MIN-MAX algorithm

— history, variants, refinements

Przemysław Klesk˛
[email protected]

Department of Methods of Artificial Intelligence and Applied Mathematics

Table of contents

1 Game theory

2 Game trees and searching

3 Games of perfect information — algorithms

4 Games of perfect information with random elements

5 Games of imperfect information

6 References
Table of contents

1 Game theory

2 Game trees and searching

3 Games of perfect information — algorithms

4 Games of perfect information with random elements

5 Games of imperfect information

6 References
Game theory

Game theory
A branch of mathematics dealing with situations of conflict (strategic
situations), where a result of a participant depends on choices made by
himself and others. Sometimes, also called the theory of rational behaviours.
Apart from computer science, applied in the fields of sociology, economics,
military (historically earlier).

Important historical works:

Émil Borel, Applications for random games (fr. Applications aux Jeux de
Hasard), 1938.
John von Neuman and Oskar Morgenstern, Theory of Games and
Economic Behavior, 1944.
 Game theory

Notions

Game
A situation of conflict, where:
at least two players can be indicated,
every player has a certain number of possible strategies to choose from (a strategy
precisely defines the way the game shall be played by the player),
result of the game is a direct consequence of the combination of strategies chosen by
players.

Strategy
Complete set of decisions (about choices or moves) that a player has to make for all possible
states the game can reach.

It is often impossible to write down (memorize) a strategy because of its size (for typical games).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 5 / 77
 Game theory

Notions

Finite game
A game for which it is guaranteed that the game shall finish.

Zero-sum game
A game in which payoffs for all players (determined by result of game) sum
up to zero.

1
For chess the convention is: 0 (loss), 1 (win), 2 (draw); zero sum can be
obtained by a linear transformation: 2x − 1).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 6 / 77
 Game theory

Minimax Theorem

Minimax Theorem (von Neuman, 1928)

For every finite two-person zero-sum game there exists at least one optimal
mixed strategy. Therefore, there exists a game value v, such that by applying
the optimal strategy the first player guarantees for himself a payoff not worse
than v, while the second player guarantees for himself a payoff not worse
than −v.

For zero-sum games the minimax solution is identical with Nash equilibrium
(a broader notion).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 7 / 77
 Game theory

Example for theorem

B chooses b1 B chooses b2 B chooses b3

A chooses a1 3 −2 3
A chooses a2 −1 0 4
A chooses a3 −4 −3 1

Matrix of payoffs: W = {wij } for zero-sum game, where players A and B

make simultanous moves.
What are minimax choices for A and B?
Is it a stable solution?
Do there exist dominated choices?

Przemysław Klesk
˛ (KMSIiMS, ZUT) 8 / 77
 Game theory

Example for theorem

B chooses b1 B chooses b2 B chooses b3
A chooses a1 3 −2 3
A chooses a2 −1 0 4
A chooses a3 −4 −3 1

The minimax choice for A is a2 , because the worst possible result for A is then −1:

max min wij = −1. (1)

i j

The minimax choice for B is b2 , because the worst possible result for B is then 0:

min max wij = 0. (2)

j i

Solution (a2 , b2 ) is not stable, because if B believes that A chooses a2 then B shall choose b1
in order to obtain the payoff −1; then, if A believes that B chooses b1 then A shall choose a1
to obtain the payaoff 3, etc.
Dominated choices: a3 and b3 — regardless of opponent’s choice, the other choices are
better (more precisely: not worse). Hence, the matrix of payoffs can be reduced by
deleting third row and third column.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 9 / 77
 Game theory

Example for theorem

B chooses b1 B chooses b2
A chooses a1 3 −2
A chooses a2 −1 0
!
  q
p 1−p ·W· = −q − 2p + 6pq. (3)
1−q

Mixed strategy is a combination of choices — pure strategies — with certain probabilities

(frequencies).
The presence of pq term indicates the existence of saddle point.
∂· ∂·
By demanding ∂p = 0 and ∂q = 0, one obtains the solution:

1 1
p= , q= . (4)
6 3
Game value: v = − 13 .
Formally, when P and Q represent mixed strategies (as vectors of probabilities), then:

max min PT · W · Q = min max PT · W · Q = v. (5)

P Q Q P

If there exist more than one optimal mixed strategy then infinitely many of them exist.
Przemysław Klesk
˛ (KMSIiMS, ZUT) 10 / 77
 Game theory

Nash Equilibrium (NEQ)

John Nash, born 1928, Nobel prize winner in 1994 (in economics).

Informally
In a multi-player game, we say that a certain set of strategies from particular
players constitutes the Nash equilibrium, if and only if each of those strategies
is the best response for all remaining ones and none of the players can gain
by changing its own strategy with other strategies kept fixed.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 11 / 77
 Game theory

Nash Equilibrium (NEQ)

Formally
In a game with n players, let Si denote the set of possible strategies for i-th player.
Let S denote the space of all strategies as the cartesian product of sets of strategies from
particular players:
S = S1 × S2 × · · · × Sn
For any set of strategies (s1 , . . . , sn ) from particular players, let Wi (s1 , . . . , sn ) determine the
payoff for the i-th player. Therefore, Wi is a function:

Wi : S → R.

Let W denote a vector function: W = (W1 , . . . , Wn ).

