Вы находитесь на странице: 1из 35

A Comprehensive

Overview of
Minimax Algorithm
A PRESENTATION BY ASHISH SABADE

Minimax Algorithm An Introduction


Minimax Algorithm is used in Artificial Intelligence in computer games
Its at the heart of almost every computer board game
Minimax chooses the path which maximizes the gain of the current player, while minimizing the
gain of the adversary
A search tree is generated, depth-rst, starting with the current game position upto the end
game position.
Can work in real time(i.e- not turn based) with timer (iterative deepening, later)

Where is the algorithm used ?


Applies to games where:
Players take turns
Have perfect information
Chess, Checkers, Tactics
Most widely used in logic games. These games are also called as zero sum games i.e a game in
which gain of players move is exactly balanced by loss incurred by opponents move

Used in economic situations


Used in social situations

Minimax Algorithm-Components
Search tree :
Squares represent decision states (i.e- after a move)
Branches are decisions (i.e- the move)
Start at root
Nodes at end are leaf nodes
Unlike binary trees can have any number of children
Depends on the game situation
Levels usually called plies (a ply is one level)
Each ply is where "turn" switches to other player
Players called Min and Max

Current board
position

Minimax Algorithm-Description(1)
Assign points to the outcome of a game
Ex: Tic-Tac-Toe: X wins, value of 1. O wins, value -1.

Max (X) tries to maximize point value, while Min (O) tries to minimize point
value
Assume both players play to best of their ability
Always make a move to minimize or maximize points

So, in choosing, Max will choose best move to get highest points, assuming Min
will choose best move to get lowest points

Minimax Algorithm-Description(2)
A tree is generated and filled with values generated from evaluation function
An evaluation function, also known as a heuristic evaluation function or static evaluation
function, is a function used by game-playing programs to estimate the value or goodness of a
position
The evaluation function is typically designed to prioritize speed over accuracy; the function looks
only at the current position and does not explore possible moves (therefore static).
E.g for chess: f(P) = 200(K-K') + 9(Q-Q') + 5(R-R') + 3(B-B'+N-N') + (P-P') - 0.5(D-D'+S-S'+I-I')
+ 0.1(M-M')
K, Q, R, B, N, P are the number of white kings, queens, rooks, bishops, knights and pawns on the board.
D, S, I are doubled, backward and isolated white pawns.
M represents white mobility (measured, say, as the number of legal moves available to White)

How is Minimax tree generated ?


Step 1:Creation of a MAX node
Step 2:Generation of children nodes (MIN) using DFS

15

MAX

Step 3:Node A1 is generated.


Step 4:Children of a A1 are generated

MIN

A1

A2 15

15

A3

Step 5:Node evaluation takes place using


evaluation function.
Values are filled in nodes
12

20

Step 6:Now A1 chooses the minimum of the (two)


child node values
Step 7:Repeat steps 3 to 6 till all children of main
nodes (A1,A2,A3) are generated and filled

Step 8:The Main mode chooses the max value


of the child node values and this tells us the
best possible move that we should take

Minimax Algorithm-Pseudo Code


Following functions are used in minimax algorithm :
Main function to call either max or min function depending upon player
int MiniMax (int depth)
Max Function returns move of max value
int Max (int depth)
Min Function returns min value
int Min (int depth)
Evaluation function returns goodness of position of player
int eval(move)

Minimax Algorithm-Main Function


int MinMax(int depth){
// White is Max, Black is Min

if (turn == WHITE)
return Max(depth);
else
return Min(depth);
}
Function call:value = MinMax(5); // search 5 plies

Minimax Algorithm-Max Function


int Max(int depth) {
int best = -INFINITY; // first move is best

if (depth == 0)
return Evaluate();
GenerateLegalMoves();
while (MovesLeft()) {
MakeNextMove();
val = Min(depth 1); // Mins turn next
UnMakeMove();
if (val > best)
best = val;

