Introduction to Game Theory - UMD Department of … Normal form... · Introduction to Game Theory...
Transcript of Introduction to Game Theory - UMD Department of … Normal form... · Introduction to Game Theory...
Nau: Game Theory 1
Introduction to Game Theory
3a. More on Normal-Form Games Dana Nau
University of Maryland
Nau: Game Theory 2
More Solution Concepts Last time, we talked about several solution concepts
Pareto optimality Nash equilibrium Maximin and Minimax Dominance Rationalizability
We’ll continue with several more Trembling-hand perfect equilibrium ε-Nash equilibrium Rationalizability Evolutionarily stable strategies
Nau: Game Theory 3
Trembling-Hand Perfect Equilibrium A solution concept that’s stricter than Nash equilibrium
“Trembling hand”: Requires that the equilibrium be robust against slight errors or “trembles” by the agents
I.e., small perturbations of their strategies
Recall: A fully mixed strategy assigns every action a non-0 probability
Let S = (s1, …, sn) be a mixed strategy profile for a game G
S is a (trembling hand) perfect equilibrium if there is a sequence of fully mixed-strategy profiles S0, S1, …, that has the following properties:
lim k→∞ Sk = S
for each Sk = (s1k, …, si
k, …, snk), every strategy si
k is a best response to the strategies S−i
k
The details are complicated, and I won’t discuss them
Nau: Game Theory 4
ε-Nash Equilibrium Another solution concept
Reflects the idea that agents might not change strategies if the gain would be very small
Let ε > 0. A strategy profile S = (s1, . . . , sn ) is an ε-Nash equilibrium if, for every agent i and for all strategies siʹ′ ≠ si,
ui (si , S−i ) ≥ ui (siʹ′, S−i ) − ε ε-Nash equilibria always exist
Every Nash equilibrium is surrounded by a region of ε-Nash equilibria for any ε > 0
This concept can be computationally useful Algorithms to identify ε-Nash equilibria need consider only a finite set of
mixed-strategy profiles (not the whole continuous space) Because of finite precision, computers generally find only ε-Nash
equilibria, where ε is roughly the machine precision
Nau: Game Theory 5
Problems with ε-Nash Equilibrium For every Nash equilibrium, there are ε-Nash equilibria that approximate it, but the
converse isn’t true There are ε-Nash equilibria that aren’t close to any Nash equilibrium
Example: the game at right has just one Nash equilibrium: (D, R) We can use strategy elimination to get it:
• D dominates U for agent 1
• On removing U, R dominates L for agent 2 (D, R) is also an ε-Nash equilibrium
But there’s another ε-Nash equilibrium: (U, L) In this equilibrium, neither agent’s payoff
is within ε of the agent’s payoff in a Nash equilibrium
Problem: In the ε-Nash equilibrium (U, L), agent 1 can’t gain more than ε by deviating
But if agent 1 deviates, agent 2 can gain more than ε by best-responding to agent 1’s deviation
Nau: Game Theory 6
Problems with ε-Nash Equilibrium Some ε-Nash equilibria are very unlikely to arise
Agent 1 might not care about a gain of ε/2, but might reason as follows: • Agent 2 may expect agent 1 to to play D since it dominates U • So agent 2 is likely to play R
• If agent 2 plays R, agent 1 does much better by playing D rather than U
In general, ε-approximation is much messier in games than in optimization problems
Nau: Game Theory 7
Rationalizability A strategy is rationalizable if a perfectly rational agent could justifiably
play it against perfectly rational opponents The formal definition is complicated
Informally, a strategy for agent i is rationalizable if it’s a best response to some beliefs that agent i could have about the strategies that the other agents will take But agent i’s beliefs must take into account i’s knowledge of the
rationality of the others. This incorporates • the other agents’ knowledge of i’s rationality, • their knowledge of i’s knowledge of their rationality, • and so on ad infinitum
A rationalizable strategy profile is a strategy profile that consists only of rationalizable strategies
Nau: Game Theory 8
Example Matching Pennies Agent 1’s pure strategy Heads is rationalizable