A game can be understood as a pair (S, W).
We say that a set of strategies (s∗1 , . . . , s∗n ) constitutes the Nash equilibrium if and only if:

∀i ∀si ∈ Si : Wi (s∗1 , . . . , s∗i , . . . , s∗n ) > Wi (s∗1 , . . . , si , . . . , s∗n ). (6)

Przemysław Klesk
˛ (KMSIiMS, ZUT) 12 / 77
 Game theory

Nash Equilibrium (NEQ)

Another way to define NEQ is that s∗i can be viewed as the solution of:

max Wi (s∗1 , . . . , si , . . . , s∗n ) (7)

si ∈Si

for all i.
The idea of NEQ can be applied to analyze or predict what happens when several players
(parties, institutions) must make decisions simultaneously, and when the outcome
depends on all those decisions. The outcome cannot be predicted when analyzing the
decisions seperately (in isolation).
NEQ does not have to indicate the best result for the group (the best sum of results) and
may seem irrational for an outside observer (e.g. prisonner dilemma, Braess paradox).
In many cases, players could improve their group result if they agree strategies different
from NEQ (e.g. business cartels instead of free-market competition).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 13 / 77
 Game theory

Braess Paradox

B
1+ n
100 2

A 0.25 D

2 1+ n
100
C

Problem
Assuming selfishness and rationality of drivers, find the expected traffic flow (the NEQ) for 100
drivers travelling from A to D in two cases: (1) when edge BC does not exist, (2) when edge BC
exists.
Treat the problem as a game where every player (driver) has 2 or 3 strategies, respectively: ABD,
ACD and possibly ABCD. The payoff is the time of travel for road selected. For AB and BD
edges, the n parameter denotes the number of players who selected the edge as a road fragment.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 14 / 77
 Game theory

Braess paradox — solution

p, q, r — number of drivers choosing strategies, respectively: ABD, ACD,

ABCD.

Case 1

p q
1 + 100 + 2 = 1 + 100 + 2;


 (8)
p + q
 = 100.

Solution: p = q = 50, road cost (game value) v = 3.5.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 15 / 77
 Game theory

Braess paradox — solution

Case 2

p+r q+r p+r q+r
1 + 100 + 2 = 1 + 100 + 2 = 1 + + 0.25 + 1 + 100 ;

 100
 (9)
p + q + r
 = 100.

Solution: p = q = 25, r = 50, road cost (game value) v = 3.75.

Drivers would travel shorter if they agreed no to use the BC fragment.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 16 / 77
Table of contents

1 Game theory

2 Game trees and searching

3 Games of perfect information — algorithms

4 Games of perfect information with random elements

5 Games of imperfect information

6 References
 Game trees and searching

Game trees

For a certain state s assume there exists n possible choices (moves,

manipulations, actions): a1 , a2 , . . . , an causing new states to arise from s,
respectively: s1 , s2 , . . . , sn . For each of those states there again exist further
possible choices. By coninuing this procedure a tree structure arises naturally.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 18 / 77
 Game trees and searching

Search problems — difficulties

too large search space — when generation of complete tree is

impossible or too expensive computationally or memory-wise
(exponential growth).
games of imperfect information — players do not have complete
information about the state of game (e.g. cards held by opponents,
opponents’ letters in Scrabble, opponents’ military units, etc.).
random factors — when random generators are an element of game
(e.g. rolling dice, shuffling cards, random events of environment).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 19 / 77
 Game trees and searching

Game complexity measures

1 State-space complexity — the number of legal positions in a game that

can be reached from the initial position. Often, this number can be
upperbounded, when it is easire to make an estimate by taking into
account illegal positions.
2 Game tree size — the number of different games that can be played;
equivalently, the number of leaves in the tree with the root representing
the initial position. The same states reached by different paths (different
moves order) are calculated multiple times. The number can be
upperbounded by taking into account the tree growth with illegal
moves included.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 20 / 77
 Game trees and searching

Game complexity measures

3 Decision complexity — the number of leaves in the smallest decision tree
which is capable to determine the game value for the initial position.
4 Game tree complexity (ang. game tree complexity) — the number of
leaves in the smallest decision tree of full width capable to determine the
gamve value for the initial position. Such a tree takes into account all
possibilities for both players. The number corresponds to the required
number of operations in a MIN-MAX search.

Decision tree (from the first player perspective)

A subtree (of the full game tree) in which all states have labels: win, draw, loss.
A state becomes labeled as a win when any of its decendants is labeled a win.
A states becomes labeled as a loss when all its decendants are labeled a loss.
A state becomes labeled as a draw when at least one of its decendants is labeled a draw and
all remaining ones are labeled a loss.
A decision tree takes into account all possibilities for one player and only one possibility (the
best response) for the second player (corresponds roughly to the number of operations in a α-β
pruning optimisitc search).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 21 / 77
 Game trees and searching

Complexity estimates for some games

board size state-space game tree

game (number of cells) complexity complexity
tic-tac-toe 9 103 105
connect 4 42 1013 1021
English checkers 32 1020 1031
hex
also: Nash or John 121 1056 ?
chess 64 1047 10123
connect 6 361 10172 10140
backgammon 28 1020 10144
Go 361 10171 10360

Przemysław Klesk
˛ (KMSIiMS, ZUT) 22 / 77
 Game trees and searching

Chinook project (checkers)

English checkers (8 × 8, kings move in any direction but by 1 cell only).

Project started in 1989. Goal: beating human in world championships of checkers.
Authors: (back, from the left) Yngvi Bjørnsson, Neil Burch, Rob Lake, Joe Culberson
(from, from the left) Paul Lu, Jonathan Schaeffer, Steve Sutphen. Universities: Rejkjavik
(Iceland), Alberta (Kanada), Hakodate (Japan).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 23 / 77
 Game trees and searching

Chinook project (checkers)

In 1990 the program was given the right to participate in championships and playing
against human.
The program lost the championships in 1992, but won in 1994. In 1996, the possibility of
Chinook’s participation was withdrawn (program was much stronger than any other
human player).
Search space of order: 5 · 1020 . Database (library) with information about the best move
(continuation) for many states.
29.04.2007 — authors of the project anounce English checkters a solved game! Black
(starting the game) have a draw guarantee with a perfect play. White is also guaranteed
with a draw, regardless of the first move by black.
Until today, it is the “largest” solved mind game.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 24 / 77
 Game trees and searching

