Knowledge and action in groups

Rohit Parikh

CLMPS
Augustr 6, 2015
Theorem 1 is joint with Çağıl Taşdemir
and Andreas Witzel

Theorem 2 (actually a corollary) is joint with Çağıl

Taşdemir
The last result generalizes an earlier joint result with Walter Dean.
The Social Level

Sometimes a social planner wants to achieve some social goal.

He needs to be aware that the plan can only succeed if the various
individuals act in a manner which is consonant with the plan and
that they will act in a certain way depending on their beliefs, on
their desires and on their possibilities of action.
The Social Level

Sometimes a social planner wants to achieve some social goal.

He needs to be aware that the plan can only succeed if the various
individuals act in a manner which is consonant with the plan and
that they will act in a certain way depending on their beliefs, on
their desires and on their possibilities of action.
So their actions can be influenced, either by influencing their
beliefs or their possibilities of action, or their desires (the last is
harder but can be achieved via advertising or brain washing).
El Farol Bar problem - the Santa Fe Bar probblem

The bar El Farol is located in Santa Fe, New Mexico

This problem was created in 1994 by W. Brian Arthur.
The problem is as follows: There is a particular, finite population
of people. Every Thursday night, all of these people want to go to
the El Farol Bar. However, the El Farol is quite small, and it’s no
fun to go there if it’s too crowded. So much so, in fact, that the
preferences of the population can be described as follows:
I If less than 60% of the population go to the bar, they’ll all
have a better time than if they stayed at home.
I If more than 60% of the population go to the bar, they’ll all
have a worse time than if they stayed at home.
Suppose there is a theory T which everyone has in possession of
and which allows them to make predictions as to whether the bar
will be crowded on the next Thursday. The theory takes previous
attendance as input and returns a prediction.
Suppose there is a theory T which everyone has in possession of
and which allows them to make predictions as to whether the bar
will be crowded on the next Thursday. The theory takes previous
attendance as input and returns a prediction.
Paradox If everyone has the same theory and uses the theory to
decide whether to go on the next Thursday then the theory will be
false.
If the theory predicts that the bar will be crowded, no one will go
and the bar will be empty.
If the theory predicts that the bar will not be crowded, then
everyone will go and the bar will be crowded.
A real life problem can arise if a highway which can handle 1,000
cars per hour is under repair and two alternate routes A and B,
each of which can handle 550 cars per hour are available.
A real life problem can arise if a highway which can handle 1,000
cars per hour is under repair and two alternate routes A and B,
each of which can handle 550 cars per hour are available.
What should one advise motorists to do?
If we advise motorists to take A, it will have a traffic jam. Ditto if
we advise them to take B.
A real life problem can arise if a highway which can handle 1,000
cars per hour is under repair and two alternate routes A and B,
each of which can handle 550 cars per hour are available.
What should one advise motorists to do?
If we advise motorists to take A, it will have a traffic jam. Ditto if
we advise them to take B.
During a gas shortge many years ago, the state of Massachusetts
solved this problem by asking motorists with odd numbered plates
to fill their cars on M, W, F and even number cars to fill on T, Th,
Sa.
Epistemic Reasoning
Inducing Beliefs: Shakespeare’s Much ado about Nothing

At Messina, a messenger brings news that Don Pedro, a Spanish

prince from Aragon, and his officers, Claudio and Benedick, have
returned from a successful battle. Leonato, the governor of
Messina, welcomes the messenger and announces that Don Pedro
and his men will stay for a month.
Inducing Beliefs: Shakespeare’s Much ado about Nothing

At Messina, a messenger brings news that Don Pedro, a Spanish

prince from Aragon, and his officers, Claudio and Benedick, have
returned from a successful battle. Leonato, the governor of
Messina, welcomes the messenger and announces that Don Pedro
and his men will stay for a month.
Beatrice, Leonato’s niece, asks the messenger about Benedick, and
makes sarcastic remarks about his ineptitude as a soldier. Leonato
explains that “There is a kind of merry war betwixt Signior
Benedick and her.”
Inducing Beliefs: Shakespeare’s Much ado about Nothing

