Link Search Menu Expand Document

Intro to Game Theory - Concise Notes

MATH70141

PDFs
Problem Sheets
Table of contents

Prelude

Strategies

Definition 1. A move refers to the action a player must make on their turn to progress from one game position to the next position

Definition 2. An outcome of a game refers to the final result of a game once the game has been played

Definition 3. A strategy for a player involves a complete description of all the moves that will be made in any game position, including responses to any random moves, and the opponent’s moves. A strategy is a program which can be followed to play the game mechanically.

Definition 4. A pure strategy is a strategy that doesn’t involve any self-imposed random chances of playing any moves.

Definition 5. Finite game - if all players in the game have a finite number of pure strategies. If at least one player has an infinite number of pure strategies, the game is called an infinite game.

Dominance, Best Response and Equilibria

Define the following notation to start with

note Note 1. Player $A$ will have pure strategies $A_{s} = {a_1, a_2, \ldots }$, the set may be finite or infinite. Similarly, player $B$ will have pure strategies $B_{s} = {b_1, b_2, \ldots }$

Denote by $g_{A}(a_{i},b_{j})$ the payoff to player $A$ when player $A$ plays pure strategy $a_{i}$ and player $B$ plays pure strategy $b_{j}$.

Definition 6. Strategy $a \in A_{s}$ is strictly dominated by another strategy $a^\prime \in A_{s}$ if \(g_{A}(a,b) < g_{A}(a^\prime,b) \quad \forall b \in B_{s}\)

Definition 7. In an $n$-player game, a strategy $s_{i} \in S_{i}$ for player $i$ is strictly dominated by another strategy $s_{i}^\prime \in S_{i}$ if \(g_{i}(s_{i},s_{-i}) < g_{i}(s_{i}^\prime,s_{-i}) \quad \forall s_{-i} \in S_{-i}\) $s_{-i}$ denotes the strategies of all players other than $i$

Definition 8. $a \in A_{s}$ is weakly dominated by $a^\prime \in A_{s}$ if \(g_{A}(a,b) \leq g_{A}(a^\prime,b) \quad \forall b \in B_{s}\) and there exists at least one $b \in B_{s}$ such that the inequality is strict

Definition 9. In an $n$-player game, a strategy $s_{i} \in S_{i}$ for player $i$ is weakly dominated by another strategy $s_{i}^\prime \in S_{i}$ if \(g_{i}(s_{i},s_{-i}) \leq g_{i}(s_{i}^\prime,s_{-i}) \quad \forall s_{-i} \in S_{-i}\) and there exists at least one $s_{-i} \in S_{-i}$ such that the inequality is strict

Definition 10. In an $n$-player game, a strategy $s_{i} \in S_{i}$ for player $i$ is payoff equivalent to another strategy $s_{i}^\prime \in S_{i}$ if \(g_{i}(s_{i},s_{-i}) = g_{i}(s_{i}^\prime,s_{-i}) \quad \forall s_{-i} \in S_{-i}\)

Definition 11. In an $n$-player game, a strategy $s_{i} \in S_{i}$ for player $i$ is a best response to a strategy profile $s_{-i} \in S_{-i}$ if \(g_{i}(s_{i},s_{-i}) \geq g_{i}(s_{i}^\prime,s_{-i}) \quad \forall s_{i}^\prime \in S_{i}\)

Proposition 12. A dominated strategy is never a best response

Equilibria

Definition 13 (Nash Equilibrium). An equilibrium of an $n$-player game is a strategy profile $s \in S$ such that \(g_{i}(s_{i},s_{-i}) \geq g_{i}(s_{i}^\prime,s_{-i}) \quad \forall s_{i}^\prime \in S_{i}\) for all players $i$.