}
return best;
}

Minimax Algorithm-Min Function


int Min(int depth) {
int best = INFINITY; // different than MAX

if (depth == 0)
return Evaluate();
GenerateLegalMoves();
while (MovesLeft()) {
MakeNextMove();
val = Max(depth 1); // Maxs turn next
UnMakeMove();
if (val < best) // different than MAX
best = val;

}
return best;
}

Minimax Algorithm-Evaluation
Function
Most Important Function in minimax algorithm
Example for tic tac toe for max player (for min player just return score *(-1) )
int eval(move){
if(game over) return 0;
else{
if(move blocking opponents win) return highest_priority; // say highest priority is 5
if(open move){
if(move in central area) return 4;
if(move in corner area) return 3;
if(move in horizontal block) return 1;
}
}
}

Properties Of Minimax Algorithm


Complete?
Yes (if tree is finite)

Optimal?
Yes (against an optimal opponent).
Can it be beaten by an opponent playing sub-optimally?
No

Time complexity?

O(bm) ,where m is depth and b is branching factor

Space complexity?
O(bm) (depth-first search, generate all actions at once)
O(m) (backtracking search, generate actions one at a time)

Alpha Beta Pruning and


Search techniques

Minimax Algorithm-Alpha Beta


Pruning(1)
MinMax searches entire tree, even if in some cases the rest can be ignored
e.g->In chess for depth of 5, 364 nodes are generated !!!

In general, stop evaluating move when find worse than previously examined move
Does not benefit the player to play that move, it need not be evaluated any further.
Save processing time without affecting final result
Hence, Alpha-Beta pruning is used to dismiss unnecessary generation of worthless nodes
We can use alpha-beta pruning to calculate values at more depth and quickly, thus enabling
computer to make best decision using less time

Minimax Algorithm-Alpha Beta


Pruning(2)
Two scores passed around in search
Alpha best score by some means
Anything less than this is no use (can be pruned) since we can already get alpha
Minimum score Max will get
Initially, negative infinity

Beta worst-case scenario for opponent


Anything higher than this wont be used by opponent
Maximum score Min will get
Initially, infinity

Recursion progresses, the "window" of Alpha-Beta becomes smaller

Beta < Alpha current position not result of best play and can be pruned

Alpha Beta-Psuedocode of Alpha


function

Alpha Beta techniques-Quiescence


Search
It is a remedy for the horizon problem faced by AI engines for various games
like chess and abalone
Emulates human behaviour by instructing a computer to search "interesting" positions to a
greater depth than "quiet" ones to make sure there are no hidden traps.
Pseudocode:
function quiescence_search(node, depth)
if node appears quiet or node is a terminal node or depth = 0
return estimated value of node else //One might use minimax or alpha-beta search here...
search children of node using recursive applications of quiescence_search
return estimated value of children //...and here
function normal_search(node, depth)
if node is a terminal node
return estimated value of node
else if depth = 0
if node appears quiet
return estimated value of node
else
return estimated value from quiescence_search(node, reasonable_depth_value)
else search children of node using recursive applications of normal_search
return estimated value of children

Alpha Beta techniques-Non-Linear


Aspiration Window
Used in order to decrease the Search Space in an Alpha-Beta Pruning algorithm
This technique can be generalized by setting a search window for each level of the tree.

The upper and lower limit of that window are defined as the highest and the lowest reasonable
value that a node at this level can have
If a nodes value falls outside that window, depending on the father type (max or min), we can
interrupt our search and return that same value for this node.

The window size should be different for every level of the tree
For higher levels->bigger window
For lower levels->smaller window

Alpha Beta techniques-Transposition


Tables
A common method that is applied when a problem consists of overlapping sub-problems is
Memoization(making a memo)
This method involves storing the result of solved sub-problems in order to use it, in the future,
without the need of making the same calculations all over again
A hash table is created accordingly. Zorbit Hashing technique is used
A unique number needs to be assigned to each different state of the game board

Transposition Tables have great impact on the performance of the algorithm.


With the use of a table with 4MegaEntries the search becomes approximately three times
faster.