At Messina, a messenger brings news that Don Pedro, a Spanish

prince from Aragon, and his officers, Claudio and Benedick, have
returned from a successful battle. Leonato, the governor of
Messina, welcomes the messenger and announces that Don Pedro
and his men will stay for a month.
Beatrice, Leonato’s niece, asks the messenger about Benedick, and
makes sarcastic remarks about his ineptitude as a soldier. Leonato
explains that “There is a kind of merry war betwixt Signior
Benedick and her.”
Various events take place and Claudio wins the hand in marriage of
Hero, Leonato’s only daughter and the wedding is to take place in
a week.
Don Pedro and his men, bored at the prospect of waiting a week
for the wedding, hatch a plan to matchmake between Beatrice and
Benedick who inwardly love each other but outwardly display
contempt for each other.
Don Pedro and his men, bored at the prospect of waiting a week
for the wedding, hatch a plan to matchmake between Beatrice and
Benedick who inwardly love each other but outwardly display
contempt for each other.
According to this strategem, the men led by Don Pedro proclaim
Beatrice’s love for Benedick while knowing he is eavesdropping on
their conversation. Thus we have, using b for Benedick, d for Don
Pedro and E for the event of eavesdropping,
Kb (E ), Kd (E ) and ¬Kb (Kd (E ))
All these conditions are essential and of course the plot would be
spoiled if we had Kb (Kd (E )) instead of ¬Kb (Kd (E )). Benedick
would be suspicious and would not credit the conversation.
The women led by Hero carry on a similar charade for Beatrice.
Beatrice and Benedick, are now convinced that their own love is
returned, and hence decide to requite the love of the other.
The play ends with all four lovers getting married.
 Benedick’s Decision problem

love nolove
propose 100 −20
nopropose −10 0

Here love means “Beatrice loves me” and nolove the other
possibility.
 The economic problem of society is (thus) not merely
a problem of how to allocate “given” resources - if
“given’ is taken to mean given to a single mind which
deliberately solves the problem set by these “data.” It is
rather a problem of how to secure the best use of
resources known to any of the members of society, for
ends whose relative importance only these individuals
know.
F. Hayek, 1945
Applications of Epistemic Reasoning to Society
Formalism

We create a language to talk about various knowledge properties in

the following way.
I An atomic predicate P is a formula
I If A, B are formulas then so are ¬A and A ∧ B
I If A is a formula and i is an agent then Ki (A) is a formula
I We may also include formulas C (A) if we wish to denote
common knowledge
Intuition

Intuitively Ki (A) means that the agent i knows the fact expressed
by the formula A. Kj Ki (A) means that j knows that i knows A.
If i, j are the only agents, then C (A) means that i knows A, j
knows that i knows A, j knows that j knows that i knows A and so
on forever.
For example suppose Ravi and Usha are playing cards. There is a
mirror behind Usha so Ravi can see her cards. But Usha does not
know this.
For example suppose Ravi and Usha are playing cards. There is a
mirror behind Usha so Ravi can see her cards. But Usha does not
know this.
And there is a mirror behind Ravi, and Usha can see Ravi’s cards.
But Ravi only knows about Usha’s mirror. And Usha only knows
about Ravi’s mirror.
For example suppose Ravi and Usha are playing cards. There is a
mirror behind Usha so Ravi can see her cards. But Usha does not
know this.
And there is a mirror behind Ravi, and Usha can see Ravi’s cards.
But Ravi only knows about Usha’s mirror. And Usha only knows
about Ravi’s mirror.
Suppose Ravi has the Queen of spades. Let Q represent this fact.
Then we have Kr (Q), of course, Ku (Q), and naturally Ku Kr (Q).
For example suppose Ravi and Usha are playing cards. There is a
mirror behind Usha so Ravi can see her cards. But Usha does not
know this.
And there is a mirror behind Ravi, and Usha can see Ravi’s cards.
But Ravi only knows about Usha’s mirror. And Usha only knows
about Ravi’s mirror.
Suppose Ravi has the Queen of spades. Let Q represent this fact.
Then we have Kr (Q), of course, Ku (Q), and naturally Ku Kr (Q).
But we do not have Kr Ku (Q).
Kripke structures

Kripke structures are used to interpret the language above.

