metaphor (Lakoff)
A
B
f
x
f(x)
x y
f(x y)
f(y)
y
f A ? B f(x y)f(x) f(y)
lt B, gt
lt A, gt
7
IntenTionality

• Intentionality significa intenzionalitàcapacit
  à di avere rappresentazioni interne del mondo
• come è (credenze)
• come vorremmo che sia (desideri)
• come vogliamo cambiarlo (intenzioni)
• come dovrebbe essere (obblighi)

8
Attenzione!

• IntenSionalità è una cosa differenteintensionale
  significa che non vale il test della
  sostituibilità
• 24/2sqrt(4)
• Pierino crede che 2 sia la metà di 4
• Pierino crede che 2 sia la radice di 4

9
FMRI
10
Bratman What Is Intention?

• 2 possibili interpretazioni della parola
• Azione lha ucciso intenzionalmente
• Stato mentale ho lintenzione di andare a
  vedere Titanic
• Proprieta
• future directed ho lintenzione di andare al
  cinema ieri
• stato mentale che influenza il comportamento
  futuro
• stabili (committment) ma revocabili non posso
  prevedere cosa accadra in futuro, quindi
  devo poter cambiare idea.

11
Why bother today to have an intention about
tomorrow?

• Perche non attendere domani sera prima di
  formarmi (ora) lintenzione di andare al cinema?
• Perche sprecare risorse per formarmi delle
  intenzioni ora, se non so cosa accadra domani
  (ad es. Mi potrebbero invitare a cena)?
• Soluzione alternativa
• utilizzare un pianificatore classico e poi
  scegliere fra le
• opzioni che propone

12
Planning Practical Reasoning?

• Scelgo un obbiettivo
• Costruisco un piano completo che lo soddisfi
• Scelgo fra le alternative tramite decision
  theory (probabilita eventi utility function)
• Limiti
• devo conoscere lo stato del mondo
• piu tempo passo a pianificare e decidere,
  maggiore e la probabilita che il mondo cambi
• ad ogni cambiamento del mondo devo ricominciare
  da capo

13
RUOLO DELLE INTENZIONI

• Questi problemi sono dovuti alla resource
  boundedness il ruolo delle intenzioni e di
  limitare lo spazio di ricerca del planning e
  quello della scelta fra le opzioni.
• Per questo motivo le intenzioni
• sono future directed e (relativamente) stabili
  se stabilisco ora una linea di azione per il
  futuro, le scelte successive dovranno essere
  compatibili con questa limito le opzioni
  possibili.
• Se ho lintenzione di andare domattina a
  prendere un libro in biblioteca, non prendero in
  considerazione lazione di andare a giocare a
  tennis

14

• Riguardano piani parziali non conoscendo in
  anticipo come puo evolvere il mondo, costruisco
  dei piani che sono incompleti temporalmente (non
  stabilisco lordinamento completo delle azioni) o
  strutturalmente (non scelgo come eseguire delle
  azioni complesse). Esempio
• Posso avere lintenzione di andare a prendere un
  libro in biblioteca, senza aver necessariamente
  costruito un piano in tutti i suoi dettagli
  (posso non aver ancora scelto se andare in
  macchina o prendere il tram, perche non so se ci
  sara sciopero).
• Costituiscono linput per ulteriori fasi di
  pianificazione

15
INTENZIONI E CREDENZE

• Intendo solo cio che credo possibile (i.e.
  esiste un piano che ritengo eseguibile)
• credo che eseguiro quello che intendo se
  intendo andare al cinema domani, credo che domani
  saro al cinema e non a casa (conduct
  coordinating pro-attitude)
• la stabilita delle intenzioni implica che, se
  fallisce il mio piano, ne cerchero un altro per
  raggiungere il mio scopo
• controllo se lesecuzione di un piano ha avuto
  successo
• non intendo quelle che credo essere le
  conseguenze indesiderate delle mie azioni
  (problema dei side effects)

16
SIDE EFFECTS

• Intend(A) ? Bel(A?B)
• Intend(B)

?
Non voglio essere committed verso B altrimenti
dovrei perseguire B anche se A e gia realizzato
(e quindi le mie credenze erano
sbagliate). Esempio Strategic bomber
Intend(bombardare la fabbrica) ma
Bel(bombardare la fabbrica ?uccidere i bimbi
della scuola) Terror bomber Intend(uccidere i
bimbi della scuola)
17
ARCHITETTURA ASTRATTA
action
Plan library
Intentions structured into plans
Opportunity analyser
means-end reasoning
Filtering process
option
Desires
Beliefs
perception
Deliberation process
intentions
Reasoner
18
Cos'è un gruppo di agenti

• Un insieme di agenti BDI che hanno deciso di
  operare in maniera coordinata al raggiungimento
  di un obiettivo (goal) per mezzo di una sequenza
  di azioni (piano). Ad ogni agente è assegnata la
  responsabilità di una parte del piano.

