Professional Documents
Culture Documents
by Meir Retter
August 2017
1 Our Story
You are a traveler who needs to get to the other side of a canyon. There are two
ways you can get across: you can take the rickety bridge across the canyon, or
go around the canyon using the safe road. The bridge is dangerous and if you
choose to take it then you will have only an r probability (0 r 1) of making
it across without plummeting down to your death. That is, r is the safety of
the bridge. The road has no such hazard associated with it.
So why not just take the road? Well, there is a complication. There is a
hitman who wants you dead, and he has a killer robot at his disposal. Just as
you must choose whether to take the bridge or the road, the hitman will choose
one of these paths along which he will send his killer robot. You and the hitman
make these choices at the same time and independently.
1
Perhaps it would be better to take the bridge then...but the hitman might
anticipate this, and send his robot along the bridge...
We will instead consider the case where this scenario take place a large
number of times in a simulation. You can imagine it as a game, where the two
players (you and the hitman), each turn, choose bridge or road. If you
survive you get a point; if you die the hitman gets a point. This happens a large
number of times, and each players goal is to obtain as many points as possible.
Now, you could of course simply play the game based on your intuition,
trying to predict what your opponent will do and trick him and whatnot, but
instead we would like to find an optimal rule that you can use to make your
decisions. Of course, this rule will have an element of randomness to it; if it
was unchanging or followed a noticeable pattern then the hitman could take
advantage of that. Also, your strategy on a given turn should take into account
what the hitman has been observed to do on previous turns (if the hitman
chooses the road every single time, for example, you would be wise to start
choosing the bridge). One last thing: the hitman is no dummy either, and uses
strategy just like you to build his own optimal decision-making rule.
A few miscellaneous things to note before we proceed:
1. The value of r is common knowledge. You know r, the hitman knows r,
you know the hitman knows r, the hitman knows you know he knows r, etc.
2. It doesnt matter to the hitman whether you die by killer robot or by
falling off the bridge. He just needs you dead.
Now lets talk strategy.
2
Now we ask a question that might seem pointless but actually is the key
to creating an optimal decision rule. If you somehow knew the value of k for
a turn, what would you select as y? Well, the values of k and r are known
constants so ps is just a function of y. We differentiate:
dps
= k + r kr
dy
Since k and r are known, dp dy is a constant. If its a positive constant,
s
then increasing y increases your probability of survival ps , and you should set
y = 1 (take the bridge). But if dpdy is negative, then the way to increase your
s
1
y>
1+r
This can be appreciated intuitively too. For example, if the bridge is very
unsafe, the killer might as well send his robot along the road. Substituting low
values of r into this formula, we see that the robot will be sent along the road
unless the hitman knows you are almost certainly taking the bridge. On the
other hand, if the bridge is safe (r = 1) then hell choose the bridge if and only
if your probability of taking the bridge is above 12 .
3
3 Applying the Decision Rule
Even though the value of k is not known to you, there are two ways that you
can practically apply this theory to the canyon game. The first is to realize that
an equilibrium is reached when at least one of the players chooses the border
value for his opponents decision rule. For example, suppose you always select
y = 1+r1
. Then dp dk = 0, that is, the hitmans choice of k has no effect on
s
r
your survival rate. Similarly, if the hitman selects k = 1+r , then dp
dy = 0. (In
s
the long term, both you and the hitman might as well choose these values as
your probabilities, because using any other value allows the opponent to take
advantage using a decision rule.) Recalling that
ps = k ky + yr kyr
1 r
we can substitute either y = 1+r , or k = 1+r , and derive your constant
resultant survival rate
r
ps =
1+r
Doing the substitutions for both y and k gives this too of course, but only
one is necessary.
A second way one might practically apply the decision rule is by proportion.
You make an arbitrary move on your first turn, and after that you keep track
of the fraction of times that your opponent took the bridge, use this as their
selected probability, and apply the inequality. It turns out that if even one of
r
the players does this then ps converges to 1+r , as before. Furthermore, if both
players do this, then the fractions of times they each will be choosing the bridge
will too converge to the values in the equilibrium described. (These facts were
verified by computer simulation.)