Kripke structure M for knowledge for n knowers consists of a space
W of states and for each knower i a relation Ri ⊆ W × W .
There is a map π from W × A −→ {0, 1} which decides the truth
value of atomic formulas at each state.
Kripke structures

Kripke structures are used to interpret the language above.

Kripke structure M for knowledge for n knowers consists of a space
W of states and for each knower i a relation Ri ⊆ W × W .
There is a map π from W × A −→ {0, 1} which decides the truth
value of atomic formulas at each state.
Usually the Ri are taken to be equivalence relations, i.e., reflexive,
symmetric and transitive.
We now define the truth values of formulas as follows:
1. M, w |= P iff π(w , P) = 1
2. M, w |= ¬A iff M, w 6|= A
3. M, w |= A ∧ B iff M, w |= A and M, w |= B
4. M, w |= Ki (A) iff (∀t)(wRi t → M, t |= A)
Ki (A) holds at w , (i knows A at w ) iff A holds at all states t
which are Ri accessible from w .
Some Consequences

If Ri is reflexive then we will get Ki (A) → A (veridicality) as a

consequence.
Moreover, regardless of the properties of Ri , we have,
1. If A is logically valid, then A is known
2. If A and A → B are known, then so is B
Some Consequences

If Ri is reflexive then we will get Ki (A) → A (veridicality) as a

consequence.
Moreover, regardless of the properties of Ri , we have,
1. If A is logically valid, then A is known
2. If A and A → B are known, then so is B

This is the well known problem of Logical Omniscience since we

are attributing knowledge properties to agents which they do not
actually have.
Some Consequences

If Ri is reflexive then we will get Ki (A) → A (veridicality) as a

consequence.
Moreover, regardless of the properties of Ri , we have,
1. If A is logically valid, then A is known
2. If A and A → B are known, then so is B

This is the well known problem of Logical Omniscience since we

are attributing knowledge properties to agents which they do not
actually have.
Still, in small settings, such assumptions are reasonable.
Axiom system

1. All tautologies of the propositional calculus

2. Ki (A → B) → (Ki (A) → Ki (B)
3. Ki (A) → A
4. Ki (A) → Ki Ki (A)
5. ¬Ki (A) → Ki (¬Ki (A))

Some of these axioms are controversial

but we will not discuss the controversy.
There are also two rules of inference. Modus Ponens, to infer B
from A and A → B. And the other is generalization, to infer Ki (A)
from A.
There are also two rules of inference. Modus Ponens, to infer B
from A and A → B. And the other is generalization, to infer Ki (A)
from A.
The second rule does not say that if A is true than i knows it.
Only that it A is a logical truth then i knows it.
These rules are complete. All valid formulas are provable using the
axioms and rules.
Reivising Kripke structures when an announcement is made

Suppose we are given a Kripke structure M. Then some formula ϕ

is announced publicly.
Reivising Kripke structures when an announcement is made

Suppose we are given a Kripke structure M. Then some formula ϕ

is announced publicly.

The new Kripke structure is then obtained by deleting all states in

M where ϕ did not hold.
Theory of Mind

A group of children are told the following story:

Maxi goes out shopping with his mother and when they come
back, Maxi helps mother put away the groceries, which include
chocolate. There are two cupboards, red and blue. Maxi puts the
chocolate in the red cupboard and goes out to play.
Theory of Mind

A group of children are told the following story:

Maxi goes out shopping with his mother and when they come
back, Maxi helps mother put away the groceries, which include
chocolate. There are two cupboards, red and blue. Maxi puts the
chocolate in the red cupboard and goes out to play.