Chinook project — Samuel legacy

Arthur Samuel wrote a checkers engine program in 1950 under a project sponsored by
IBM.
In 1952, a genetic element of self-training was added — two instances of the program
were playing against one another with repetitions. The thus evolved program was
beating amateurs and intermediate players.
After program presentation in 1956 for IBM stakeholders, the IBM stock quotes rose by 15
points.
In 1962 the program played a public match against Robert Nealy (a blind checkers
master), in which the program won. The win was given much publicity. Nealy was not a
world-class master.
In effect, a false belief was spread, that English checkers were already a solved game at
the time. Bjørnsson had troubles obtaining his grant for Chinook research in the 80s
because of that.
A year later, Samuel’s program lost a rematch: 1 loss, 5 draws. In 1966 the program lost 8
consecutive games against top level players: Derek Oldbury and Walter Hellman.
Przemysław Klesk
˛ (KMSIiMS, ZUT) 25 / 77
Table of contents

1 Game theory

2 Game trees and searching

3 Games of perfect information — algorithms

4 Games of perfect information with random elements

5 Games of imperfect information

6 References
 Games of perfect information — algorithms

MIN-MAX algorithm

Procedure mmEvaluateMax(s, d, D)
1 If s is a terminal then return h(s) (position evaluation).
2 v := −∞.
3 For all states t being descendants of s:
1 v := max{v, mmEvaluateMin(t, d + 21 , D)}.
4 Return v.

Procedure mmEvaluateMin(s, d, D)
1 If s is a terminal then return h(s) (position evaluation).
2 v := ∞.
3 For all states t being descendants of s:
1 v := min{v, mmEvaluateMax(t, d + 21 , D)}.
4 Return v.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 27 / 77
 Games of perfect information — algorithms

α-β pruning algorithm

Many independent discoverers: Samuel (1952), McCarthy (1956), Newell and Simon
(1958).
During analysis two values are propagated down and up the tree:
α — guaranteed (so far) payoff for maximizing player,
β — guaranteed (so far) payoff for minimizing player.
Out-most execution for root forces α = −∞, β = ∞.
Children-nodes (and their subtrees) are analyzed while α < β.
Whenever α > β, one should stop considering successive children (and their subtrees) —
they will not affect the outcome for the whole tree; they would be a result of a
non-optimal play by some of players.
Inoptimistic case, the gain in complexity with respect to MIN-MAX is from O(bD ) to
 √ 
O bD/2 = O bD , where b — branching factor (constant or average). E.g. for chess b ≈ 40.
Owing to the gain one may search deeper.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 28 / 77
 Games of perfect information — algorithms

α-β pruning algorithm ( fail-hard version

returns a result within [α, β])
Procedure alphaBetaEvaluateMax(s, d, D, α, β)
1 If s is a terminal then return h(s) (position evaluation).
2 For all states t being descendants of s:
1 v := alphaBetaEvaluateMin(t, d + 21 , D, α, β).
2 If β 6 v then return β. (cut-off)
3 α := max{α, v}.
3 Return α.

Procedure alphaBetaEvaluateMin(s, d, D, α, β)
1 If s is a terminal then return h(s) (position evaluation).
2 For all states t being descendants of s:
1 v := alphaBetaEvaluateMax(t, d + 12 , D, α, β).
2 If v 6 α then return α. (cut-off)
3 β := min{β, v}.
3 Return β.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 29 / 77
 Games of perfect information — algorithms

Illustration for α-β pruning — example 1

α = −∞, 5
β=∞
MAX

α = −∞ α=5
β = ∞, 5 β = ∞, 10, 5
MIN

α = −∞, 5 α = −∞, 6 α = 5, 7, 10 α=5

β=∞ β=5 β=∞ β = 10
MAX

5 4 6 ∗ 7 10 4 4

Przemysław Klesk
˛ (KMSIiMS, ZUT) 30 / 77
 Games of perfect information — algorithms

Why is optimistic complexity O(bD/2)?

In traditional MIN-MAX:
O(b · b · · · b) = O(bD ). (10)
| {z }
D-times b

In α-β with even number of tree levels, optimistically we have:

 
O(b · 1 · b · 1 · · · b · 1) = O bD/2 . (11)
| {z }
D/2-times b

Explanation: we need to build all possible children for the first player, but we assume
moves are optimally ordered therefore, in each child, alread the first move of the second
player causes a cut-off (α > β) and further moves are discarded as non-optimal ones. And
so forth recursively.
3
There exist estimates for the average case (random order of children), yielding O(b 4 D ).
In chess for: b = 40 and D = 12 (12 half-moves), the proportion of visited states for
pessimistic ordering to visited states for optimistic ordering is 406 , i.e. of 109 order.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 31 / 77
 Games of perfect information — algorithms

α-β pruning algorithm (fail-soft version, result

can fall outside [α, β])
Procedure fsAlphaBetaEvaluateMax(s, d, D, α, β)
1 If s is a terminal then return h(s) (position evaluation).
2 For all states t being descendants of s:
1 v := fsAlphaBetaEvaluateMin(t, d + 21 , D, α, β).
2 α := max{α, v}.
3 If α > β then return α. (cut-off)
3 Return α.

Procedure fsAlphaBetaEvaluateMin(s, d, D, α, β)
1 If s is a terminal then return h(s) (position evaluation).
2 For all states t being descendants of s:
1 v := fsAlphaBetaEvaluateMax(t, d + 21 , D, α, β).
2 β := min{β, v}.
3 If α > β then return β. (cut-off)
3 Return β.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 32 / 77
 Games of perfect information — algorithms

Knuth-Moore theorem (1975)

Article: Knuth D.E., Moore R.W., „An Analysis of Alpha-Beta Pruning”, Artificial Intelligence,
1975.

„Theorem about α-β window”

Let v∗ denote the true (exact) game result obtained via MIN-MAX procedure. Let v denote the result of
fsAlphaBeta (fail-soft) procedure executed for a root node with parameters α, β. Then, possible are three
cases:
1 α<v<β ⇒ v = v∗ ,
2 v 6 α (failing low) ⇒ v∗ 6 v (v is an upper bound for v∗ ),
3 β 6 v (failing high) ⇒ v6 v∗ (v is a lower bound fo v∗ ).

In particular, a consequence: fsAlphaBeta(root, −∞, ∞) = v∗ .

