1. Game Theory Just last week:
CMU poker AI player Libratus beats top human poker players in heads up no-limit Texas Hold’em. A monumental achievement! (Compare to Chess (1997), Jeopardy (2009), Go (2016)).

2. Matrix Games
A Matrix Game is given by a payoff matrix 𝐴 =( 𝑎 𝑖𝑗 )∈ 𝑅 𝑚×𝑛 .
Row player (Minnie) chooses row 𝑖∈ 1,…,𝑚 (without seeing Max's move).
Column player (Max) chooses column 𝑗∈{1,…𝑛} (without seeing Minnies's move).
Max gains 𝑎 𝑖𝑗 “dollars” and Minnie loses 𝑎 𝑖𝑗 “dollars”.
Matrix games are zero-sum games as Max gains exactly what Minnie loses.
Negative numbers can be used to model money transferred in the other direction.
Warning: In many texts and old exam problems, the Row player is the maximizer and the Column player is the minimizer.

3. Column Players (Max's) Optimal randomized Strategy
Optimal randomized strategy and guaranteed lower bound on exp. gain for Max is ( 𝑝 1 , 𝑝 2 ,…, 𝑝 𝑛 ;𝑔) which is a solution to the LP: max 𝑔 s.t. 𝑗=1 𝑛 𝑎 𝑖𝑗 𝑝 𝑗 ≥𝑔 𝑖=1,…,𝑚 𝑗=1 𝑛 𝑝 𝑗 =1 𝑝 𝑗 ≥0 𝑗=1,…,𝑛

4. Row Players (Minnie's) Optimal randomized Strategy
Optimal randomized strategy and guaranteed upper bound on exp. loss for Minnie is ( 𝑞 1 , 𝑞 2 ,…, 𝑞 𝑚 ;𝑙) which is a solution to the LP: min 𝑙 s.t. 𝑖=1 𝑚 𝑎 𝑖𝑗 𝑞 𝑖 ≤𝑙 j=1,…,𝑛 𝑖=1 𝑚 𝑞 𝑖 =1 𝑞 𝑖 ≥0 i=1,…,𝑚

5. Crucial observation
Max's program and Minnie's program are each others duals!

6. Beyond simultaneous-move games - Kuhn's three-card poker (Section 11
Two players, Max and Minnie, each pays a dollar to get dealt a card from a 3-card deck with an Ace, a King and a Queen. Max either checks or bets an additional dollar. If Max checks, Minnie may either check or bet. If Max bets, Minnie may either fold or call. If Minnie bets, Max may either fold or call. If the last player to act folds, the player betting wins the pot. If the last player to act calls, the player with the highest card wins the pot.

7. Analyzing three-card poker
In three-card poker, the players do not act simultaneously and a random card is dealt to them. How does the theory apply to this kind of game? A deterministic strategy may be described by an instruction for what to do in every possible situation that may arise during the game.

8. A deterministic strategy for Max
If I get an Ace, I shall bet.
If I get a King, I shall check. If Minnie then bets, I shall call.
If I get a Queen, I shall check. If Minnie then bets I shall fold.
A deterministic strategy for Max (or Minnie) is a tree structure.

9. Matrix game formulation of three-card poker
For each deterministic strategy 𝑗 for Max and deterministic strategy 𝑖 for Minnie, define 𝑎 𝑖𝑗 as the expected (under the random cards dealt) gain for Max when these two deterministic strategies are used. A randomized strategy is a probability distribution over all deterministic strategies. Game matrix: 64×27 matrix. Using simplifications: Down to 4×8 matrix.

10. Unique optimal randomized strategy for Max
If I get the ace, with probability I shall bet. With probability , I shall check and call if Minnie bets.
Slowplaying!
If I get the king, I shall check. If Minnie bets, I shall call with probability and fold with probability
If I get a queen, I shall check with probability If I check and Minnie bets, I fold. But with probability I shall start by betting.
Bluffing!
Value of the game for Max: − (Intuitive reason: Minnie has a positional advantage.)

11. Conclusions for poker
Concealing your hand by bluffing and slowplaying is necessary rational behavior - even against completely rational opponents (who knows that you are rational).

12. General conclusions
The theory of matrix games can be used to analyze games where several rounds of actions are made, information is leaked “bit by bit" and external randomness enters the game.
... but the game matrix may become very large. The idea described gives a matrix with size exponential in the size of the game tree.
An alternative LP formulation gives a linear program of size linear in the game tree.
Heads-up, no-limit Texas Hold’em: 10160 nodes —more than the number of atoms in the known universe.

13. Game theory applied to algorithm analysis
Consider a randomized algorithm as follows:
Allowed to “flip coins" and continue execution based on the outcome.
Must always provide a correct answer when terminating (Las Vegas algorithm).
Example?
Reasonable measure of complexity: Expected execution time, on a worst case instance.
How to analyse this?

14. Formulation as a Game
We may view a randomized algorithm as a probability distribution on deterministic algorithms.
Requirements for this to work:
Only a finite number of possible deterministic algorithms.
Only a finite number of possible inputs.
Consider the following game:
Max chooses an input.
Minnie chooses an algorithm.
Minnie pays Max for every unit of execution time.

15. Yao's Principle
Yao's Principle: The expected complexity of the best randomized algorithm on a worst-case input equals the weighted average complexity for a worst-case distribution of inputs using the best deterministic algorithm.
Proof: Both are the value of the game we defined.

16. Using Yao's Principle
Suppose we wish to prove a lower bound on the complexity of a randomized algorithm.
By Yao's principle we may instead construct a “hard" input distribution and analyze the average case complexity of the best deterministic algorithm on this (a sometimes easier task).
The best lower bound possible can be obtained this way!

17. Example: Sorting
How many comparisons must be used to sort n elements? Ω(𝑛 log⁡𝑛) What about a randomized algorithm? By Yao's principle we may analyze a best deterministic algorithm on a chosen input distribution. We know that a deterministic sorting algorithm must use Ω 𝑛 log 𝑛 comparisons in the worst case. Almost the same proof shows also that Ω 𝑛 log 𝑛 comparisons are needed on the average to sort a permutation selected uniformly at random.