Let’s look at the chain of beliefs
For agent 1, Heads is a best response to agent 2’s pure strategy Heads, … … and believing that 2 would also play Heads is consistent with 2’s
rationality, for the following reasons 2 could believe that 1 would play Tails, to which 2’s best response is
Heads; … … and it would be rational for 2 to believe that 1 would play Tails, for
the following reasons: • 2 could believe that 1 believed that 2 would play Tails, to which
Tails is a best response; …
–1, 1 1,–1
1,–1 –1, 1
Heads Tails
Heads
Tails
Nau: Game Theory 9
Strategies that aren’t rationalizable Prisoner’s Dilemma Strategy C isn’t rationalizable for agent 1
It isn’t a best response to any of agent 2’s strategies
The 3x3 game we used earlier M is not a rationalizable strategy for agent 1
It is a best response to one of agent 2’s strategies, namely R
But there’s no belief that agent 2 could have about agent 1’s strategy for which R would be a best response
5, 0 1, 1
3, 3 0, 5
Nau: Game Theory 10
Comments The formal definition of rationalizability is complicated because of the
infinite regress But we can say some intuitive things about rationalizable strategies
Nash equilibrium strategies are always rationalizable So the set of rationalizable strategies (and strategy profiles) is always
nonempty In two-player games, rationalizable strategies are simply those that survive
the iterated elimination of strictly dominated strategies In n-agent games, this isn’t so
Rather, rationalizable strategies are those that survive iterative removal of strategies that are never a best response to any strategy profile by the other agents
Example: the p-beauty contest
Nau: Game Theory 11
The p-Beauty Contest At the start of my first class, I asked you to do the following:
Choose a number in the range from 0 to 100 Write it on a piece of paper, along with your name In a few minutes, I’ll ask you to pass your papers to the front of the
room After class, I’ll compute the average of all of the numbers The winner(s) will be whoever chose a number that’s closest to 2/3 of
the average I’ll announce the results in a subsequent class
This game is famous among economists and game theorists It’s called the p-beauty contest I used p = 2/3
Nau: Game Theory 12
The p-Beauty Contest Recall that in n-player games,
Rationalizable strategies are those that survive iterative removal of strategies that are never a best response to any strategy profile by the other agents
In the p-beauty contest, consider the strategy profile in which everyone else chooses 100 Every number in the interval [0,100) is a best response Thus every number in the interval [0,100) is rationalizable
Nau: Game Theory 13
Nash Equilibrium for the p-Beauty Contest Iteratively eliminate dominated strategies
All numbers ≤ 100 => 2/3(average) < 67 => any strategy that includes numbers ≥ 67 isn’t a best response to any
strategy profile, so eliminate it The remaining strategies only include numbers < 67
=> for every rationalizable strategy profile, 2/3(average) < 45
=> any strategy that includes numbers ≥ 45 isn’t a best response to any strategy profile, so eliminate it
Rationalizable strategies only include numbers < 45 => for every rationalizable strategy profile, 2/3(average) < 30
. . . The only strategy profile that survives elimination of dominated strategies:
Everybody chooses 0
Therefore this is the unique Nash equilibrium
Nau: Game Theory 15
Another Example of p-Beauty Contest Results
Average = 32.93 2/3 of the average = 21.95 Winner: anonymous xx
Nau: Game Theory 16
We aren’t rational Most of you didn’t play Nash equilibrium strategies
We aren’t game-theoretically rational agents Huge literature on behavioral economics going back to about 1979
Many cases where humans (or aggregations of humans) tend to make different decisions than the game-theoretically optimal ones
Daniel Kahneman received the 2002 Nobel Prize in Economics for his work on that topic
Nau: Game Theory 17
Choosing “Irrational” Strategies
Why choose a non-equilibrium strategy?