19
I vantaggi della cooperazione fra agenti

• La cooperazione fra agenti BDI permette
• di raggiungere obiettivi altrimenti impossibili
• di raggiungere i risultati desiderati con una
  maggiore efficienza

20
... gli svantaggi della cooperazione

• L'interdipendenza fra le azioni dei diversi
  agenti di un gruppo è più difficile da gestire
• E' necessario coordinare i diversi compiti
  tramite mezzi di comunicazione
• Si rischia di lavorare a vuoto se la coesione del
  gruppo viene meno (quindi è necessario comunicare
  anche i cambiamenti di commitment)

21
Tipi di azioni

• Sollevare un tavolo
• Giocare una partita di calcio
• Preparare una cena
• Cercare un oggetto
• (fare qualsiasi cosa che un altro agente possa
  impedire)

22
Autonomous agents

• "...to pose a goal to oneself is something
  about which no external legislation can
  interfere...". An agent "cannot undergo any
  obligation other than what he gives himself on
  his own. (...) only by this means it is possible
  to reconcile this obligation (even if it were an
  external obligation) with our will".
• Kant (Die Metaphysik der Sitten, 1794)

23
Stato dell'arte sulla cooperazione

• Bratman, Castelfranchi
• Cohen e Levesque (1991) i membri di un gruppo
  comunicano i cambiamenti di commitment
• Grosz e Kraus (1996) supporto reciproco fra gli
  agenti (helpful behaviour)

24
Bratman

• His objective is to define the notion of shared
  cooperative activity in terms of individual
  mental states and their interrelationship,
  without resorting to collective form of
  intentions that go beyond the mind of individuals
  and without introducing ad hoc mental states
  characterizing cooperative behaviors
• A shared intention is not an attitude in the mind
  of some super-agent consisting literally of some
  fusion of the two agents.

25
VS

• The examples that contrast with this broadly
  individualistic" approach are Searle, 1990
  (collective intentional behavior is a primitive
  phenomenon"), Tuomela, and Gilbert, 1960 (the
  cooperating agents form a plural subject which
  is no more reducible)

26
Key features

• Mutual responsiveness Each participating agent
  attempts to be responsive to the intentions and
  actions of the other knowing that the other is
  attempting to be similarly responsible. Where
  responsiveness" means keeping an eye to the
  behavior of the other and to act on the
  expectations that an agent has on the partner's
  behavior. That is, the agents are reactive" to
  the changes in the environment caused by the acts
  of the partners.

27
Key features II

• Commitment to the joint activity The
  participants each have an appropriate commitment
  (though perhaps for different reasons) to the
  joint activity, and their mutual responsiveness
  is in the pursuit of this commitment

28
Key features III

• meshing the subplans of the different agents
  that aim at performing the joint activity must
  mesh" with each other. If they don't, the
  agents cannot opt out of the group but must work
  to achieve a meshing.

29
Key features IV

• helpful behavior the agents provide support to
  the partner without further incentives. This
  requirement prevents agents to opt out of the
  group instead of providing further help. If an
  agent can provide help without undermining his
  own contribution to the joint activity, he has to
  do so without any other request, solicitation or
  promise of reward by his partners.This new
  commitment ensures also that the partner who
  needs help does not give up his commitment
  because he is despairing to receive support from
  the others.