While Maxi is gone, mother takes the chocolate out of the red
cupboard, uses some of it to bake a cake, and then puts the rest in
the blue cupboard.
Now Maxi comes back from play and wants the chocolate
Now Maxi comes back from play and wants the chocolate
Where will Maxi look for the chocolate?
Now Maxi comes back from play and wants the chocolate
Where will Maxi look for the chocolate?
Children at the age of five or more say, In the red cupboard.
Now Maxi comes back from play and wants the chocolate
Where will Maxi look for the chocolate?
Children at the age of five or more say, In the red cupboard.
But children up to the age of three or four say, Maxi will look in
the blue cupboard.
Now Maxi comes back from play and wants the chocolate
Where will Maxi look for the chocolate?
Children at the age of five or more say, In the red cupboard.
But children up to the age of three or four say, Maxi will look in
the blue cupboard.
What three year old children lack, according to psychologists
Premack and Woodruff is a Theory of Mind
Knowledge leads to action

So far a great deal of work has been done on reasoning about

knowledge, studying how knowledge is transferred and
updated, and even some paradoxical theories of knowledge.
Knowledge leads to action

So far a great deal of work has been done on reasoning about

knowledge, studying how knowledge is transferred and
updated, and even some paradoxical theories of knowledge.
But we do not acquire knowledge, or transmit it to others just
to pass the time.
Knowledge leads to action

So far a great deal of work has been done on reasoning about

knowledge, studying how knowledge is transferred and
updated, and even some paradoxical theories of knowledge.
But we do not acquire knowledge, or transmit it to others just
to pass the time.
Knowledge means influence and power.
It is a commonplace that what we do depends on what we
know. And given that most of us have at least the rudiments
of a theory of mind (cf. Premack and Woodruff) we also know
that what others do will depend on what they know.
Is Knowledge always beneficial?

Kamien, Tauman and Zamir consider the following example. A

black or white card is chosen from a deck and player 1 is invited to
guess its color. After 1 makes her choice, which is announced,
player 2 is invited to make a choice. The payoffs are as follows:
I If both players guess correctly, then both get 2.
I If neither player guesses correctly, then both get 0.
I If only one player guesses correctly, then the correct player
gets 5 and the other player gets 0.
 I Suppose neither player knows the color then player 1 should
choose randomly, player 2 should choose a different color and
the expected payoff for both is 2.5 (half of 5+0)
I If player 1 is allowed to see the card, then the dominant
strategy for her is to announce the correct color, player 2
should choose the same color and the expected payoff for
both (a certain payoff in fact) is 2.
So the knowledge of player 1 makes her worse off.
However, player 1 is not harmed by the fact that she knows the
color but by the fact that player 2 knows that 1 knows the color.
Neyman shows that if we can make one player know more but
prevent other players from having more knowledge, then the player
who knows more cannot lose.
We now consider how the same game may be played differently
depending on the information available to the agents and their
temperaments.
We now consider how the same game may be played differently
depending on the information available to the agents and their
temperaments.
If there is a knowledge manipulator - KM who can control how
much information the various agents can have then that agent can
influence the way the game is played and the outcome.
Wife and husband
In the last figure we assume that the wife moves first and the
husband after.
In the last figure we assume that the wife moves first and the
husband after.
We consider various scenarios involving the husband’s
knowledge and temperament. We assume that the wife knows
the husband’s payoffs and temperament and he does not know
hers.
Case 1) Husband does not know wife’s move (and she knows this).
a) He is aggressive. Then being aggressive, he will choose S
(Stravinsky) for his move since the highest possible payoff is 3.
Anticipating his move, she will also choose S, and they will end up
with payoffs of (2,3).
b) If the husband is conservative, then not knowing what his wife
chose, he will choose B since the minimum payoff of 1 is better
than the minimum payoff of 0. Anticipating this, the wife will also
choose B and they will end up with (3,2).
2) Finally if the husband will know what node he will be at, then
the wife will choose B, the husband will also choose B and they
will end up at (3,2).
We consider now the question of how KM can create these various
knowledge scenarios of the last example.
KM is capable of creating all these three situations by means of
signals, as well as the one we did not mention where the husband
does not know but the wife does not know that he will not.
For case 1a), s(H − l) = (l, a) and s(H − r ) = (r , a). The wife
knows (if she did not already) which node they are at, but the
husband will not.
For case 2, s(H − l) = (l, l) and s(H − r ) = (r , r ). Both will know
which node they are at.
Finally if KM wants the wife to be in doubt whether the husband
knows, he could make s(H − l) = {(l, l), (l, a)} and
s(H − r ) = {(r , r ), (r , a)}. Then if the wife chose left and receives
an l, she will not know if the husband got an l or the neutral a. If
KM does send (l, l) then the husband will know, but will also know
that his wife did not know whether he would know.
Theorem
Any knowledge situation represented by a finite Kripke structure M
can be created in a single signaling step.
Proof

