blue evolution & learning in games econ 243b · evolution & learning in games econ 243b...
TRANSCRIPT
Evolution & Learning in GamesEcon 243B
Jean-Paul Carvalho
Lecture 7:Evolutionary Stability
1 / 21
Evolutionary Stable States (ESS)Maynard Smith and Price (1973) defined the notion of an evolutionarystable strategy as immune to invasion by mutants:
I Their focus was on monomorphic populations: every memberplays the same strategy, which can be a mixed strategy.
I We are concerned with a polymorphic population of agents eachprogrammed with a pure strategy.
I We have seen the equivalence of these two problems.
Hence we can adapt the concept of an evolutionary stable strategy toa population setting:
I The term we shall use is evolutionary stable state (ESS).
2 / 21
Invasion
Let the state be x =
x1x2. . .xn
.
Consider a game F, where F(x) =
F1(x)F2(x)
. . .Fn(x)
.
Consider an invasion of mutants who make up a fraction ε ofthe post-entry population.
The shares of each strategy in the mutant population are
represented by y =
y1y2. . .yn
.
3 / 21
Invasion
Therefore, the post-entry population state is:
xε = (1− ε)x + εy =
(1− ε)x1 + εy1(1− ε)x2 + εy2
. . .(1− ε)xn + εyn
.
The average payoff in the incumbent population in thepost-entry state is x′F
((1− ε)x + εy
).
The average payoff in the mutant population in the post-entrystate is y′F
((1− ε)x + εy
).
4 / 21
Uniform Invasion Barrier
The average payoff in the incumbent population is higher if:
(y− x)′F((1− ε)x + εy
)< 0. (1)
State x is said to admit a uniform invasion barrier if thereexists an ε > 0 such that (1) holds for all y ∈ X− {x} andε ∈ (0, ε).
That is, for all possible mutations y, as long as the mutantpopulation is less than fraction ε of the postentry population,the incumbent population receives a higher average payoff.
5 / 21
ESS
DEFINITION. State x ∈ X is an evolutionary stable state (ESS)of F if there exists a neighborhood O of x such that:
(y− x)′F(y) < 0 for all y ∈ O− {x}. (2)
In other words, if x is an ESS, then for any state y sufficientlyclose to x, a population playing x will receive a larger averagepayoff in state y than a population playing y (i.e. x is a betterreply to y than y is to itself).
Note that this considers invasions of other states y by x ratherthan invasions of x by other states. Hence it is not clear, atpresent, why this should be a stability condition.
6 / 21
ESS and Invasion Barriers
Theorem 7.1. State x ∈ X is an evolutionary stable state (ESS)if and only if it admits a uniform invasion barrier.
Thus if x is stable in the face of an arbitrarily large populationof entrants who mutate to a nearby state, then it is stable in theface of a sufficiently small population of entrants who mutateto an arbitrary state.
7 / 21
ESS and NE
What is the relationship between ESS and NE?
DEFINITION. Suppose that x ∈ X is a NE. Then (y− x)′F(x) ≤ 0for all y ∈ X.
In addition, suppose there exists a neighborhood of x that doesnot contain any other NE.
Then x is an isolated NE.
Proposition 7.2. Every ESS is an isolated NE.
8 / 21
Proof
Let x be an ESS of F, O be the nhd posited in (2) andy ∈ X− {x} (not necessarily in O).
Then for all ε > 0 sufficiently small, the postentry statexε = εy + (1− ε)x is in O.
Given x is an ESS, this implies that:
(xε − x)′F(xε) < 0(εy + (1− ε)x− x)′F(xε) < 0
ε(y− x)′F(xε) < 0(y− x)′F(xε) < 0.
(3)
9 / 21
Proof
Taking ε→ 0 yields:
(y− x)′F(x) ≤ 0,
by the continuity of F. That is, x is a NE.
To establish that x is isolated, note that if w ∈ O− {x} were aNE then (w− x)′F(w) ≥ 0, contradicting the supposition that xis an ESS [by (2)]. �
The converse of Proposition 7.2 is not true.I The mixed equilibrium of a two-strategy coordination
game is a counterexample.