Iterative Deletion of Dominated Strategies

Proposition 14. In an $N$-player game, with strategy sets $S_1, S_2, \ldots , S{N}$, let $s_{i}, s_{i}^\prime$ be two strategies for player $i$. Suppose $s_{i}^\prime$ weakly dominates or is payoff equivalent to $s_{i}$. Consider game $G^\prime$ with identical payoffs as $G$ but where $S_{i}$ is replaced by $S_{i} - {s_{i}}$, Then:_

  1. Any Nash equilibrium of $G^\prime$ is a Nash equilibrium of $G$

  2. If $s{i}$ is dominated by $s_{i}^\prime$, then $G$ and $G^\prime$ have the same equilibria_

Proposition 15. Consider game $G$ that upon performing iterative deletion of dominated strategies, results in game $G^\prime$ with a single strategy profile. Then, the single strategy profile is the unique equilibrium of $G$.

Mixed Equilibria

Mixed Strategies

Definition 16. A mixed strategy for a player is a self-imposed randomization over the player’s pure strategies. A mixed strategy is a probability distribution over the pure strategies. A mixed strategy $\alpha$ for player $A$ is denoted as \(\begin{aligned} \alpha &= (p_1, p_2, \ldots , p_n), \quad \text{or} \\ \alpha &= p_1 a_1 + p_2 a_2 + \ldots + p_n a_n, \quad \text{where} \quad \sum_{i=1}^{n} p_i = 1, \quad 0 \leq p_i \leq 1 \end{aligned}\) We extend the pure strategy set $A_{s}$ to the more general mixed strategy set, $\mathbb{A}_{s}$ - the infinite set of all possible $\alpha$ for player $A$.

Definition 17. Let player $A$ have pure strategy set $A_{s} = {a_1, \ldots , a_{n} }$ and player $B$ have pure strategy set $B_{s} = {b_1, \ldots , b_{m} }$.

If player $A$ choses to play the mixed strategy $\alpha = (p_1, \ldots , p_{n} ) \in \mathbb{A}{s}$ and player $B$ choses to play the mixed strategy $\beta = (q_1, \ldots , q{m} ) \in \mathbb{B}{s}$, then the expected payoff to player $A$ is \(g_{A}(\alpha,\beta) = \sum_{i=1}^{n} \sum_{j=1}^{m} p_i q_j g_{A}(a_i,b_j)\) If $A{s}, B_{s}$ are infinite sets then the summation is replaced by integration. \(g_{A}(\alpha ,\beta ) = \int_{x} \int_{y} g_{A}(x,y) f_{A}(x) f_{B}(y)\, dx \, dy\) where $f_{A}(x), f_{B}(y)$ are the probability density functions of the mixed strategies $\alpha, \beta$ respectively.

Definition 18. A pair of mixed strategies $\alpha^{\ast}$ for $A$ and $\beta^{\ast}$ for $B$, are said to be in mixed equilibrium if

\[\begin{aligned} g_{A}(\alpha^{\ast},\beta^{\ast}) &\geq g_{A}(\alpha,\beta^{\ast}) \quad \forall \alpha \in \mathbb{A}_{s} \\ \text{and} \quad g_{B}(\alpha^{\ast},\beta^{\ast}) &\geq g_{B}(\alpha^{\ast},\beta) \quad \forall \beta \in \mathbb{B}_{s} \end{aligned}\]

Finding mixed equilibria by considering Pure strategies

Proposition 19. _For any mixed strategies $\alpha^{\ast}$ of player $A$ and $\beta^{\ast}$ of player $B$, then \(\begin{aligned} \mathop{\max}_{\alpha \in \mathbb{A}_{s}}\{ g_{A}(\alpha,\beta^{\ast})\} &= \mathop{\max}_{a \in A_{s}}\{ g*{A}(a,\beta^{\ast})\},\\ \mathop{\max}*{\beta \in \mathbb{B}_{s}}\{ g_{B}(\alpha^{\ast},\beta)\} &= \mathop{\max}_{b \in B_{s}}\{ g*{B}(\alpha^{\ast},b)\} \end{aligned}\)*

Definition 20. Let $c$ a constant. A mixed strategy $\alpha^{\ast}$, for player $A$ is an equaliser strategy if \(g_{A}(\alpha^{\ast},b) = c \quad \forall b \in \mathbb{B}_{s}\)