The theorem is useful for building more advanced search algorithms: Negascout, MTD-f based on
so called zero search windows.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 33 / 77
 Games of perfect information — algorithms

Quiescence algorithm

Tries to mimic the intuition of human players by: expanding loud nodes/leaves, and not
expanding quiet nodes/leaves (instead, immediate return of position evaluation).
Partially solves the horizon effect problem.
We call a position quiet if no sudden changes of position evaluation between given state
and its descendants occur (e.g. takes/captures).
An assesment if given state is quiet or not may not be easy; it may require a heuristic in
itself. Importantly, such an assement must be faster than expanding a new tree level.
Quiescence does not have to be applied only at leaves level, but already sooner. Current
depth can be used as an element for assesment of quietness, i.e. the deeper we are the
larger tendency to leave quiet states not expanded.
Exact description e.g. in: D. Laramée, Chess Programming Part V: Advanced Search, 2000.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 34 / 77
 Games of perfect information — algorithms

Sorting children nodes in α-β pruning

It is worth to sort children nodes, especially “at the top of the tree”,
where it is fairly cheap (there are fewer states at higher levels) and
might result in greater savings deeper.

Sorting heuristics
in chess: „captures first”,
in many card games e.g. bridge: „extreme cards first, middle ones
later”; e.g. a hand A, D, 8, 6, 5, 2 can be sorted to: A, 2, D, 5, 8, 6; the order
can be also arranged accoring to the position of the player within a trick
(e.g. the last player typically plays a high card first, the second player
typically plays a low card first, etc.),
sorting according to position evaluation — evaluate and sort children
immediately based on their position evaluation, before running the
recurrence downwards.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 35 / 77
 Games of perfect information — algorithms

Sorting children nodes in α-β pruning

Sorting heuristics (. . . continued)

„refutation table” — a table memorizing so called (refutation moves) — the ones causing
cut-offs, usually at shallow levels; this allows to consider these moves in first order in next
search iterations (progressive search or iterative search). Historically, first chess programs
with serious memory limitations typicall kept b2 refutation moves for root grand children,
or possibly b2 + b4 by considering two more levels. Also called best (principal) continuation
table.
„killer heuristic” — memorizes for each level a short list of killer moves (usually 2, 3
moves) causing cut-offs and taking those into account in first order for other states at the
same level; intuition: if some move is good for some state it might be also good for
another (similar) state.
If a non-killer moved caused a cut-off it is introduced to the list in the place of the
“weakest” killer-move so far (counting approaches).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 36 / 77
 Games of perfect information — algorithms

Transposition table
The name comes from chess and represents the possibility of obtaining
the same position (state) by different sequences of moves.
If the downwards recurrence for such a state has already been
calculated then one can save time by using a ready-made result.
Often, implemented as a hash map (time efficience, analogical to Closed
set in A∗ , BFS, etc.). The keys in hash map are states themselves or their
abbreviations — hash codes (e.g.. in chess positions of at most 32 pieces
are required plus information about castling and en passant capture
possibilities).
Conditions for reusing a state from the transposition table:
Depth of the state in transposition table not deeper than for the tested state (so that
the ready score comes from an analysis of a tree of equal or greater depth).
α-β window for the state in transposition table must be not more narrow than
current ones (so that the ready score was not affected by more cut-offs)
Sometimes applied as the book of openings or endgames (chess,
checkers).
Przemysław Klesk
˛ (KMSIiMS, ZUT) 37 / 77
 Games of perfect information — algorithms

Scout algorithm)

Historically, an idea due to J. Pearl’a (1980): for reconnaissance, one

may preliminarily and less expensively test if the current payoff can be
improved. Two recursive procedures eval(·) and test(·); the second
returns a boolean inficating if an improvement is possible.
The idea developed further by A. Reinefeld, „An Improvement to the
Scout Tree Search Algorithm”, ICCA Journal, (1983); introduced are so
called zero α-β windows (also: null windows, scout windows).
If payoffs are integers, a zero window takes place when

α + 1 = β. (12)

The idea “cooperates” with the Knuth-Moore theorem.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 38 / 77
 Games of perfect information — algorithms

Scout algorithm

Definition
We say that a given α-β window succeeded if v returned by the fsAlphaBeta procedure (fail-soft) is
such that: α < v < β. It implies (Knuth-Moore) that the true game value v∗ equals v.

Definition
We say that a given α-β window failed low if v returned by the fsAlphaBeta procedure (fail-soft) is
such that: v 6 α. It implies (Knuth-Moore) that v is an upperbound on the true game value:
v∗ 6 v.

Definition
We say that a given α-β window failed high if v returned by the fsAlphaBeta procedure (fail-soft) is
such that: β 6 v. It implies (Knuth-Moore) that v is a lowerbound on the true game value: v 6 v∗ .

Przemysław Klesk
˛ (KMSIiMS, ZUT) 39 / 77
 Games of perfect information — algorithms

Scout algorithm

The more narrow the imposed window is, the greater chance to generate more cut-offs.
Only the first child of each state is analyzed by a full α-β window, the second and
successive children are analyzed by a zero window, i.e. α-(α + 1) or (β − 1)-β, respectively
for MAX, MIN states.
A zero window must fail either way.
If a zero window imposed on a child of MAX state failed low then we do not have to care
— the payoff for the maximizing player could not be improved within this child
(computational gain, probably greater number of cutoffs should appear within the
subtree of that child).
If a zero window imposed on a child of MAX failed high then we have to repeat the
search for that child (computational loss) with a wider window v-β in order to obain an
exact result for given subtree. Remark: still, the window (for the repeated calculation) is
more narrow than the original one: α-β.
The last two remarks are suitably opposite for MIN states.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 40 / 77
 Games of perfect information — algorithms

