GAMES AGAINST NATURE Topic 3 Games Against Nature

  • Slides: 22
Download presentation
GAMES AGAINST NATURE Topic #3

GAMES AGAINST NATURE Topic #3

Games Against Nature • In game theory, for reasons that will be explained later,

Games Against Nature • In game theory, for reasons that will be explained later, the alternatives (e. g. , LEFT and RIGHT) that each player chooses from are usually called strategies, – but they may also be called choices, actions, etc. • A Game Against Nature is – a 1 -player game, in which a single rational self-interested player must choose a strategy, and – the outcome and the player’s payoff depends on both his chosen strategy and the “choice” made by a totally disinterested nature. • A Game Against Nature part of what is generally called decision theory (rather than game theory) because there is only one player who – makes a rational choice, and – is interested in (gets a payoff, i. e. , a gain or loss, from) the outcome.

A Decision Problem without Risk or Uncertainty Choices for P 1 => s 1

A Decision Problem without Risk or Uncertainty Choices for P 1 => s 1 s 2 s 3 s 4 s 5 • Outcomes x y z v w => Preferences/Payoffs/Utility u 1(x) = 4 u 2(y) = 5 u 3(z) = 3 u 4(v) = 4 u 5(w) = 3 In this case, Nature makes no choice, so – – the decision problem is conceptually trivial, i. e. , P 1 should choose the strategy that gives him the best outcome, – in this case s 2. However, such decision problems are often difficult in practice, because the decision maker may not know how to predict or evaluate the payoffs from the different outcomes, – but this problem is essentially one of information and research, not strategic choice. • Decision/game theory focus on the problem of strategic choice, and assumes that all players know their payoffs. •

A Literal Game Against Nature • A payoff matrix is set up as a

A Literal Game Against Nature • A payoff matrix is set up as a crosstabulation of P 1’s strategies and Nature’s contingencies (or another player’s strategies). – This is the simplest possible non-trivial matrix , with two rows and two columns. • Numbers in each cell of a payoff matrix indicate the payoff (or at least rank the payoffs) to each player given by the resulting outcome. – In this case, there is only one number in each cell because there is only one player, and – nature is indifferent among outcome and has no payoffs.

The Maximax Principle • We now identify several different decision principles (some of which

The Maximax Principle • We now identify several different decision principles (some of which may be poor principles) that player P 1 might use in order to pick his strategy. • The Maximax Principle (informally, “aim for the best”): – choose the strategy that can give you the best outcome. – In this case, maximax says “choose s 2. ” • Note however that s 2 can also give your worst outcome, – so maximax can mean “aim for the best --- but get the worst. ”

The Maximin Principle • Maximin Principle (informally, “avoid the worst”): – For each strategy,

The Maximin Principle • Maximin Principle (informally, “avoid the worst”): – For each strategy, determine the worst outcome (minimum payoff) you could possibly get if you select that strategy; • this is called the security level for each strategy. – In this case , • the security level of s 1 is 2, while • the security level of s 2 is 1. – Then choose the strategy that gives you the highest security level, • i. e. , “the maximum of the minimums” or “maximin. ” – In this case, s 1 is P 1’s maximin strategy, and by using it P 1 can guarantee a payoff of at least 2, regardless of what Nature does.

Maximax vs. Maximin • For reasons that will be made evident later, the term

Maximax vs. Maximin • For reasons that will be made evident later, the term ‘minimax’ is used more commonly than ‘maximin. ’

The Principle of Insufficient Reason • The Principle of Insufficient Reason (informally, “maximize the

The Principle of Insufficient Reason • The Principle of Insufficient Reason (informally, “maximize the average payoff”): – choose the strategy the gives you the highest average payoff (averaged across contingencies). – In this case, s 1 and s 2 have the same average value (2. 5). • This principle has two big problems: – It assumes that the payoffs are cardinal [or interval], whereas • the payoffs displayed in the matrix above are probably just ordinal. – It in effect assumes that all contingencies are equally probable.

Decision Making (cont. ) • In general, Nature’s choice of contingencies may be made

Decision Making (cont. ) • In general, Nature’s choice of contingencies may be made – by an objective and known (to the player) random mechanism (e. g. , at a Las Vegas gaming table); • This is called decision making under (objective) risk; – or by some process concerning which the player has subjective probability estimates (e. g. , whether it will rain or shine); • This is called decision making under (subjective) uncertainty. • In either event, it is unlikely that all of nature’s contingencies are equally probable.

Principle of Expected Utility Maximization • Principle of Expected Utility (or Payoff) Maximization: –

Principle of Expected Utility Maximization • Principle of Expected Utility (or Payoff) Maximization: – The payoff matrix has been modified to show cardinal (or interval), rather than merely ordinal, payoffs. • Choose the strategy that gives you the highest expected payoffs, weighting each contingency by its (objective or subjective) probability. – This requires that • payoffs are cardinal (or interval) in nature, and • the decision maker knows the probability over nature’s contingencies. – Given p =. 5, the expected payoff from s 1 = 6 and from s 2 is 5, so P 1 should choose S 1. – If p =. 25 and p =. 75, the expected payoff from s 1 is 7 and from s 2 is 7. 5, so P 1 should choose s 2. 1 1 2 2

Expected Utility Maximization (cont. ) • If rain is certain, P 1 definitely should