Similarly for player $B$

Proposition 21. In a 2-player game, if $\alpha^{\ast}$ is an equaliser strategy for $A$ using $B$’s payoffs and $\beta^{\ast}$ is an equaliser strategy for $B$ using $A$’s payoffs, then $(\alpha^{\ast},\beta^{\ast})$ is a mixed equilibrium

Geometry of Games

note Note 2. Define the convex hull of a set of points as the smallest convex set that contains all the points. For a set of points ${x_1, \ldots, x_{n} }$ with each $x_{i} \in \mathbb{R}^m$, form their convex hull as \(C = \left\{ \sum_{i=1}^{n} \lambda_{i} x_{i} \mid \lambda_{i} \geq 0, \sum_{i=1}^{n} \lambda_{i} = 1 \right\}\)

Existence of an equilibrium

Theorem 22 (Nash, 1951). Every finite game has at least one mixed equilibrium

Finding equilibria by checking subgames

The upper envelope method

Degenerate games

Definition 23 (Degenerate game). A 2-player game is said to be degenerate if some player has a mixed strategy that assigns positive probability to exactly $k$ pure strategies so that the other player has more than $k$ pure strategies.

Zero-sum games

Max-min and Min-max Strategies

Definition 24. A max-min strategy $\hat{\alpha} \in \mathbb{A}_{s}$ of player $A$ is a strategy such that \(\mathop{\min}_{\beta \in \mathbb{B}_{s}}\{ g_{A}(\hat{\alpha},\beta)\} = \mathop{\max}_{\alpha \in \mathbb{A}_{s}} \left\{ \mathop{\min}_{\beta \in \mathbb{B}_{s}}\{ g_{A}(\alpha,\beta)\} \right\}\) assuming that the maxima and minima exist. This also defines the max-min payoff to player $A$

Definition 25. A min-max strategy $\hat{\beta} \in \mathbb{B}_{s}$ of player $B$ is a strategy such that \(\mathop{\max}_{\alpha \in \mathbb{A}_{s}}\{ g_{B}(\alpha,\hat{\beta})\} = \mathop{\min}_{\beta \in \mathbb{B}_{s}} \left\{ \mathop{\max}_{\alpha \in \mathbb{A}_{s}}\{ g_{B}(\alpha,\beta)\} \right\}\) This also defines the min-max payoff to player $B$

Proposition 26. In a zero-sum game, for $\alpha \in \mathbb{A}{s}$, then \(\mathop{\min}_{\beta \in \mathbb{B}_{s}}\{ g_{A}(\alpha,\beta)\} = \mathop{\min}_{b \in B_{s}}\{ g_{A}(\alpha,b)\}\) Similarly for $\beta \in \mathbb{B}_{s}$, then \(\mathop{\max}_{\alpha \in \mathbb{A}_{s}}\{ g_{B}(\alpha,\beta)\} = \mathop{\max}_{a \in A_{s}}\{ g_{B}(a,\beta)\}\)_

Relationship of Equilibria and Max-min/Min-max Strategies

Proposition 27. *In a finite zero-sum game with $\hat{\alpha} \in \mathbb{A}{s}, \hat{\beta} \in \mathbb{B}{s}$ then $(\hat{\alpha},\hat{\beta})$ is a mixed equilibrium if and only if $\hat{\alpha}$ is a max-min strategy for $A$ and $\hat{\beta}$ is a min-max strategy for $B$, and

\(\mathop{\max}_{\alpha \in \mathbb{A}_{s}} \left\{ \mathop{\min}_{\beta \in \mathbb{B}_{s}}\{ g_{A}(\alpha,\beta)\} \right\} = \mathop{\min}_{\beta \in \mathbb{B}_{s}} \left\{ \mathop{\max}_{\alpha \in \mathbb{A}_{s}}\{ g_{B}(\alpha,\beta)\} \right\}\)*

The Minimax theorem of Von Neumann

Theorem 28 (Von Neumann, 1928). In a finite zero-sum game then \(\mathop{\max}_{\alpha \in \mathbb{A}_{s}} \left\{ \mathop{\min}_{\beta \in \mathbb{B}_{s}}\{ g_{A}(\alpha,\beta)\} \right\} = v = \mathop{\min}_{\beta \in \mathbb{B}_{s}} \left\{ \mathop{\max}_{\alpha \in \mathbb{A}_{s}}\{ g_{B}(\alpha,\beta)\} \right\}\) where $v$ is the unique max-min payoff to $A$ (and cost to $B$), called the value of the game.