Scout algorithm

Procedure scoutMax(s, d, D, α, β)
1 If s is a terminal then return h(s) (position evaluation).
2 b := β.
3 For all states t being descendants of s:
1 v := scoutMin(t, d + 12 , D, α, b).
1
2 If t is not the first child and D − d > 2 · 2 and b 6 v (failing high) then:
1
1 v := scoutMin(t, d + 2 , D, v, β). (repeat search with a wider window)
3 α := max{α, v}.
4 If α > β then return α. (cut-off)
5 b := α + 1.
4 Return α.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 41 / 77
 Games of perfect information — algorithms

Scout algorithm

Procedure scoutMin(s, d, D, α, β)
1 If s is a terminal then return h(s) (position evaluation).
2 a := α.
3 For all states t being descendants of s:
1 v := scoutMax(t, d + 12 , D, a, β).
1
2 If t is not the first child and D − d > 2 · 2 and v 6 a (failing low) then:
1
1 v := scoutMax(t, d + 2 , D, α, v). (repeat search with wider window)
3 β := min{β, v}.
4 If α > β then return β. (cut-off)
5 a := β − 1.
4 Return β.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 42 / 77
 Games of perfect information — algorithms

Illustration for Scout — example 1

α = −∞, 5
β=∞
MAX

α = −∞ α=5
β = ∞, 5 β = 6, 5
MIN

α = −∞, 5 α = 4, 6 α = 5, 7 α=5
β=∞ β=5 β=6 β=6
MAX

5 4 6 ∗ 7 ∗ 4 4

Przemysław Klesk
˛ (KMSIiMS, ZUT) 43 / 77
 Games of perfect information — algorithms

Illustration for Scout — example 2

α = −∞, 5, 6 (6 6 v∗ ), fail high, bo 6 = b > v = 6

β=∞
MAX

α = −∞
β = ∞, 5
α=5 ?
β=6
MIN

α = −∞, 5 α = 4, 6 α = 5, 7 α = 5, 8
β=∞ β=5 β=6 β=6
MAX

5 4 6 ∗ 7 ∗ 4 8

Przemysław Klesk
˛ (KMSIiMS, ZUT) 44 / 77
 Games of perfect information — algorithms

Scout algorithm

The condition D − d > 2 · 12 checks if we are at a depth of 2 halfmoves

away or deeper from the search horizon. If so then it is not necessary to
repeat the search despite fail situation, because the algorithm works
accurately at such depths.
The algorithm works well in the progressive search scenario when
cooperating with children-sorting heuristics, especially with killer
heuristic, when the best move (and best path) are often considered first.
Due to that fact, the scenario is often called Principal Variation Search.
Experiments indicate that computational gains implied by zero
windows are more frequent cut-offs are typically greater than
computational losses induced by repeated searches.
Reinefeld’s experiments showed that for trees with a branching factor
within b ∈ {20, . . . , 60} (e.g. chess), Scout visitis on average about 20%
fewer tree leaves. Tests for depths: 4, 5 halfmoves.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 45 / 77
 Games of perfect information — algorithms

Negamax algorithm

Fact:

∀n ∈ N ∀a1 , a2 , . . . , an min{a1 , a2 , . . . , an } = − max{−a1 , −a2 , . . . , −an }.

Owing to the above one may simplify implementation by replacing two

twin-procedures by a single one in all algorithms: Negamax (in fact
negaAlphaBeta), Negascout, etc.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 46 / 77
 Games of perfect information — algorithms

Negamax algorithm

Procedure negaMax(s, d, D, α, β, color)

1 If s is a terminal then return color · h(s) (position evaluation).
2 For all states t being descendants of s:
1 α := max{α, −negaMax(t, d + 12 , D, −β, −α, −color)}.
2 If α > β then return α.
3 Return α.

The out-most call for the root is:: negaMax(root, 0, D, −∞, ∞, 1).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 47 / 77
 Games of perfect information — algorithms

Illustration for Negamax — example 1

α = −∞, 5
β=∞
MAX

α = −∞, −5 α = −∞, −10, −5

β=∞ β = −5
-MAX

α = −∞, 5 α = −∞, 6 α = 5, 7, 10 α=5

β=∞ β=5 β=∞ β = 10
MAX

5 4 6 ∗ 7 10 4 4 −h(s)

Przemysław Klesk
˛ (KMSIiMS, ZUT) 48 / 77
 Games of perfect information — algorithms

Negascout algorithm

Procedure negascout(s, d, D, α, β, color)

1 If s is a terminal then return color · h(s) (position evaluation).
2 b := β.
3 For all states t being descendants of s:
1 v := −negascout(t, d + 21 , D, −b, −α, −color).
1
2 If t is not the first child and D − d > 2 · 2 and b 6 v (failing high) then:
1
1 v := −negascout(t, d + 2 , D, −β, −v). (repeated search with a wider window)
3 α := max{α, v}.
4 If α > β then α. (cut-off)
5 b := α + 1.
4 Return α.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 49 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (1)

Master thesis: Mateusz Bożykowski, Implementation of self-teaching program for checkers, WI,
2009.

Checkers: international (a.k.a. Polish) (100 squares, players have 20 pawns each),
Brazilian (64 squares, 12 pawns per player), English (64 squares, 12 pawns, kings moving
by 1 square).
Implementation of α-β pruning including transpostion table and Quiescence.
Multiple programs playing checkers according to different heuristics (position evaluation
functions) and competing withing a genetic evolution.
Individuals can be identified with heuristics evaluation positions (various AIs). The
simplest heuristics materialistic, symmetrical:

h = w1 Pp + w2 Kp − w1 Po − w2 Ko , (13)

where P, K denote the number of pawns and kings, respectively; whereas indexes p, o
stand for player and opponent, respectively. The parameters under genetic optimization
are w1 , w2 .
Parameters coded as integers, initially picked on random from {−100, . . . , 100}. During
evolution, the parameter have not gone outside the initial range.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 50 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (2)