The knowledge manipulator (KM) picks a world w in M and sends

player i the signal (M, Xiw ) where Xiw = {v |wRi v } and Ri is the
accessibility relation of player i. This tells us the local history of i.
The global history Hw is (M, w , X1w , ..., Xnw ). An atomic formula p
holds at Hw iff it holds at w . We now show by induction on the
complexity of the formula F that M, w |= F iff Hw |= F where we
use the [?] history-based semantics to define Hw |= F
We have already noticed that atomic formulas behave correctly (by
stipulation) and truth functions are clear. Consider F = Ki (B).
 I M, w |= Ki (B) iff
I (∀v )(wRi v → M, v |= B) iff
I (IH) (∀v )(wRi v → Hv |= B) iff
I (∀Hv )(di (Hv ) = di (Hw ) → Hv |= B)
I iff Hw |= Ki (B)
We recall that di is the projection for i of a global history, and two
histories Hw , Hv here have the same projection for i iff Xiw = Xiv
iff wRi v 2
Election campaigning by a candidate

This section presumes that the candidate has made some

statements in the past and as a consequence the voters can
assume that if she is elected then the state of the world (or nation)
will belong to a certain set Z of (possible) states, those which
agree with the statements she has made.1

1
If the statements which she has already made constitute a set T , then Z is
the set of those states ω which satisfy T .
A voter may have different views about the different states in Z
finding some good and others not so good. Also, different voters
may have different views about the same state.
If the candidate reveals more of her views on some issues, then the
set of states compatible with her views will shrink. Some voters
may be displeased, finding that their favorite states are no longer
compatible with her new stance. Other voters may be pleased
seeing that some states they feared are no longer in the running.
How should the candidate speak so as to improve her overall
position given these forces pulling in different directions?
Let V denote the set of voters, and Ω the set of the states of the
world. Both V , Ω are assumed to be compact subsets of some
Euclidean space. Let the satisfaction function s : V × Ω 7→ R
represent the extent to which voter v ∈ V likes the state ω ∈ Ω.
What is the candidate’s current average degree of satisfaction
among all voters? It is
R R
ω∈Z v ∈V s(v , ω)dv dω
α= (1)
µ(Z )
namely the average
R value of s(v , ω) over all voters v and all states
ω in Z . Here v ∈V s(v , ω)dv is the extent to which a particular
world ω is liked by the average voter. A is the average of such
averages over all worlds in Z .
The candidate is now wondering whether she should make a
statement A or its negation ¬A. At the moment we are assuming
that she has no restrictions as to what she can say with popularity
being her only concern so she is free to say A or its negation. Later
on we will consider restrictions on what she can say.
Let X and Y be the two disjoint subsets of Z where X is the set of
states where A is true and Y the set where A is false, ie. where
¬A is true. Z is X ∪ Y .
Then the average satisfaction on X ∪ Y could be rewritten as
R R
ω∈X ∪Y v ∈V s(v , ω)dv dω
α= (2)
µ(X ∪ Y )

where µ(X ∪ Y ) is the probability that the event X ∪ Y occurs.

Note: Here an event is simply a set of states.
If we write the values as α(X ), α(Y ), α(X ∪ Y ) then α(X ∪ Y ) is
a convex function of α(X ), α(Y ).
At least one of these two values is at least as great as α(X ∪ Y ).
Thus at least one of the statements A and ¬A will either benefit
her (raise her level of satisfaction) or at least leave her level the
same. 2
Corollary: A candidate is best off being as explicit as she can.
Proof: For consider all possible theories (sets of statements) which
she could utter. And let T be the best theory she could utter. If T
is not complete, i.e. leaves some question A open, then there is an
extension of T which includes either A or its negation and is no
worse than T . 2.