Expected Utility Maximization (cont. ) • If rain is certain, P 1 definitely should take his umbrella • If shine is certain, P 1 definitely should not take his umbrella • How probable does rain have to be so that P 1 (if he is maximizing expected payoffs) should choose to take his umbrella? • Put otherwise, what is the probability rain that leaves P 1 indifferent between his two strategies (because they have the same expected payoffs)? • Let p (0 < p < 1)be the probability of rain, so the probability of shine is (1 -p). • • Expected Payoff from S 1 = Expected Payoff from S 2 4 p + 8(1 -p) = 0 p + 10(1 -p) 6 p = 2 p = 1/3 (and p = 2/3) In which case the expected payoff from s 1 = 4/3 + 16/3 = 20/3 = 7. 333 , The expected payoff from s 2 = 0/3 = 20/3 = 7. 333

Eisenhower Before D-Day • D-Day weather decision made by Supreme Allied Commander. – D-Day

Eisenhower Before D-Day • D-Day weather decision made by Supreme Allied Commander. – D-Day June 5 was postponed due to a storm in the channel but another storm was predicted to arrive shortly after the first. – RAF Meteorologist James Martin Stagg predicted there would be a brief window on June 6 with a break in weather. • Eisenhower: “OK – let’s go. ” – Marginal weather predicted for June 6 may have been advantageous in terms the (Zero-Sum) Game Against the Germans (as opposed to the Game Against Nature). – To invade only with favorable tides and full moon entailed disadvantages in terms of the Game Again the Germans but evidently more than compensating advantages in terms of the Game Against Nature. • The Game Against the Germans was, in this respect, “strictly determined” (as defined in Topic #5).

Eisenhower Before D-Day (cont. )

Eisenhower Before D-Day (cont. )

Eisenhower Before D-Day (cont. )

Eisenhower Before D-Day (cont. )

Cardinal Payoffs ≈ $$ • • Mostly, you can think of cardinal payoffs as

Cardinal Payoffs ≈ $$ • • Mostly, you can think of cardinal payoffs as being essentially equivalent to money, but not if a player is not risk-neutral, i. e. , utility = expected $$$ Consider a game with two players plus Nature: – P 1 is you, deciding whether to buy home-owner’s (fire) insurance; – P 2 is an insurance company, deciding whether to sell you a policy; – Nature decides whether your home will burn down in the next year. • If you and the insurance company have the same beliefs about the probability that your house will burn down in the year, is there a premium price such that you will buy and they will sell – No, if your payoffs = expected $$$. – You probably are not risk-neutral but rather risk-averse. – But the insurance company, insuring millions of widely dispersed house, can afford to be risk neutral. • If you win the state lottery, are you indifferent between these two prizes? – Prize 1: $1, 000 – Prize 2: Another lottery ticket, which gives you a. 5 chance of winning $2, 000 and a. 5 chance of winning $0. • But you were risk-acceptant to buy the lottery ticket in the first place,

Decision Problems – Some Practice

Decision Problems – Some Practice

Decision Problems – More Practice • But can you see another and very persuasive

Decision Problems – More Practice • But can you see another and very persuasive decision principle lurking in this example?

The Dominance Principle • The Dominance Principle (D & N, Rule 2, p. 66):

The Dominance Principle • The Dominance Principle (D & N, Rule 2, p. 66): – If strategy A gives you at least as good a payoff in every contingency as strategy B and a better outcome in at least one contingency, then • do not choose strategy B (as it is dominated by strategy A); • choose an undominated strategy instead. – In the payoff matrix above, s 2 dominates both s 1 and s 3, so P 1 should not choose s 1 or s 3. – Put otherwise, S 2 is at least as good a “reply” as s 1 (or s 3) in every contingency and is a better “reply” in at least one contingency.

The Dominance Principle (cont. ) • Important note: the Dominance Principle says: – if

The Dominance Principle (cont. ) • Important note: the Dominance Principle says: – if strategy A dominates strategy B, don’t use strategy B. • The Dominance Principle does not say: – if strategy A dominates strategy B, use strategy A, because • there may be some strategy C that dominates A, or (more likely) • there may be some strategy D that is undominated and which gives a larger payoff than A in some contingencies. • In the matrix above, s 2 and s 4 are both undominated, – that is, each is a best reply in some contingency.

The Dominance Principle (cont. ) • In the modified payoff matrix above, s 2

The Dominance Principle (cont. ) • In the modified payoff matrix above, s 2 now dominates s 4 (as well as s 1 and s 3), – so s 2 is a dominant strategy. • A dominant strategy if a best reply in every contingency. • Corollary of Dominance Principle: If you have a dominant strategy, use it. • Notice that the Dominance Principle does not – require cardinal payoffs in order for it to be applied, and – does not depend on the probabilities of the contingencies.

The Dominance Principle (cont. ) • The Dominance Principle is consistent with every decision

The Dominance Principle (cont. ) • The Dominance Principle is consistent with every decision principle discussed earlier. • If s 1 dominates s 2, – – s 2 cannot be uniquely maximax; S 2 cannot be uniquely minimax S 2 cannot have a higher average payoff S 2 cannot have a higher expected payoff. • But typically players may have many (or only) undominated strategies and – only rarely do players have dominant strategies. • Therefore, while the Dominance Principle is very good advice where it applies, it very often does not apply.

The Dominance Principle (cont. ) • The Dominance Principle appears to be a good

The Dominance Principle (cont. ) • The Dominance Principle appears to be a good principle to follow, not only in Games Against Nature, but also in Games Against Other Players. – This is not true of the other Decision Principles. – But notice that RIGHT dominates LEFT in the Social Dilemma Game (and compare with the Prisoner’s Dilemma Game in Topic #4). • Two examples of dominant strategies (one trivial, one tragic): – Following another car into an almost full parking garage. – Passengers on United #93 deciding to charge the hijackers. • A threat to use a dominant strategy is highly credible (as will be discussed in Section II of the course).