Finding solutions in small zero-sum games

Proposition 29. Consider 2 zero-sum games $G, G^\prime$, where $G^\prime$ is obtained from $G$ by deleting a weakly dominated strategy of one of the players. Then any equilibrium of $G^\prime$ is also an equilibrium of $G$, and $G$ and $G^\prime$ have the same value.

Cooperative Games

Bargaining sets

Definition 30. Bargaining (Negotiation) set, $S$, resulting from a 2-player game in strategic form is the convex hull of all payoff pairs, with the added constraint that \(\forall (x,y) \in S, \quad x \geq t_{A}, \quad y \geq t_{B}\) where $t_{A}, t_{B}$ are the max-min payoff of player $A$ and $B$ respectively. Known as $A$ and $B$’s security level or threat level.

Call $(t_{A}, t_{B})$ the threat point

Bargaining Axioms

Definition 31 (Axioms for bargaining solution). For a bargaining set $S$ with threat point $(t_{A}, t_{B})$, a Nash bargaining solution $N(S) = (X,Y)$ is said to satisfy the following axioms:

(a) Efficiency - $(X,Y) \in S$

(b) Pareto optimality - $(X,Y)$ are Pareto optimal, i.e. $\forall (x,y) \in S$ if $x \geq X$ and $y \geq Y$, then $(x,y) = (X,Y)$

(c) Invariant under payoff scaling, meaning if $a,c > 0$ and $b,d \in \mathbb{R}$ and we define $S^\prime$ to be the bargaining set \(S^\prime = \{ (ax + b, cy + d) \mid (x,y) \in S\}\) with threat point $(at_{A} + b, ct_{B} + d)$, then $N(S^\prime) = (aX + b, cY + d)$

(d) Symmetry - If $t_A = t_B$ and $(x,y) \in S$ implies $(y,x) \in S$ then we must have $X = Y$

(e) Independence of irrelevant alternatives - If $S,T$ are bargaining sets with the same threat point and $S \subset T$, then either $N(S) = N(T)$ or $N(T) \notin S$

The Nash Bargaining Solution

Theorem 32. Under the axioms of bargaining solution, (a)-(e) above. Every bargaining set $S$ that contains a point $(x,y)$ with $x > t_A, y > t_B$, has a unique Nash bargaining solution $N(S) = (X,Y)$

Obtained as the unique point $(x,y) \in S$ that maximises the Nash product \((x - t_{A})(y - t_{B})\)

Congestion Games

Components of a Congestion Game

Definition 33. A congestion network has the following components:

  1. A finite set of nodes

  2. A finite set of directed edges, each edge, e, an ordered pair written $AB$ from node $A$ to node $B$

  3. Each edge $e$ has an associated cost function $c_{e}(x)$ giving value when there are $x$ users on edge $e$, with $c_{e}(x)$ weakly increasing in $x$ \(x \leq y \implies c_{e}(x) \leq c_{e}(y)\)

Definition 34. To form a congestion game, we need the following components:

  1. A congestion network

  2. $N$ users of network with each user having a origin node, $O_{i}$ and a destination node $D_{i}$

  3. A strategy of user $i$ is a path $P_{i}$ from $O_{i} \to D_{i}$. Given strategy $P_{i}$ for each user $i$, the flow on edge $e$ is the number of users using edge $e$ \(f_{e} = \left\lVert \{i : e \in P_{i}\}\right\rVert\)

