Date post: | 31-Dec-2015 |
Category: |
Documents |
Upload: | ferdinand-anderson |
View: | 32 times |
Download: | 3 times |
Chapter 12 & Module E
Decision Analysis
&
Game Theory
Components of Decision Making (D.M.)
Decision alternatives - for managers to choose from.
States of nature - that may actually occur in the future regardless of the decision.
Payoffs - payoff of a decision alternative in a state of nature.The components are given in Payoff Tables.
A Payoff Table (It shows payoffs of different decisions at different states of nature)
Investment States of Nature
decision Economy Economy
alternatives good bad
Apartment $ 50,000 $ 30,000
Office 100,000 - 40,000
Warehouse 30,000 10,000
Types of Decision Making (D.M.) - 1
Deterministic D.M. (D.M. under certainty):– Only one “state of nature”,– Payoff of an alternative is known,– Examples:
Problems for LP, IP, transportation, and network flows.
Types of Decision Making (D.M.) - 2
D.M. without probabilities (D.M. under uncertainty):– More than one states of nature;– Payoff of an alternative is not known at the
time of making decision;– Probabilities of states of nature are not
known.
Types of Decision Making (D.M.) - 3
D.M. with probabilities (D.M. under risk)– More than one states of nature;– Payoff of an alternative is not known at the
time of making decision;– Probabilities of states of nature are known
or given.
Types of Decision Making (D.M.) - 4
D.M. in competition (Game theory)– Making decision against a human
competitor.
Decision Making without Probabilities
No information about possibilities of states of nature.
Five criteria (approaches) for a decision maker to choose from, depending on his/her preference.
Criterion 1: Maximax Pick the maximum of the maximums of
payoffs of decision alternatives. (Best of the bests)Investment States of Nature max
decision Economy Economy payoffs
alternatives good bad (bests)
Apartment $ 50,000 $ 30,000 $50,000
Office 100,000 - 40,000 100,000
Warehouse 30,000 10,000 30,000
Decision:
Whom Is MaxiMax for?
MaxiMax method is for optimistic decision makers who tend to grasp every chance of making money, who tend to take risk, who tend to focus on the most fortunate outcome of an alternative and overlook the possible catastrophic outcomes of an alternative.
Criterion 2: Maximin Pick the maximum of the minimums of
payoffs of decision alternatives. (Best of the worsts)Investment States of Nature min
decision Economy Economy payoff
alternatives good bad (worsts)
Apartment $ 50,000 $ 30,000 $30,000
Office 100,000 - 40,000 - 40,000
Warehouse 30,000 10,000 10,000
Decision:
Whom Is MaxiMin for?
MaxiMin method is for pessimistic decision makers who tend to be conservative, who tend to avoid risks, who tend to be more concerned about being hurt by the most unfortunate outcome than the opportunity of being fortunate.
Criterion 3: Minimax Regret
Pick the minimum of the maximums of regrets of decision alternatives. (Best of the worst regrets)
Need to construct a regret table first.
Regret of a decision under a state of nature= (the best payoff under the state of nature)
– (payoff of the decision under the state of nature)
Investment States of Nature decision Economy Economy alternatives good bad
Apartment $ 50,000 $ 30,000Office 100,000 - 40,000Warehouse 30,000 10,000
Payoffs
Investment States of Nature max decision Economy Economy regret alternatives good bad
Apartment $ 50,000 $ 0 $50,000Office 0 70,000 70,000Warehouse 70,000 20,000 70,000
Regrets
Decision:
Whom Is MiniMax Regret for?
MiniMax regret method is for a decision maker who is afraid of being hurt by the feeling of regret and tries to reduce the future regret on his/her current decision to minimum. “I concern more about the regret I’ll have than how much I’ll make or lose.”
Criterion 4: Hurwicz
Pick the maximum of Hurwicz values of decision alternatives. (Best of the weighted averages of the best and the worst)
Hurwicz value of a decision alternative
= (its max payoff)() + (its min payoff)(1-)
where (01) is called coefficient of optimism.
Investment States of Nature decision Economy Economy alternatives good bad
Apartment $ 50,000 $ 30,000Office 100,000 - 40,000Warehouse 30,000 10,000
Payoffs
Investment decision Hurwicz Values alternatives
Apartment 50,000(0.4)+30,000(0.6) = 38,000Office 100,000(0.4)40,000(0.6) = 16,000Warehouse 30,000(0.4)+10,000(0.6) = 18,000
Hurwicz Values with =0.4
Decision:
Whom Is Hurwicz Method for?