30
Key features V

• Limits of bargaining Though we share the
  intention to paint together we might have
  conflicting preferences about who scrapes and who
  paints, or about what color paint to use. Such
  conflicts call for bargaining in some form not
  bargaining about whether to paint together but,
  rather, bargaining about howwe are to paint
  together

31
Key features VI

• No benevolence Though we share the intention to
  paint together we might have conflicting
  preferences about who scrapes and who paints, or
  about what color paint to use. Such conflicts
  call for bargaining in some form not bargaining
  about whether to paint together but, rather,
  bargaining about howwe are to paint together

32
Delegation

• In delegation an agent needs or likes an action
  of another agent and includes the action in his
  own plan. For delegating, an agent has the
  following attitudes
• he believes that the delegated agent can do and
  will do the action
• he has the goal that the other does the action
• he relies on the action he restrains himself
  from doing it and coordinates his own behavior
  with the predicted action.

33
Castelfranchi

• Goal adoption social action is more than the
  mere coordination implicit in predicting the
  behavior of others. The leap to strong social
  action is exemplified by the situation in which
  one helps another agent in reaching his goals
  since he knows that the other has such a goal
  and, perhaps the other is dependent on him.
• an agent considers his addressees as agents
  having a mental state composed of beliefs and
  intentions.

34
Anticipatory coordination

• Given the interferences arising from sharing the
  same environment, Castelfranchi claims that
• No agents could really plan" (also partially)
  its behavior in a multi-agent world without some
  anticipatory coordination. There is a
  co-evolutionary coupling between planning in a
  multi-agent world and mind-reading ability.
• where for mind-reading ability it is intended the
  understanding of goals and plans of the other.

35
Different levels

• literal help x adopts what y has delegated to
  him.
• over-help x maintains the delegated plan but
  goes beyond it.
• critical help x modifies the plan remaining in
  the limit of the delegated goal.
• over-critical help xgoes beyond the goal and
  modifies ys plan.
• hypercritical help x adopts goals that y didn't
  considered but he gives up the delegated goal and
  plan.

36
Cohen and Levesque TEAMWORK

• Any theory of joint action should indicate when
  communication is necessary

37
INTENTIONS FOR MULTIAGENTS

• Se invece di un singolo agente si ha a che fare
  con un gruppo che deve eseguire un certo compito,
  e necessaria una nuova nozione di intenzione che
  permetta di coordinare lattivita del gruppo.
• Lintenzione resta uno stato mentale del singolo
  agente, ma include delle nuove condizioni. Cosi
  come le azioni sono sempre eseguite da un singolo
  agente.
• In particolare, per coordinare un gruppo di
  agenti e necessaria la comunicazione fra loro,
  in modo che la loro azione non prenda strade
  divergenti.
• Non e sufficiente avere lintenzione di eseguire
  la propria parte del piano e il mutual belief che
  gli altri hanno tale intenzione.

38
PERMANENT GOAL

• Achieavement goal (vs maintaineance)
• A-Goalx(P) ? Goalx(later(P)) ? Belx(?P)
• Committment (fanatical) goal permanente a certe
  condizioni
• P-Goalx(P) ? A-Goalx(P) ? Belx(?P) ?
• before(Belx(P) ? Belx(??P)),
• ?Goalx(later(P))
• Relative committment
• P-R-Goalx(P,Q) ? A-Goalx(P) ? Belx(?P) ?
• before(Belx(P) ? Belx(??P) ? Belx(?Q) ),
• ?Goalx(later(P))

39

• Esempio del convoglioCohen Levesque, 91
• A non sa la strada per andare a casa ma sa che B
  per andare a casa propria passa davanti alla sua.
• A segue B senza dirglielo ma B guida troppo
  veloce e A si perde.
• B sa che qualcuno lo seguira per andare a casa
  pero, se A si ferma per un guasto B crede che A
  sia arrivato e accellera.
• A e B sanno mutualmente A arriva a casa e si
  ferma. Credendo che ci sia un guasto B si ferma e
  aspetta
• A a meta strada capisce dove si trova e da quel
  punto puo procedere da solo ma B continua voler
  passare davanti a casa di A e a procedere non
  troppo forte.

40

• The solution of Cohen and Levesque is to propose
  a weaker notion of cooperation after one of the
  partners has dropped his private intention to do
  his part in the group, he will adopt the
  intention to make mutually known the reason why
  he discharged the intention either because it is
  impossible to reach the joint goal or because it
  has already been achieved.
• Note that, when an agent does not believe that
  the goal is achievable, the mutual beliefs of the
  group do not hold anymore. Anyway, the group will
  achieve the mutual belief about the impossibility
  of their action only when the agent communicates
  his discovery.