  4. The cost to user $i$ of using path $P_{i}$ is the sum of the costs of the edges in $P_{i}$ \(\text{Cost}_{i}(P_{i}) = \sum_{e \in P_{i}} c_{e}(f_{e})\)

Definition 35. Say $P_{i}$ a best response for user $i$ if against strategies $P_{j}$, $j \neq i$, then \(\sum_{e \in P_{i}} c_{e}(f_{e}) \leq \sum_{e \in P_{i} \cap Q_{i}} c_{e}(f_{e}) + \sum_{e \in P_{i} / Q_{i}} c_{e}(f_{e} + 1)\) holds for every possible alternative path $Q_{i}$ for user $i$

Definition 36. In a congestion game with $N$ users strategies $P_1, P_2, \ldots , P_{N}$ of all $N$ users define an equilibrium if each strategy is a best response to the other strategies. i.e if the above inequality holds for all $i$

Existence of Equilibrium in Congestion Games

Theorem 37. Every congestion game has at least one equilibrium

Price of Anarchy

Definition 38. The price of anarchy of a congestion game is the ratio of the cost of the worst equilibrium to the cost of the best possible solution \(\text{PoA} = \frac{\text{Worst average cost per user in any equilibrium}}{\text{Average cost per user in social optimum} } = \frac{\max_{P} \sum_{i} \text{Cost}_{i}(P_{i})}{\min_{P} \sum_{i} \text{Cost}_{i}(P_{i})}\)

Proposition 39. For atomic flow congestion games, the price of anarchy is at most $5/2$

Proposition 40. For split-able flow congestion games, the price of anarchy is at most $4/3$

Combinatorial Games

These are 2-player, perfect information games with no chance moves. They come in 2 types:

  • Partizan games - where the players have different sets of moves

  • Impartial games - where the players have the same set of moves

The Ending Condition

A combinatorial game ends when there are no legal moves left for any player. The game is then said to be in a terminal position. This is a necessary condition for a game to be a combinatorial game.

The Normal Play Convention

The normal play convention is that the player who cannot move loses the game. This is a necessary condition for a game to be a combinatorial game.

Nim and Impartial Games

Definition 41. An option of a game position in a combinatorial game is a position that can be reached in one move from the player to move.

Winning and Losing Positions

Impartial games, game positions belong to one of 2 classes:

  • Winning positions - the player to move has a winning move

  • Losing positions - the player to move has no winning move

Proposition 42. In an impartial game, a game position is losing if and only if all its options are winning positions. A game is winning if and only if at least one of its options is a losing position; moving to that position is a winning move.

Proposition 43. A Nim position is losing if and only if the Nim sum equals zero for all columns in the binary representation of the position; such a position is called a zero position. A Nim position is winning if and only if the Nim sum is not zero.

Top-down induction

Partial and Total Orders

Definition 44. A binary relation $\simeq$ on a set $S$ is a partial order if, for all $x,y,z \in S$, we have:

  • Reflexivity - $x \simeq x$

  • Antisymmetry - $x \simeq y$ and $y \simeq x$ implies $x = y$

  • Transitivity - $x \simeq y$ and $y \simeq z$ implies $x \simeq z$

If in addition to the above, for all $x,y \in S$, we have:

  • Comparability - $x \simeq y$ or $y \simeq x$

then $\simeq$ is a total order

Definition 45. For a given partial order $\simeq$ on a set $S$, we define the strict order $\sim$ corresponding to $\simeq$ by; for all $x,y \in S$: \(x \sim y \iff x \simeq y \text{ and } x \neq y\)

Definition 46. An element $x \in S$ is maximal if there is no $y \in S$ such that $x \sim y$

Back to Top-Down Induction

Definition 47. Consider a set $S$ of games, defined by a starting game and all the games that can be reached from it via any sequence of moves of the players. For two games; $G,H \in S$, we call $H$ simpler than $G$, denoted with the binary relation $H \leq G$, if there is a sequence of moves that leads from $G$ to $H$. We allow for $G = H$ where this sequence is empty.