It is difficult to fairly assign a numeric fitness or rank to such due to the possibility of
three-ways-draw: A wins against B, B wins agains C, C wins against A.
Tournament selection comes to mind. Difficulties: (1) frequent draws (who should be
then selected to next population?), (2) possibility of losing the best individual, (3) pointing
out the best individual in the final population.
Final approach: tournaments for population sizes being powers of 2 n = 2m . Individuals
paired randomly into matches, n/2 of population filled up with winners (in case of a draw
added was 1 child of crossed parents), the rest of population filled up iteratively with
winners of matches between winners added before.
The winner of the very last match (the winner among winners) considered the best
individual in the final population.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 51 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (3)

Linear cross-over — draw a randome number α ∈ (0, 1), a child C of parents A and B is
obtained as:
wi (C) = αwi (A) + (1 − α)wi (B). (14)
Uniform cross-over — for all wi we randomly decide from which parent it comes (and
copy it unchanged). Additionally, mutation is suggested.
Mutation of constant radius — each gene (weights) is added a random value from
{−20, . . . , 20}. The probability of mutation is lineaerly decreased from 0.9 in the first
iteration to 0.3 in the last.
Different depths for tree analysis were set up based on the GA iteration (the later iteration
the more accurate the analysis should be).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 52 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (4)

Studied heuristics:
materialistic, unsymmetrical

h = w1 Pp + w2 Kp + w3 Pp + w4 Kp (15)

materialistic-positional, symmetrical (in general h = hp − ho )

hs = w1 Ps + w2 Ks + w3 Cs + w4 Ss + w5 OTs + w6 PTs + w7 KDs , (16)

where: C — number of pieces in the board center, S — number of pieces at sides, OT —

number of pawns at promotion row for opponent (defence of promotion), PT — number
of pawns one square away from own promotion, KD — number of kings on main
diagonal.
extended materialistic-positional, symmetrical (in general h = hp − ho )

 
Ds
hs = w1 Ps + · · · + w7 KDs + w8 Ms + w9 + w10 KD2s , (17)
Ps + Ks

where: M ∈ {0, 1} — extra reward for turn to move, D — number of doubled pieces
(touching by corners), KD2 — number of kings on opposite double diagonal.
Przemysław Klesk
˛ (KMSIiMS, ZUT) 53 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (5)

Studied heuristics:
materialistic-row-wise — pawns have different values according to rows they occupy.

N−1
X
hs = wi Pi + wN K, (18)
i=1

where: N — number of rows on the board, Pi — number of pawns in i-th row.

materialistic-structural, symmetrical — created based on an observation that row-wise
heuristics was significantly weaker than extended materialistic-positional (an attempt to
improve the row-wise heuristics), and an observation that the number of doubled pieces
had a negative impact on evaluation:

hs = w1 Ps + w2 Ks + w3 OTs + w4 PTs + w5 Is + w6 Fs (19)

where: I — number of immortal (non-capturable) pieces, F — number of frozen pieces.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 54 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (6)

Obtained results of optimization:

materialistic, unsymmetrical:

Pp = 5, Kp = 12, Po = −7, Ko = 10.

Comment: aggressive play with pawns — capturing opponent’s pawn increases the
evaluation by 2. Careful play with kings in endgame, because own kings are worth more
than opponents’.
materialistic-positional, symmetric:

Ps = 24, Ks = 65, Cs = 1, Ss = 1, OTs = −11, PTs = 27, KDs = 0.

Comment: surprisingly the pawns defending promotion line are evaluated negatively,
and kings on main diagonal as immaterial.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 55 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (7)

Obtained optimization results:
extended materialistic-positional, symmetrical:

Ps = 5, Ks = 16, Cs = 0, Ss = 0, OTs = 0, PTs = 6, KDs = 0, Ms = 0, Ds = −7, KD2s = 0.

Comment: surprisingly, most of parameters were zeroed; of relevance seem to be: pawns
just before promotion and penalty for doubled pieces.
extended materialistic-row-wise, symmetrical:

P1 = 2, P2 = 1, P3 = 2, P4 = 2, P5 = 2, P6 = 2, P7 = 1, P8 = 3, P9 = 6, K = 12.

Comment: interestingly, values in rows from 3 to 6 are equal; a pawn in 8-th row already
starts to be worth more.
extended materialistic-structural, symmetrical:

Ps = 13, Ks = 85, OTs = 0, TTs = 6, Is = 1, Fs = −1.

Comment: one can note that immortal pieces have a positive impact, frozen ones a
negative impact (before both those elements were ‘hidden’ in the doubling parameter D).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 56 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (7)

Comparison against free checkers programs:

Dam 2.0 — highly assessed program, developed since 1987. Lack of possibility to set up
the tree depth — inaccurate comparison. Test: best AI from Warcabnik with 7 halfmoves
horizon vs. successive levels of Dam 2.0. Warcabnik winse agains levels Beginner A, B,
draws with Beginner C, D, starts losing against Beginner E (in fact, it loses endgames
involving kings).
Little polish v0.7 — a program by Piotr Beling for Brazilian checkers. The opponent was
set up limited to 1 s per move. Warcabnik drew, despite the opponent was at times
analyzing even 18 halfmoves deep. In matches against stronger AIs (> 1 s) Warcabnik was
losing.
Windames 3D — allows to set up 9 difficulty levels. Warcabnik wins with first three and
loses against next ones. Since times to move at fourth level are about equal, one may
suspect that the depth of analysis is then similar.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 57 / 77
 Games of perfect information — algorithms

Warcabnik (Mateusz Bożykowski) (8)

Comparison against free checkers programs:

Warcaby v.1.2 — a program by Marcin Źółtkowski for Polish and Brazilian checkers. It
allows to set up the tree depth (fair comparison). Tests for 6 halfmoves (max. allowed for
opponent). Warcabnik won in both variants of checkers and was much faster (approx. 2 s
per move, while opponent was taking 30 s per move). By setting 1 halfmove less
Warcabnik was still able to draw in Brazilian checkers.
Gui Checkers 1.05+ — a program by Jon Kreuzer for English checkers. It allows to set up
both depth and time. Tests for 10 halfmoves — Warcabnik was losing even after allowing
him for 1 halfmove more. Conclusion: studied heuristics are not sufficient.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 58 / 77
 Games of perfect information — algorithms

Bridge — “double dummy” problem

Double dummy
A version of bridge as a game of perfect information. Helpful for analysis of optimal bridge play
at the moment the whole board (all four hands) is known. Existing programs: Deep finesse, GIB.