Hurwicz method is for an extreme risk taker (=1), an extreme risk averter (=0), and a person between the two extremes ( is somewhere between 1 and 0) .
Criterion 5: Equal Likelihood
Pick the maximum of the average payoffs of decision alternatives. (Best of the plain averages)
Average payoff of a decision alternative
natureofstatesofnumber
ealternativthatofpayoffsallofsum
Investment States of Nature decision Economy Economy alternatives good bad
Apartment $ 50,000 $ 30,000Office 100,000 - 40,000Warehouse 30,000 10,000
Payoffs
Investment decision Average Payoffs alternatives
Apartment (50,000+30,000) / 2 = 40,000Office (100,00040,000) / 2 = 30,000Warehouse (30,000+10,000) / 2 = 20,000
Average Payoffs
Decision:
Whom Is Equally Likelihood for?
Equally likelihood method is for a decision maker who tends to simply use the average payoff to judge an alternative.
Dominated Alternative If alternative A’s payoffs are lower than
alternative B’s payoffs under all states of nature, then A is called a dominated alternative by B.
A dominated alternative can be removed from the payoff table to simplify the problem.
Investment States of Nature decision Economy Economy alternatives good bad
Apartment $ 50,000 $ 30,000Office 100,000 - 40,000Warehouse 30,000 10,000
Decision Making with Probabilities
The probability that each state of nature will actually occur is known.
States of Nature
Investment Economy Economy
decision good bad
alternatives 0.6 0.4
Apartment $ 50,000 $ 30,000
Office 100,000 - 40,000
Warehouse 30,000 10,000
Criterion:Expected Payoff
Select the alternative that has the largest expected payoff.
Expected payoff of an alternative:
n=number of states of nature
Pi=probability of the i-th state of nature
Vi=payoff of the alternative under the i-th state of nature
n
i
ii PV1
*
Example
Decision Alt’s
Econ
Good
0.6
Econ
Bad
0.4 Expected payoff
Apartment 50,000 30,000
Office 100,000 -40,000
Warehouse 30,000 10,000
Expected Opportunity Loss (EOL)
Each decision alternative has an EOL which is the expected value of the opportunity costs (regrets).
The alternative with minimum EOL has the highest expected payoff.
Investment States of Nature decision Economy Economy alternatives good bad
Apartment $ 50,000 $ 30,000Office 100,000 - 40,000Warehouse 30,000 10,000
Payoffs 0.6 0.4
Investment States of Nature decision Economy Economy alternatives good bad
Apartment $ 50,000 $ 0Office 0 70,000Warehouse 70,000 20,000
Opp Loss Table 0.6 0.4
Example (cont.)
EOL (apartment)
= 50,000*0.6 +0*0.4 = 30,000 EOL (office)
=0*0.6+70,000*0.4 = 28,000 EOL (warehouse)
= 70,000*0.6+20,000*0.4 = 50,000
Minimum EOL = 28,000 that is associated with Office.
(Max Exp. Payoff) vs. (Min EOL)
The alternative with minimum EOL has the highest expected payoff.
The alternative selected by (Max expected payoff) and by (Min EOL) are always same.
Expected Value of Perfect Information (EVPI)
It is a measure of the value of additional information on states of nature.
It tells up to how much you would pay for additional information.
An ExampleIf a consulting firm offers to provide “perfect information
about the future with $5,000, would you take the offer?
States of Nature
Investment Economy Economy
decision good bad
alternatives 0.6 0.4
Apartment $ 50,000 $ 30,000
Office 100,000 - 40,000
Warehouse 30,000 10,000
Another Example You can play the game for many times. What is your rational strategy of “guessing”? Someone offers you perfect information about
“landing” at $65 per time. Do you take it? If not, how much you would pay?