Proposition 48. The binary relation $\leq$ (‘simpler than’) on a set $S$ of games is a partial order

Proposition 49. Every non-empty subset, $T$, of $S$ has a minimal element

Theorem 50 (Top-down induction). Consider a set $S$ with a partial order $\simeq$ such that every non-empty subset of $S$ has a minimal element. Let $P(x)$ be a statement about an element $x \in S$ that may be true or false. Assume that $P(x)$ holds whenever $P(y)$ holds for all $y \in S$ such that $y \sim x$. Then $P(x)$ is true for all $x \in S$. That is \((\forall x :\ ( \forall y \sim x : P(y)) \implies P(x)) \implies (\forall x : P(x))\)

Game Sums

Definition 51. Suppose that $G$ and $H$ are games with options $G_1, \ldots , G_n$ and $H_1, \ldots , H_m$ respectively. Then the game sum $G + H$ is the game with options $G_1 + H, \ldots , G_n + H, G + H_1, \ldots , G + H_m$

Proposition 52. Denoting the losing game with no options by $0$, then for any games $G,H$ and $J$ we have

  • Commutativity of +: \(G + H = H + G\)

  • Associativity of +: \((G + H) + J = G + (H + J)\)

  • Identity of +: \(G + 0 = G\)

Equivalence of Games

Definition 53. Two games $G$ and $H$ are called equivalent, written $G \equiv H$, if and only if for any other game $J$, the game sum $G+J$ is losing if and only if $H + J$ is losing

Lemma 54. The binary relation of equivalence, $\equiv$, is an equivalence relation between games, this means that it is:

  • Reflexive - $G \equiv G$

  • Symmetric - $G \equiv H$ implies $H \equiv G$

  • Transitive - $G \equiv H$ and $H \equiv J$ implies $G \equiv J$

Proposition 55. Two Nim piles are equivalent if and only if they have the same size

Proposition 56. $G$ is a losing game if and only if $G \equiv 0$

Corollary 57. Any two losing games are equivalent

Lemma 58. For all games $G,H$ and $K$ we have: \(G \equiv H \implies G + K \equiv H + K\)

Lemma 59. Let $J$ be a losing game. Then $G + J \equiv G$ for any game $G$

Proposition 60 (The Copycat Principle). $G + G \equiv 0$ for any impartial game $G$

Lemma 61. For impartial games $G$ and $H$, then $G \equiv H$ if and only if $G + H \equiv 0$

Notation for Nim Piles

Definition 62. If $G$ is a single Nim pile with $n \geq 0$ tokens in it, then we denote this game by $*n$. This game is specified by its $n$ options, defined recursively as \(*0, *1, \ldots , *(n-1)\)

Definition 63. If $G \equiv *m$ for an impartial game $G$, then $m$ is called the Nim value of $G$

The Mex Rule

Definition 64. For a finite set of natural numbers $S$, the minimum excluded number of $S$, written $mex(S)$, is defined as \(mex(S) = \min \{n \in \mathbb{N}\mid n \notin S\}\) In other words, $mex(S)$ is the smallest non-negative integer not contained in $S$ e.g. $mex({0,1,3,4,6}) = 2$

Theorem 65 (The Mex Rule). Any impartial game $G$ has Nim value $m$, where $m$ is uniquely determined as follows; for each option $H$ of $G$, let $H$ have Nim value $s_{H}$, and let $S = {s_{H}: H \text{ is an option of } G}$. Then $m = mex(S)$, that is, $G \equiv *(mex(S))$

Sums of Nim Piles

Definition 66. If $*k \equiv *m + *n$, then we call $k$ the Nim sum of $m$ and $n$, and write $k = m \oplus n$

Theorem 67. Let $n \in \mathbb{Z}^+$, and represent $n$ as a unique sum of powers of $2$, i.e. write $n = 2^{a} + 2^{b} + 2^{c} + \ldots$, where $a > b > c > \ldots \geq 0$. Then \(*n \equiv *2^{a} \oplus *2^{b} \oplus *2^{c} \oplus \ldots\)

$$