Bachelor thesis: Katarzyna Kubasik, Application of game tree searching algorithms for finding
minimax points in “double dummy” bridge, WI, 2011.
Implementation of α-β pruning with the use transposition table.
Despite 4 players: N, E, S, W; alternate players (N, S) and (E, W) constitute pairs which
can be identified with two players: maximizing and minimizing.
A play by each player constitutes a new level in the search tree. The full tree has
4 · 13 = 52 levels.
When searching MAX, MIN do not have to alternate — one has to check which side took
the last trick (this side will again have the next move).
Improving elements: checking current sequences (e.g. a configuraiton 6, 4, 2, becomes a
sequence once other players have used the cards 5 and 3); sorting moves according to
heuristics: „extreme cards first, middle ones later”.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 59 / 77
 Games of perfect information — algorithms

Example of “double dummy” problem

♠ A10x
♥ A10x
♦ xxxx
♣ xxx
♠ KQJ ♠ xxxxx
N
♥ KQJ ♥ xxxxx
W E
♦ xxx S ♦ xx
♣ xxxx ♣x
♠ xx
♥ xx
♦ AKQJ
♣ AKQJ10

Contract: 6 no trump by NS. First lead by W: K♠. How NS should play to

take 12 tricks?

Przemysław Klesk
˛ (KMSIiMS, ZUT) 60 / 77
 Games of perfect information — algorithms

“Double dummy” problem — solution (1)

Player N should duck (not take with ace) the first trick. After any continuation by W, a squeeze
in spades and clubs shall take place in the endgame. E.g. after a continutation of Q♠.

♠ 10
♥ A10
♦
♣
♠J
N
♥ KQ
W E immaterial
♦ S
♣
♠
♥ xx
♦
♣ 10
S now playes 10♣ and W is squeezed. Without the initial duck at trick one the squeeze won’t
take place — the first play by N has a consequence 40 levels deeper in the tree!
Przemysław Klesk
˛ (KMSIiMS, ZUT) 61 / 77
 Games of perfect information — algorithms

“Double dummy” problem — solution (2)

After another continuation (an other second lead by W), e.g. K♥ the following endgame occurs:

♠ A10
♥ 10
♦
♣
♠ QJ
N
♥J
W E immaterial
♦ S
♣
♠ xx
♥
♦
♣ 10

Again, S plays 10♣ and W is squeezed.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 62 / 77
Table of contents

1 Game theory

2 Game trees and searching

3 Games of perfect information — algorithms

4 Games of perfect information with random elements

5 Games of imperfect information

6 References
 Games of perfect information with random elements

Expectiminimax algorithm (Michie, 1966)

Dedicated for games where result partially depends on players’ choices

and partially on random elements: e.g. dice roll, coin toss, polling a
card from a shuffled deck, etc.
Idea: apart from regular nodes performing MIN, MAX operations, one
should introduce CHANCE nodes, which perform arithmetic average
or weighted average operations.
Places of occurence of CHANCE nodes depends on game rules. In
particular, if e.g. everly player rolls dice before his move than the
following interlace shall take place: CHANCE - MAX - CHANCE - MIN
- ···.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 64 / 77
 Games of perfect information with random elements

Illustration for Expectiminimax

Przemysław Klesk
˛ (KMSIiMS, ZUT) 65 / 77
 Games of perfect information with random elements

Expectiminimax algorithm

evaluateState(s, d, D) (mixed recursive procedure)

1 Calculate position evaluation h := h(s).
2 If h = ±∞ or d = D then return h.
3 If s is a MIN state:
1 v := ∞.

2 For all t being descendants of s: v := min{v, evaluateState(t, d + 12 , D)}.

4 If s is a MAX state:
1 v := −∞.

2 For all t being descendants of s: v := max{v, evaluateState(t, d + 12 , D)}.

5 If s is a CHANCE state:
1 v := 0.

2 For all t being descendants of s: v := v + P(t) · evaluateState(t, d + 21 , D).

6 Return v.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 66 / 77
 Games of perfect information with random elements

Example: backgammon

Przemysław Klesk
˛ (KMSIiMS, ZUT) 67 / 77
 Games of perfect information with random elements

Example: backgammon

2+6−1

Two dices are rolled — number of possibilities: 21 = 2 . Branching at CHANCE level:
b = 21.
For n = 15 pawns, in effect of a „typical roll” (non “double”) a player can either select one
pawn to move (outcomes sum), or select two pawns (individual outcomes). Number of
possibilities: n(n − 1) = 210.
In case of “doubles” (the same outcome on both dices) a player has 4 single moves (of
same value) at disposal. Number of possibilities: 4+n−1 = 3060. Doubles occur with 16


4
probability.
Field blockages significantly reduce the number of possibilities. The average branching is
estimated to be approx. 400.
As depth increases the probability of some state decreases exponentially fast — therefore,
long-term forcasts are of little value.
The TDGammon program visits only 4 halfmoves of depth but has a very complex
heuristics for position evaluation (sufficient for a play at master’s level).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 68 / 77
Table of contents

1 Game theory

2 Game trees and searching

3 Games of perfect information — algorithms

4 Games of perfect information with random elements

5 Games of imperfect information

6 References
 Games of imperfect information

Imperfect information

E.g. in card games, where we do not know opponents’ cards.