Land on ‘Head’ Land on ‘Tail’
Guess ‘Head’ $100 - $60
Guess ‘Tail’ - $80 $150
Calculating EVPI
EVPI
= EVwPI – EVw/oPI
= (Exp. payoff with perfect information) –
(Exp. payoff without perfect information)
Expected payoff with Perfect InformationEVwPI
where n=number of states of nature hi=highest payoff of i-th state of nature
Pi=probability of i-th state of nature
n
i
ii Ph1
Example for Expected payoff with Perfect Information
States of Nature
Investment Economy Economy
decision good bad
alternatives 0.6 0.4
Apartment $ 50,000 $ 30,000
Office 100,000 - 40,000
Warehouse 30,000 10,000
hi 100,000 30,000
Expected payoff with perfect information
= 100,000*0.6+30,000*0.4 = 72,000
Expected payoff without Perfect Information
Expected payoff of the best alternative selected without using additional information. i.e.,
EVw/oPI = Max Exp. Payoff
Example for Expected payoff without Perfect Information
Decision Alt’s
Econ
Good
0.6
Econ
Bad
0.4 Expected payoff
Apartment 50,000 30,000 42,000
*Office 100,000 -40,000 *44,000
Warehouse 30,000 10,000 22,000
Expected Value of Perfect Information (EVPI) in above Example
EVPI
= EVwPI – EVw/oPI
= 72,000 - 44,000
= $28,000
Example Revisit
0.5 0.5
Land on “Head” Land on “Tail”
Guess “Head” $100 -$60
Guess “Tail” -$80 $150
Up to how much would you pay for a piece of information about result of “landing”?
EVPI is equal to (Min EOL)
EVPI is the expected opportunity loss (EOL) for the selected decision alternative.
Maximum average payoff per game
Alt. 1, Guess “Head”
Alt. 2, Guess “Tail”
EMV
EMV
regr
et regr
et
average payoff
average payoff
EO
L EO
LAlternatives
$
125
20
35
EVPI is a Benchmark in Bargain
EVPI is the maximum $ amount the decision maker would pay to purchase perfect information.
Value of Imperfect Information
Expected value of imperfect information
= (discounted EVwPI) – EVw/oPI
= (EVwPI * (% of perfection)) – EVw/oPI
Decision Tree Decision tree is used to help make a series of
decisions. A decision tree is composed of decision nodes
(square), chance nodes (circle), and payoff nodes (final or tip nodes).
A decision tree reflects the decision making process and the possible payoffs with different decisions under different states of nature.
Making Decision on a Decision Tree
It is actually a process of marking numbers on nodes.
Mark numbers from right to left. For a chance (circle) node, mark it with its
expected value. For a decision (square) node, select a
decision and mark the node with the number associated with the decision.
Example p.551
Example p.558-559
Game Theory
Game theory is for decision making under competition.
Two or more decision makers are involved, who have conflicting interests.
Two-Person Zero-Sum Game
Two decision makers’ benefits are completely oppositei.e., one person’s gain is another person’s loss
Payoff/penalty table (zero-sum table):– shows “offensive” strategies (in rows) versus
“defensive” strategies (in columns);– gives the gain of row player (loss of column
player), of each possible strategy encounter.
Example 1 (payoff/penalty table)
Athlete Manager’s Strategies
Strategies (Column Strategies)
(row strat.) A B C
1 $50,000 $35,000 $30,000
2 $60,000 $40,000 $20,000
Two-Person Constant-Sum Game
For any strategy encounter, the row player’s payoff and the column player’s payoff add up to a constant C.
It can be converted to a two-person zero-sum game by subtracting half of the constant (i.e. 0.5C) from each payoff.
Example 2 (2-person, constant-sum)
During the 8-9pm time slot, two broadcasting networks are vying for an audience of 100 million viewers, who would watch either of the two networks.
Payoffs of nw1 for the constant-sum of 100(million)
Network 2
Network 1 western Soap Comedy
western 35 15 60
soap 45 58 50
comedy 38 14 70
An equivalent zero-sum table
Network 2
Network 1 western Soap Comedy
western -15 -35 10
soap - 5 8 0
comedy -12 -36 20
Equilibrium Point
In a two-person zero-sum game, if there is a payoff value P such that
P = max{row minimums} = min{column
maximums}
then P is called the equilibrium point, or saddle point, of the game.
Example 3 (equilibrium point)
Athlete Manager’s Strategies
Strategies (Column Strategies)
(row strat.) A B C
1 $50,000 $35,000 $30,000
2 $60,000 $40,000 $20,000
Game with an Equilibrium Point: Pure Strategy
The equilibrium point is the only rational outcome of this game; and its corresponding strategies for the two sides are their best choices, called pure strategy.
The value at the equilibrium point is called the value of the game.
At the equilibrium point, neither side can benefit from a unilateral change in strategy.
Pure Strategy of Example 3
Athlete Manager’s Strategies
Strategies (Column Strategies)
(row strat.) A B C
1 $50,000 $35,000 $30,000
2 $60,000 $40,000 $20,000
Example 4 (2-person, 0-sum)
Row
Players Column Player Strategies
Strategies 1 2 3
1 4 4 10
2 2 3 1
3 6 5 7
Mixed Strategy
If a game does not have an equilibrium, the best strategy would be a mixed strategy.