Alpha Beta techniques-Iterative


Deepening
It is better to complete the search at a shallower, thus safer depth

and then steadily increase the depth, while checking the remaining time
Just before the time limit is reached, the search for the best possible move at the increased
depth is aborted and the move from the previously completed depth search is retrieved.
By using a Transposition Table, many of the evaluation functions values of the shallower search
are stored and can be retrieved while searching deeper.

Combined Move Ordering (CMO)


It is true that the more well sorted the children of a node are, the more efficient the Alpha-Beta
pruning becomes
If they are sorted by estimating which move is expected to be better, then the probability of
earlier pruning is increased.
The gain from sorting the child-nodes is more significant at higher levels of the tree, because
pruning of deep sub-trees can happen earlier.

Results of CMO against others


Y axis:number of nodes traversed
X-axis:Depth

Board game of
ABALONE
was used for tests

Randomized Minimax Algorithm


(Rminimax)

Need of Random Minimax Algorithm


Traditional minimax algorithm is very useful but it has some drawbacks
Drawbacks:1) Too deterministic (predictable)
2) Does not take into account the human error factor i.e. it assumes a player will always be perfectly
rational
3) Controlling strength of a player is not possible (only one level of difficulty achievable)

A random minimax will randomize the choice among all possible paths of the game tree is
introduced

Random Minimax Technique RSPF


Rminimax relies heavily on the randomized shortest-path model
Let G be a directed graph containing a source node with index 1

and a destination or goal node with index n


A non-negative local cost ckk >= 0 is associated to each of the arcs
The set of all paths (including cycles) that go from 1 to n is denoted as P1n
P1n is composed by a sequence of arcs k->k
Let the total cost C() of a path be the sum of these local costs along
A probability will be assigned to each path, favouring nearly optimal
paths having a low cost

-> parameter which


controls entropy

When ->infinity, entropy is 0


When ->0, almost equal
probability is
assigned to all
paths

Rminimax Algorithm
Require:
G: The generated game tree obtained with the MINIMAX
algorithm. The root of the game is k 1.
> 0: The degree of randomization of the tree ( for a
perfect rational player; 0 for an almost completely random
player).
ckk 0: The cost of each arc of the tree.
1. Assign zbn= 1 for each n N.
2. Recursively compute zbk .
3. Compute the corresponding pkk value.
4. return pkk: the transition probabilities for the next play.

1 AI (CPU) i.e MAX


2 Player i.e MIN

max

12

Normal Minimax
( CPU with max move)
9

min

12

Perfect move calculated every time

18

10

12

30

Human error not taken into


account
No randomness

*Numbers illustrate evaluation score of the game


board i.e. game state

CPU move

10

Random Minimax
( CPU with random move)
Wrong
move !!

10

12

Correct
move !!

18

10

12

30

*Numbers illustrate evaluation score of the game


board i.e. game state

Case 1: ->infinity
optimal minimax strategy
(previous example)
Case 2: ->0
random strategy
(current example)

Note that CPU does not play optimal


move because of Rminimax

(Probable) Applications of Rminimax


The research work on random minimax algorithm is still in process
However it is gaining more and more popularity in game programming because of its non
deterministic nature
It is expected that soon various multiplayer games and online games (having artificial
intelligence) will implement Rminimax algorithm
Also further research is expected to make the AI behaviour compatible to the players behaviour
by modifying
Also this framework can be applied to Monte-Carlo search technique

Resulting curves for the


Rminimax algorithm applied to
the game Tic-Tac-Toe.
The horizontal axis represents
the variation of 1 for player 1
while the vertical axis shows the
number of victories of 1 over
2,out of 100 games.
when 1 >> 2, player 1 wins
while for 1 << 2 , it is player 2
who leads the game

References
1)Rminimax: An Optimally Randomized MINIMAX Algorithm
IEEE TRANSACTIONS ON CYBERNETICS, VOL. 43, NO. 1, FEBRUARY 2013

2) Exploring Optimization Strategies in Board Game Abalone for Alpha-Beta Search


Computational Intelligence and Games (CIG), 2012 IEEE Conference

3)Research on techniques of computer games


Control and Decision Conference (CCDC), 2013

THANK YOU

Вам также может понравиться