Gaussian Cheap Talk Game with Quadratic Cost Functions: When

Gaussian Cheap Talk Game with Quadratic Cost
Functions:
When Herding between Strategic Senders Is a Virtue
Farhad Farokhi? , André Teixeira? , and Cédric Langbort†
?
†
KTH Royal Institute of Technology, Sweden
University of Illinois Urbana-Champaign, USA
American Control Conference
Thursday June 5, 2014
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
1 / 14
Crowd-Sourcing Estimation
Crowd-sourcing estimation:
- Indirect: Use their devices, e.g., Mobile Millennium;
- Direct: Ask them to report, e.g., Waze.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
2 / 14
Crowd-Sourcing Estimation
What if I intentionally
under-estimate?
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
2 / 14
Crowd-Sourcing Estimation
What if I intentionally
under-estimate?
What if I intentionally
over-estimate?
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
2 / 14
Cheap Talk Game
Cheap Talk Game
A game in which better informed senders are communicating with a
receiver, who ultimately takes a decision regarding the social welfare (e.g.,
negotiations in organizations, voting in subcommittees in congress, etc).
[Crawford & Sobel, 82]
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
3 / 14
Cheap Talk Game
Cheap Talk Game
A game in which better informed senders are communicating with a
receiver, who ultimately takes a decision regarding the social welfare (e.g.,
negotiations in organizations, voting in subcommittees in congress, etc).
[Crawford & Sobel, 82]
In our example:
• Better informed senders: Crowd;
• Receiver: Traffic estimation application (e.g., Waze);
• Decision regarding the social welfare: Traffic estimate.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
3 / 14
Quadratic Gaussian Cheap Talk Game
θ1
S1
θ2
S2
x
..
θN
y1
y2
yn
R
x̂((yi )N
i=1 )
SN
• x ∼ N (0, Vxx );
• Receiver cost: E{kx − x̂k2 }.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
4 / 14
Quadratic Gaussian Cheap Talk Game
θ1
S1
θ2
S2
x
..
θN
y1
y2
yn
R
x̂((yi )N
i=1 )
SN
At the first step, we deploy N strategic sensors:
• Sensor i cost: E{k(x + θi ) − x̂k2 };
• Sensor i has perfect measurements of x, θi (nothing about others);
• θ = (θi )N
i=1 ∼ N (0, Vθθ ).
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
4 / 14
Quadratic Gaussian Cheap Talk Game
θ1
S1
θ2
S2
x
..
θN
y1
y2
yn
R
x̂((yi )N
i=1 )
SN
At the second step, sensors transmit scalar signals:
>
• yi = γi (x, θi ) ∈ R where γi (x, θi ) = a>
i x + bi θi + vi ;
• vi ∼ N (0, Vvi vi );
• The set of such mappings is Γi (isomorph to Rnx × Rnx × R≥0 ).
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
4 / 14
Quadratic Gaussian Cheap Talk Game
θ1
S1
θ2
S2
x
..
θN
y1
y2
yn
R
x̂((yi )N
i=1 )
SN
At the third step, the receiver announces its estimate:
• x̂ = x̂(y1 , . . . , yn ) where x̂ ∈ Ψ;
• Ψ is the set of all Lebesgue-measurable functions from RN to Rnx .
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
4 / 14
Quadratic Gaussian Cheap Talk Game
θ1
S1
θ2
S2
x
..
θN
y1
y2
yn
R
x̂((yi )N
i=1 )
SN
At the fourth step, the cost functions are realized:
• Receiver: E{kx − x̂k2 };
• Sensor i: E{k(x + θi ) − x̂k2 }.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
4 / 14
Independent Senders
Equilibrium
A tuple (x̂∗ , (γi∗ )N
i=1 ) ∈ Ψ × Γ1 × · · · × ΓN constitutes an equilibrium in
affine strategies if
x̂∗ ∈ arg min E{kx − x̂((γj∗ (x, θj ))N
j=1 )k2 },
x̂∈Ψ
γi∗
∈ arg min E{k(x + θi ) − x̂(γi (x, θi ), (γj∗ (x, θj ))j6=i )k2 }, ∀i.
γi ∈Γi
As always, equilibrium is tuple of actions (i.e., policies) in which no one
(i.e., the receiver and the sensors) can gain by unilaterally deviating from
her action.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
5 / 14
Do we have an equilibrium?
Theorem
Assume that Vxθ = 0, Vθi θi = Vθθ , and Vθi θj = 0 for j 6= i. There exists a
symmetric equilibrium in affine strategies where the receiver follows
x̂∗ (y) = E{x|(y1 + · · · + yN )/N }
and sender Si , 1 ≤ i ≤ N , employs the affine policy
γ ∗ (x, θi ) = a∗> x + b∗> θi ,
where
b∗
a∗
s
=
1
1 + (N − 1)ξ1> ξ1
"
−1/2
N Vθθ
0
0
−1/2
Vxx
#
ξ
>
and ξ = ξ1> ξ2>
is the normalized eigenvector (i.e., kξk2 = 1)
corresponding to the smallest eigenvalue of the matrix
"
#
−1/2 −1/2
0
−Vθθ Vxx
.
−1/2 −1/2
−Vxx Vθθ
−Vxx
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
6 / 14
How good is the equilibrium?
Corollary
Assume that Vxθ = 0, Vθi θi = Vθθ , and Vθi θj = 0 for j 6= i. At the
presented symmetric equilibrium
E{kx − x̂∗ ((γ ∗ (x, θj ))N
j=1 )k2 } = Vxx −
1
U,
α + βN
where α, β ∈ R≥0 and U ∈ Rnx ×nx such that 0 < U ≤ (α + β)Vxx .