Theoretically, one might calculate the probability of every possible hand (situation) — “a
dice with a great number of faces, rolled once at game start”.
Idea: calculate payoffs for all players on all hands and choose the play (move) with the
greatest expected payoff.
Example: GIB — currently, strongest bridge playing problem; generates at least 100 hands
(Monte-Carlo sampling) compliant with the so-far information and chooses the play
which on average leads to the greates number of taken tricks.
Special case: if there exists a play which on every hand leads to the greatest payoff than
this play is optimal.
Remark: the above intuion can be false and may lead to errors if ‘the inequality’ is not
strict (i.e. there exist other playes of equally good average) and if we do not observe
variance.
In fact, it may be necessary to run multiple repetitions of Monte-Carlo sampling, each
repetition made before successive plays (before each successive search).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 70 / 77
 Games of imperfect information

False intuition — example 1

Exemplary 4-card endgame. “No trump” contract played. Players must

follow the suit played.
Ceratin (known for sure) cards for player A: K♠, Q♠, A♦, A♣
Ceratin (known for sure) cards for player B: A♠, J♠, A♥.
Assume the fourth card for player B is unknown: K♦ or K♣.
Evaluate the play of K♠ from the perspective of A.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 71 / 77
 Games of imperfect information

False intuition — example 1

Possible hand 1: (player A) K♠, Q♠, A♦, A♣ : A♠, J♠, A♥, K♦ (player B).

Exemplary sequence of correct plays:

1 A: K♠, B: A♠. Remaining: (A) Q♠, A♦, A♣ : J♠, A♥, K♦ (B).
2 B: A♥, A: A♣. Remaining: (A) Q♠, A♦ : J♠, K♦ (B).
3 B: J♠, A: Q♠. Remaining: (A) A♦ : K♦ (B).
4 A: A♦, B: K♦

Both sides take 2 tricks. Score: 0 (?).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 72 / 77
 Games of imperfect information

False intuition — example 1

Possible hand 2: (A) K♠, Q♠, A♦, A♣ : A♠, J♠, A♥, K♣ (B).

Exemplary sequence of correct plays:

1 A: K♠, B: A♠. Remaining: Q♠, A♦, A♣ (A) : J♠, A♥, K♣ (B).
2 B: A♥, A: A♦. Remaining: Q♠, A♣ (A) : J♠, K♣ (B).
3 B: J♠, A: Q♠. Remaining: A♣ (A) : K♣ (B).
4 A: A♣, B: K♣

Both sides take 2 tricks. Score: 0 (?).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 73 / 77
 Games of imperfect information

False intuition — example 1

In fact, we have: (player A) K♠, Q♠, A♦, A♣ : A♠, J♠, A♥, K∗ (player B).

Problem — it is necessary to guess which ace to dispose of at the moment B

plays his A♥.
1 A: K♠, B: A♠. Remaining: Q♠, A♦, A♣ (A) : J♠, A♥, K∗ (B).
2 B: A♥, A: A♦/A♣ (?) Remaining: Q♠, A♣ (A) : J♠, K∗ (B).

1 1
Score: 2 · (−1) + 2 · 0 = − 21 .

Remark: Other first plays by A: A♦ i A♣ are not associated with positive

variation while also leading to the payoff 0.

Przemysław Klesk
˛ (KMSIiMS, ZUT) 74 / 77
 Games of imperfect information

False intuition — example 2

?
♠ KQJxx ♠ Axxxx
N
♥ AKQ ♥ xxx
W E
♦ AJ10 S ♦ Kxx
♣ AQ ♣ xx
?

Contract: 6 spades played by WE. First lead N: x♠. Key missing cards: Q♦,
K♣. Does there exist a play guaranteeing 12 tricks?

Przemysław Klesk
˛ (KMSIiMS, ZUT) 75 / 77
 Games of imperfect information

False intuition — example 2 — solution

?
♠ xx ♠ xx
N
♥ ♥
W E
♦ AJ10 S ♦ Kxx
♣ AQ ♣ xx
?

Optimal play: We trump out opponents — spades (by playing three rounds of spades if need
be), we play three rounds of hearts. An endgame as above shall take place. Now, we play A♣ and
Q♣ and we give up Q♣ voluntarily! Regardles of continuation by N or S, 12 tricks is guaranteed.

An other play based on attempts to catch K♣ at S (odds: ≈ 50%) and to catch Q♦ at S or N (odds:
≈ 50%) leads to the following expected number of tricks: ≈ 14 11 + 42 12 + 41 13 = 12, but with
variance: ≈ 14 (12 − 11)2 + 42 (12 − 12)2 + 41 (12 − 13)2 = 12 .

Przemysław Klesk
˛ (KMSIiMS, ZUT) 76 / 77
 References

Some references

1 J. von Neuman and O. Morgenstern, Theory of Games and Economic Behavior, 1944 (see:
http://press.princeton.edu/titles/7802.html).
2 Chinook project wesite: http://webdocs.cs.ualberta.ca/~chinook.
3 D.E. Knuth, R.W. Moore, An Analysis of Alpha-Beta Pruning, Artificial Intelligence, 1975 (patrz:
http://www.eecis.udel.edu/~ypeng/articles/An Analysis of Alpha-Beta Pruning.pdf).
4 A. Reinefeld, „An Improvement to the Scout Tree Search Algorithm”, ICCA Journal, 1983 (see:
http://www.top-5000.nl/ps/An improvement to the scout tree search algorithm.pdf)
5 D. Larameé, Chess Programming I-V, 2000. http://www.gamedev.net/page/resources/_/reference/programming/
artificial-intelligence/gaming/chess-programming-part-i-getting-started-r1014)
6 M. Bożykowski, Implementation of self-teaching program for checkers, master thesis, WI ZUT, 2009.
7 K. Kubasik, Application of game tree searching algorithms for finding minimax points in “double dummy” bridge, bachelor
thesis, WI ZUT, 2011.
8 P. Beling, Practical aspects of logical games programming, master thesis, Polytechnics of Łódź, 2006.
9 Expectiminimax tree, Wikipedia, (see: http://en.wikipedia.org/wiki/Expectiminimax_tree).
10 Materials on GIB — bridge playing program, (see:
http://www.greatbridgelinks.com/gblSOFT/Reviews/SoftwareReview090301.html and
http://www.gibware.com/).

Przemysław Klesk
˛ (KMSIiMS, ZUT) 77 / 77