Limitations in reasoning ability • Didn’t calculate the Nash equilibrium correctly • Don’t know how to calculate it • Don’t even know the concept
Hidden payoffs • Other things may be more important than winning
› Want to be helpful › Want to see what happens › Want to create mischief
Agent modeling (next slide)
Nau: Game Theory 18
Agent Modeling
A Nash equilibrium strategy is best for you if the other agents also use their Nash equilibrium strategies
In many cases, the other agents won’t use Nash equilibrium strategies If you can forecast their actions accurately, you may be
able to do much better than the Nash equilibrium strategy
I’ll say more about this in Session 9 Incomplete-information games
Nau: Game Theory 19
Evolutionarily Stable Strategies An evolutionarily stable strategy (ESS) is a mixed strategy that’s “resistant to
invasion” by new strategies This concept comes from evolutionary biology
Consider how various species’ relative “fitness” causes their proportions of the population to grow or shrink
For us, an organism’s fitness = its expected payoff from interacting with a random member of the population
An organism’s strategy = anything that might affect its fitness
• size, aggressiveness, sensory abilities, intelligence, …
Suppose a small population of “invaders” playing a different strategy is added to a population
The original strategy is an ESS if it gets a higher payoff against the mixture of the new and old strategies than the invaders do
Nau: Game Theory 20
Evolutionary Stability Let G be a symmetric 2-player game
Recall that the matrix shows u(r,r') = payoff for r against r'
A strategy r' invades a strategy r at level x if fraction x of the population uses r' and fraction (1–x) of the population uses r
fitness(r) = expected payoff for r against a random member of the population = (1–x)a + xb
Similarly, fitness(r') = (1–x)c + xd
r is evolutionarily stable against r' if there is an ε > 0 such that for every x < ε, fitness(r) > fitness(r')
i.e., (1–x)a + xb > (1–x)c + xd
As x → 0, (1–x)a + xb → a and (1–x)c + xd → c For sufficiently small x, the inequality holds if either a > c, or a = c and b > d
Thus r is evolutionarily stable against r' iff one of the following holds: a > c
a = c and b > d
r'
r
r' r
a
c
b
d
Nau: Game Theory 21
Evolutionary Stability More generally
We’ll use a mixed strategy s to represent a population that is composed of several different species
We’ll talk about s’s evolutionary stability against all other mixed strategies s is evolutionarily stable iff for every mixed strategy s' ≠ s,
one of the following holds:
• u(s,s) > u(s',s) • u(s,s) = u(s',s) and u(s,s') > u(s',s')
s is weakly evolutionarily stable iff for every mixed strategy sʹ′ ≠ s, one of the following holds:
• u(s,s) > u(s',s) • u(s,s) = u(s',s) and u(s,s') > u(s',s')
Includes cases where the original strategy and invading strategy have the same fitness, so the population with the invading strategy neither grows nor shrinks
r'
r
r' r
a
c
b
d
Nau: Game Theory 22
Example
The Hawk-Dove game 2 animals contend for a piece of food The animals are chosen at random from the entire population
• Each animal may be either a hawk (H) or a dove (D) The prize is worth 6 to each Fighting costs each 5
When a hawk meets a dove, the hawk gets the prize without a fight: payoffs 6, 0
When 2 doves meet, they split the prize without a fight: payoff 3, 3 When 2 hawks meet, they fight (–5 for each), each with a 50% chance
of getting the prize ((0.5)(6) = 3): payoffs –2,–2 It’s easy to show that this game has a unique Nash equilibrium (s, s),
where s = (3/5, 2/5) i.e., 60% hawks, 40% doves
Nau: Game Theory 23
Example To confirm that s is also an ESS, show that, for all sʹ′ ≠ s,
u1(s, sʹ′) = u1(sʹ′, s) and u1(s, sʹ′) > u1(sʹ′, sʹ′)
u1(s,sʹ′) = u1(sʹ′,s) is true of any mixed strategy equilibrium with full support
To show u1(s,sʹ′) > u1(sʹ′,sʹ′), find the sʹ′ that minimizes f (sʹ′) = u1(s,sʹ′) − u1(sʹ′,sʹ′)
s = “play H with probability 3/5, D with probability 2/5”
sʹ′ = “play H with probability p, D with probability 1–p”
u1(s,s') = (3/5)[–2p + 6(1–p)] + (2/5)[0p + 3(1–p)]
u1(s',s') = p[–2p + 6(1–p)] + (1–p)[0p + 3(1–p)]
so f (sʹ′) = u1(s,sʹ′) − u1(sʹ′,sʹ′) is quadratic in p
Set d f(s')/d p = 0, solve for p => p = 3/5
• So the unique minimum occurs when sʹ′ = s
Nau: Game Theory 24
Evolutionary Stability and Nash Equilibria Theorem. Let G be a symmetric 2-player game, and s be a mixed strategy. If s
is an evolutionarily stable strategy, then (s, s) is a Nash equilibrium of G.
Proof. By definition, an ESS s must satisfy u(s,s) ≥ u(sʹ′,s), i.e., s is a best response to itself, so it must be a Nash equilibrium.
Theorem. Let G be a symmetric 2-player game, and s be a mixed strategy. If (s,s) is a strict Nash equilibrium of G, then s is an evolutionarily stable strategy.
Proof. If (s,s) is a strict Nash equilibrium, then u(s,s) > u(sʹ′,s).
This satisfies the first of the two alternative criteria of an ESS