Game without an Equilibrium Point
A player may benefit from unilateral change for any pure strategy. Therefore, the game would get into a loop.
To break loop, a mixed strategy is applied.
Example:
Company I Company II Strategies
Strategies B C
2 8 4
3 1 7
Mixed Strategy
A mixed strategy for a player is a set of probabilities each for an alternative of the player.
The expected payoff of row player (or the expected loss of column player) is called the value of the game.
Example:
Company I Company II Strategies
Strategies B C
2 8 4
3 1 7
Let mixed strategy for company I be
{0.6, 0.4}; and for Company II be
{0.3, 0.7}.
Equilibrium Mixed StrategyAn equilibrium mixed strategy
makes expected values of any player’s individual strategies identical.
Every game contains one equilibrium mixed strategy.
The equilibrium mixed strategy is the best strategy.
How to Find Equilibrium Mixed Strategy
By linear programming (as introduced in book)
By QM for Windows, – we use this approach.
Both Are Better Off at Equilibrium
At equilibrium, both players are better off, compared to maximin strategy for row player and minimax strategy for column player.
No player would benefit from unilaterally changing the strategy.
A Care-Free Strategy
The row player’s expected gain remains constant as far as he stays with his mixed strategy (no matter what strategy the column player uses).
The column player’s expected loss remains constant as far as he stays with his mixed strategy (no matter what strategy the row player uses).
Unilateral Change from Equilibrium by Column Playerprobability 0.1 0.9
B C
0.6 Strat 2 8 4
0.4 Strat 3 1 7
Unilateral Change from Equilibrium by Column Playerprobability 1.0 0
B C
0.6 Strat 2 8 4
0.4 Strat 3 1 7
Unilateral Change from Equilibrium by Row Player
probability 0.3 0.7
B C
0.2 Strat 2 8 4
0.8 Strat 3 1 7
A Double-Secure Strategy
At the equilibrium, the expected gain or loss will not change unless both players give up their equilibrium strategies.
– Note: Expected gain of row player is always equal to expected loss of column player, even not at the equilibrium, since 0-sum)
Both Leave Their Equilibrium Strategies
probability 0.8 0.2
B C
0.5 Strat 2 8 4
0.5 Strat 3 1 7
Both Leave Their Equilibrium Strategies
probability 0 1
B C
0.2 Strat 2 8 4
0.8 Strat 3 1 7
Penalty for Leaving Equilibrium
It is equilibrium because it discourages any unilateral change.
If a player unilaterally leaves the equilibrium strategy, then– his expected gain or loss would not change,
and– once the change is identified by the competitor,
the competitor can easily beat the non-equilibrium strategy.
Find the Equilibrium Mixed Strategy
Method 1: As on p.573-574 of our text book. The method is limited to 2X2 payoff tables.
Method 2: Linear programming. A general method.
Method we use: Software QM.
Implementation of a Mixed Strategy Applied in the situations where the mixed
strategy would be used many times. Randomly select a strategy each time
according to the probabilities in the strategy.
If you had good information about the payoff table, you could figure out not only your best strategy, but also the best strategy of your competitor (!).
Dominating Strategy vs. Dominated Strategy
For row strategies A and B: If A has a better (larger) payoff than B for any column strategy, then B is dominated by A.
For column strategies X and Y: if X has a better (smaller) payoff than Y for any row strategy, then Y is dominated by X.
A dominated decision can be removed from the payoff table to simplify the problem.
Example:
Company I Company II Strategies
Strategies A B C
1 9 7 2
2 11 8 4
3 4 1 7
Find the Optimal Mixed Strategy in 2X2 Table
Suppose row player has two strategies, 1 and 2, and column player has two strategies, A and B.
For row player:
Let p be probability of selecting row strategy 1. Then the probability of selecting row strategy 2 is (1-p).
Represent EA and EB by p, where EA (EB) is the expected payoff of the row player if the column player chose column strategy A (B).
Set EA = EB , and solve p from the equation.
For column player:
Let p be probability of selecting column strategy A. Then the probability of selecting column strategy B is (1-p).
Represent E1 and E2 by p, where E1 (E2) is the expected payoff of the row player if the column player chose column strategy A (B).
Set E1 = E2 , and solve p from the equation.