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
7 / 14
How good is the equilibrium?
Corollary
Assume that Vxθ = 0, Vθi θi = Vθθ , and Vθi θj = 0 for j 6= i. At the
presented symmetric equilibrium
E{kx − x̂∗ ((γ ∗ (x, θj ))N
j=1 )k2 } = Vxx −
1
U,
α + βN
where α, β ∈ R≥0 and U ∈ Rnx ×nx such that 0 < U ≤ (α + β)Vxx .
• Sadly, this implies that
lim E{kx − x̂∗ ((γ ∗ (x, θj ))N
j=1 )k2 } = Vxx .
N →∞
• This equilibrium is not good for anyone! It is even worse for the
sensors in comparison to the receiver.
lim E{kx + θi − x̂∗ ((γ ∗ (x, θj ))N
j=1 )k2 } = Vxx + Vθθ .
N →∞
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
7 / 14
What went wrong?
• All the sensors are strategic (benevolent users);
• All the sensors measure x perfectly (looking vs. measuring);
• There is no correlation between the private information (shopping
street vs. residential area);
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
8 / 14
What went wrong?
• All the sensors are strategic (benevolent users);
• All the sensors measure x perfectly (looking vs. measuring);
• There is no correlation between the private information (shopping
street vs. residential area);
It is not all doom and gloom!
• We are dealing with Humans and not Econs (bounded rationality,
intuition, etc);
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
8 / 14
Herding Senders
Herding Equilibrium
A tuple (x̂∗ , γ ∗ ) ∈ Ψ × Γ constitutes a herding equilibrium in affine
strategies if
x̂∗ ∈ arg min E{kx − x̂((γ ∗ (x, θj ))N
j=1 )k2 },
x̂∈Ψ
∗
γ ∈ arg min E{k(x + θi ) − x̂(γ(x, θi ), (γ(x, θj ))j6=i )k2 }, ∀i.
γ∈Γ
As opposed to before, the senders are constrained to imitate each other.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
9 / 14
Do we have a herding equilibrium?
Theorem
Assume that nyi = 1 for all i, Vxθ = 0, and Vθθ = 0. There exists a herding
equilibrium in affine strategies where the receiver follows
x̂∗ (y) = E{x|(y1 + · · · + yN )/N }
and sender Si , 1 ≤ i ≤ N , employs a linear policy
γ ∗ (x, θi ) = a∗> x + b∗> θi
where
b∗
a∗
" √
=
−1/2
N Vθθ
0
0
−1/2
Vxx
#
ζ,
and ζ is the normalized eigenvector (i.e., kζk2 = 1) corresponding to the smallest
eigenvalue of the matrix
#
"
−1/2 −1/2
0
− √1N Vθθ Vxx
.
−1/2 −1/2
− √1N Vxx Vθθ
−Vxx
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
10 / 14
When herding becomes a virtue
Proposition
Assume that Vxθ = 0, Vθi θi = Vθθ , and Vθi θj = 0 for j 6= i. At a herding
equilibrium
lim E{kx − x̂∗ ((γ ∗ (x, θj ))N
j=1 )k2 } = 0.
N →∞
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
11 / 14
When herding becomes a virtue
Proposition
Assume that Vxθ = 0, Vθi θi = Vθθ , and Vθi θj = 0 for j 6= i. At a herding
equilibrium
lim E{kx − x̂∗ ((γ ∗ (x, θj ))N
j=1 )k2 } = 0.
N →∞
Actually, the rate of converge is faster than employing nonstrategic sensors
with measurement noise!
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
11 / 14
Numerical Example
Let us fix nx = 1 and nyi = 1 for all i = 1, . . . , N and assume that
Vxθ = 0, Vθθ = 1, and Vxx = 1.
• Independent Senders
Receiver’s cost:
0.2763N
,
E{kx − x̂(y)k22 } =
0.7236 + 0.2763N
• Herding Senders
Receiver’s cost:
2
p
E{kx − x̂(y)k22 } =
2
N + N ( N (N + 4) + 2) + 2
• Nonstrategic Senders
Senders transmit yi = x + ui where ui are i.i.d. zero-mean Gaussian
random variables so that E{u2i } = σ.
Receiver’s cost:
σ
E{kx − x̂(y)k22 } =
σ+N
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
12 / 14
Numerical Example
1
10
0
10
Estimation Error
−1
10
−2
10
−3
10
−4
10
−5
10
−6
Independent Senders
Herding Senders
Nonstrategic Senders
10
−7
10
0
10
1
2
10
10
3
10
N
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
13 / 14
Conclusion and Future Works
Conclusions
• Including strategic sensors in networked estimation;
• Characterized an equilibrium and studied its efficiency;
• “Better to employ a handful of naively-strategic but accurate sensors
than many nonstrategic but noisy sensors”.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
14 / 14
Conclusion and Future Works
Conclusions
• Including strategic sensors in networked estimation;
• Characterized an equilibrium and studied its efficiency;
• “Better to employ a handful of naively-strategic but accurate sensors
than many nonstrategic but noisy sensors”.
Future Work
• Asynchronous communication;
• Arbitrary communication graphs;
• Dynamic or repeated cheap talk game.
Cédric Langbort (UIUC)
Gaussian Cheap Talk Game ...
Thursday June 5, 2014
14 / 14