10 / 21
More on ESS and Nash
Therefore, ESS is stronger than NE.
In particular, an ESS satisfies the additional property:
Suppose there exists a state y which is an alternative best replyto x, i.e. (y− x)′F(x) = 0.—Then (y− x)′F(y) < 0, i.e. x is a better reply to y than y is toitself.
Therefore:
I A strict NE is an ESS.
I A polymorphic population state (equivalent to a mixedNE) cannot be strict and hence must satisfy the additionalproperty.
11 / 21
More on ESS and Nash
In the case in which agents are matched uniformly at random toplay a normal form game (the case we have been focussing on),then it is easy to see why the additional property is required.
Suppose (y− x)′F(x) = 0, i.e. y is an alternative best reply to x.
Then:
(y− x)′F(εy + (1− ε)x) = ε(y− x)′F(y) + (1− ε) (y− x)′F(x)︸ ︷︷ ︸=0
= ε(y− x)′F(y).(4)
Therefore, (y− x)′F(y) must be negative for (1) to hold andhence, by Theorem 7.1, for x to be an ESS.
12 / 21
Example: Hawk Dove
Hawk DoveHawk −2
−20
4Dove 4
00
0
ESS: x = ( 23 ,
13 ). ESS payoff = 0.
13 / 21
Example: Hawk Dove
I Consider a mutation y such that y1 > x1 = 23 .
I Check that (y− x)′F((1− ε)x + εy
)< 0 for all such y:
(y1− x1)[−2((1− ε)x1 + εy1
)+ 4((1− ε)(1− x1)+ ε(1− y1)
)].
I This equals:
(y1 − x1)ε[−2y1 + 4(1− y1)]
because −2× 23 + 4× (1− 2
3 ) = 0. This in turn equals:
(y1 − x1)ε[4− 6y1]
which is negative because y1 > 23 by hypothesis.
I A similar argument can be applied to the case y1 < x1.Hence x is an ESS.
14 / 21
The Prisoners’ Dilemma
C DC 3
35
0D 0
51
1
NE/ESS: x = (0, 1).
Therefore, an ESS is not necessarily efficient.
15 / 21
Not Every Game has an ESS
A B CA 1
10
22
0B 2
01
10
2C 0
22
01
1
x = ( 13 ,
13 ,
13 ) is the unique NE and therefore the only possible
ESS.
16 / 21
Not Every Game has an ESS
Note that x is a polymorphic population state (equivalent to amixed strategy), so any basis vector (pure strategy) is analternative best reply to x.
Check that the additional property holds: (e1 − x)′F(e1) < 0,where e1 = (1, 0, 0), i.e. the pure-strategy A.
This is not the case: x′F(e1) = e′1F(e1) = 1.
17 / 21
The Iterated Prisoners’ Dilemma
I Two players engage in a series of PD games.
I The engagement ends after the current round withprobability δ < 1
2 . We call this the stopping probability.
I Consider a population in which three strategies arepresent:I C—always cooperate,I D—always defect,I T—tit-for-tat, i.e. start by cooperating, thenceforth
cooperate in period t if partner cooperated in t− 1.
18 / 21
Expected PayoffsWithin each pairing:
C D TC 3
δ 0 3δ
D 5δ
1δ 4 + 1
δ
T 3δ
1δ − 1 3
δ
Over all pairings:
FC(x) = (xC + xT)3δ
FD(x) = xC5δ + xD
1δ + xT
(4 + 1
δ
)FT(x) = (xC + xT)
3δ + xD(
1δ − 1)
19 / 21
All-T is not an ESS
Let x = (xD, xC, xT) = (0, 0, 1).Consider any alternative state y such that yD = 0.
(y− x)′F(y) = (0 yC yT − 1)
FD(y)FC(y)FT(y)
= yCFC(y) + (yT − 1)FT(y)= [yC + (yT − 1)] 3
δ (recall that yD = 0)
= [yC + (1− yC − 1)] 3δ (because yT = 1− yC)
= 0.
This violates (2). Hence all-T is not an ESS.This is a case of evolutionary drift.
20 / 21
Vector Field
C
D
T21 / 21