41
JOINT INTENTION

• Un gruppo di agenti ha un joint persistent goal
  di ottenere P relativo a Q sse
• tutti i membri del gruppo credono che P e falso
• tutti mutualmente credono che tutti hanno ?P come
  goal
• e creduto mutualmente che continueranno ad avere
  un weak achievement goal fino a che non sara
  conoscenza mutua fra loro che P e vero o che P
  e impossibile da raggiungere o che Q e
  diventato falso.
• Weak achievement goal P Un agente ha il goal ?P
  e crede che P e falso oppure crede che che P e
  vero, irrilevante o impossibile da raggiungere e
  ha il goal che tale situazione sia mutua
  conoscenza di tutti i membri del gruppo.

42
cosa manca ?
43
Tambe's STEAM System

• Tambe proposes an implemented model of
  cooperation based on the SOAR architecture which
  has as basic building blocks the notion of joint
  intention by Cohen and Levesque and the Shared
  plan operator of Grosz and Kraus

44
Tambe

• Since communication is a significant overhead
  when many agents work together and it is a risk
  in hostile environments, it must be tailored to
  the situation.
• The selectivity in communication is achieved via
  a decision-theoretic approach. By introducing the
  notion of utility, Tambe adapts the prescription
  of the Cohen and Levesque framework of
  communication, or, more precisely, the
  prescription to attain mutual belief.

45
Shared Plans of Grosz and Kraus

• Their goal is to provide a specification of the
  mental state of the participants in a
  collaborative activity that handles complex
  actions and is comprehensive in its treatment of
  partiality of belief and intention
• They define how the snapshots" of the agents'
  mental state should look like from the initial
  minimal state of having a partial shared plan to
  the one of having a full one that represents the
  stopping condition for the planning process.

46
Partial plan
47
Shared plans

• In order to have a shared plan each agent must
  have
• a mutual belief about a (partial) recipe,
• individual intentions that its own part of the
  action be done
• individual intentions that the collaborators
  succeed in doing the (identified) constituent
  subactions
• individual or collaborative plans for the
  subactions
• - collaborative plans for having a full recipe

48
Intend-that vs Intend-to

• Among the most important achievements is the
  introduction and axiomatization of the notion of
  intending a state of affairs (intend-that) while
  the standard notion of intention allows only
  intending single agent action, by means of
  intend-that agents are allowed to have an
  attitude towards actions by intending that a
  given action is successful in particular, agents
  can intend that a certain multi-agent action is
  executed successfully

49
Intend-that
50
Partial Shared Plan
51
Tuomela

• The possibility of ascribing goals, beliefs, and
  actions to collectives relies on the idea that
  collectives can be taken to resemble persons. I
  shall here accept this idea, although the analogy
  does not go very far (partly because collectives
  are not capable of performing primary actions but
  only act via their members and representatives).
  Following common-sense examples, I will accept
  ... that both factual and normative beliefs can
  be ascribed (somewhat metaphorically) to groups,
  both formal and informal, structured and
  unstructured.

52
Groups as agents
53
(No Transcript)
54
Costruction of social reality

• Searle argues that there is a distinction between
  two types of rules, a distinction which also
  holds for formal rules like those composing
  normative systemsSome rules regulate
  antecedently existing forms of behaviour. For
  example, the rules of polite table behaviour
  regulate eating, but eating exists independently
  of these rules. Some rules, on the other hand, do
  not merely regulate an antecedently existing
  activity called playing chess they, as it were,
  create the possibility of or define that
  activity. The activity of playing chess is
  constituted by action in accordance with these
  rules. Chess has no existence apart from these
  rules. The institutions of marriage, money, and
  promising are like the institutions of baseball
  and chess in that they are systems of such
  constitutive rules or conventions

55
Counts as

• According to Searle, institutional facts like
  marriage, money and private property emerge from
  an independent ontology of brute natural facts
  through constitutive rules of the form such and
  such an X counts as Y in context C" where X is
  any object satisfying certain conditions and Y is
  a label that qualifies X as being something of an
  entirely new sort. Examples of constitutive rules
  are X counts as a presiding official in a
  wedding ceremony", this bit of paper counts as
  a five euro bill" and this piece of land counts
  as somebody's private property".

56
(No Transcript)
57
(No Transcript)
58
(No Transcript)