Minimax Search

Technical Game Development II

Professor Charles Rich
Computer Science Department
[email protected]

Reference: Millington, Section 8.2

IMGD 4000 (D 10)

Minimax Search
Minimax is at the heart of almost every
computer board game
Applies to games where:
Players take turns
Have perfect information
Chess, Checkers, Tactics

But can work for games without perfect

information or with chance
Poker, Monopoly, Dice

Can work in real-time (i.e., not turn based) with

timer (iterative deepening, later)
IMGD 4000 (D 10)

The Game Tree

e.g,. Tic-Tac-Toe

Note: -just showing top part of tree

-symmetrical positions removed (optimization example)
3

IMGD 4000 (D 10)

The Game Tree

Level 0 (First Player)
Level 1 (Second Player)
Level 2 (First Player)

Nodes in tree represent states

e.g., board configurations, positions

Arcs are decisions that take you to a next state

e.g., moves

Technically a directed acyclic graph

may have joins but no cycles

Levels called plies (plural of ply)

players alternate levels (or rotate among >2 players)
IMGD 4000 (D 10)

Naive Approach
1. Exhaustively expand tree

naive because tree may be too big

e.g., chess
typical board position has ~35 legal moves
for 40 move game, 3540 > number atoms in universe

2. Choose next move on a path that leads to

your winning

assumes your opponent is going to cooperate

and let you win
on his turn, he most likely will choose the worst
case for you!
IMGD 4000 (D 10)

Minimax Approach
assume both/all players play to the best of their
ability
define a scoring method (see next)
from the standpoint of a given player (lets call
him Max):
choose move which takes you to the next state with
highest expected score (from your point of view)
assuming the other player (lets call her Min-nie) will
on her move choose the next state with the lowest
score (from your point of view)
IMGD 4000 (D 10)

(Static) Evaluation Function

assigns score to given state from point of
view of given player
scores typically integers in centered range
e.g., [-100,+100] for TTT
e.g., [-1000,+1000] for chess

extreme values reserved for win/lose

this is typically the easy case to evaluate
e.g., for first player in TTT, return +100 if board has three
Xs in a row or -100 if three Os in a row
e.g., checkmate for chess

what about non-terminal states?

IMGD 4000 (D 10)

(Static) Evaluation Function

much harder to score in middle of the game
score should reflect likelihood a player will win from
given state (board position)
but balance of winning/losing isnt always clear (e.g.,
number/value of pieces, etc.)
e.g., in Reversi, best strategy is to have fewest counters in
middle of game (better board control)
generic local maxima problem with all hill climbing search
methods

static evaluation function is where (most) gamespecific knowledge resides

IMGD 4000 (D 10)

Naive Approach
1. Apply static evaluation to each next state
2. Choose move to highest scoring state
If static evaluation function were perfect, then
this is all you need to do

perfect static evaluator almost never exists

using this approach with imperfect evaluator
performs very badly

The solution?

Look ahead!
9

IMGD 4000 (D 10)

Minimax Looking Ahead

5

Max
Min

3 9

5 6 7

Max

Its Maxs turn at the start of the game (root of the tree)
There is only time to expand tree to 2nd ply
Maxs static evaluation function has been applied to all leaf states
Max would like to get to the 9 point state
But if Max chooses leftmost branch, Min will choose her move to get to 3
=> left branch has a value of 3
If Max chooses rightmost branch, Min can choose any one of 5, 6 or 7
(will choose 5, the minimum)
=> right branch has a value of 5
Right branch is largest (the maximum) so choose that move

IMGD 4000 (D 10)

10

Minimax Bubbling Up Values

Maxs turn (root of tree)

Circles represent Maxs turn, Squares represent Mins turn
Values in leaves are result of applying static evaluation function
Red arrows represent best (local) move for each player
Blue arrow is Maxs chosen move on this turn
IMGD 4000 (D 10)

11

Minimax Algorithm
def MinMax (board, player, depth, maxDepth)
if ( board.isGameOver() or depth == maxDepth )
return board.evaluate(player), null
bestMove = null
if ( board.currentPlayer() == player )
bestScore = -INFINITYNote: makeMove returns copy of board
else bestScore = +INFINITY
(can also move/unmove--but dont execute graphics!)
for move in board.getMoves()
newBoard = board.makeMove(move)
score = MinMax(newBoard, player, depth+1, maxDepth)
if ( board.currentPlayer() == player )
if ( score > bestScore ) # max
bestScore = score
bestMove = move
Note: test works for multiplayer
else
case also
if ( score < bestScore ) # min
bestScore = score
bestMove = move
return bestScore, bestMove
MinMax(board, player, 0, maxDepth)

IMGD 4000 (D 10)

12

Negamax Version
for common case of
two player
zero sum

single static evaluation function

returns + or - same value for given board position,
depending on player

IMGD 4000 (D 10)

13

Negamax Algorithm
def NegaMax (board, depth, maxDepth)
if ( board.isGameOver() or depth == maxDepth )
return board.evaluate(), null
bestMove = null
bestScore = -INFINITY
for move in board.getMoves()
newBoard = board.makeMove(move)
score = NegaMax(newBoard, depth+1, maxDepth)
score = -score
# alternates players
if ( score > bestScore )
bestScore = score
bestMove = move
return bestScore, bestMove
NegaMax(board, 0, maxDepth)

IMGD 4000 (D 10)

14

Pruning Approach
Minimax searches entire tree, even if in some cases it is clear that
parts of the tree can be ignored (pruned)
Example:

You won a bet with your enemy.

He owes you one thing from a collection of bags.
You get to choose the bag, but your enemy chooses the thing.
Go through the bags one item at a time.
First bag: Red Sox tickets, sandwich, $20

Hell choose sandwich

Second bag: Dead fish,
Hell choose fish.
Doesnt matter what the rest of the items in this bag are ($500, Yankees
tickets )
No point in looking further in this bag, since enemys dead fish is already
worse than sandwich

IMGD 4000 (D 10)

15

Pruning Approach
In general,
Stop processing branches at a node when you find a
branch worse than result you already know you can
achieve
This type of pruning saves processing time without

affecting final result

i.e., not a heuristic like the evaluation function in A*

IMGD 4000 (D 10)

16

Pruning Example

From Maxs point of view, 1 is already lower than 5, which he

knows he can achieve, so there is no need to look farther at
sibling branches
Note that there might be large subtrees below nodes labeled
2 and 3 (only showing the top part of tree)

IMGD 4000 (D 10)

17

Alpha-Beta Pruning
Keep track of two scores:
Alpha best score by any 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

As recursion progresses, the "window" of Alpha-Beta

becomes smaller
(Beta < Alpha) => current position not result of best play and
can be pruned

IMGD 4000 (D 10)

18

Alpha-Beta NegaMax Algorithm

def ABNegaMax (board, depth, maxDepth, alpha, beta)
if ( board.isGameOver() or depth == maxDepth )
return board.evaluate(), null
bestMove = null
bestScore = -INFINITY
for move in board.getMoves()
newBoard = board.makeMove(move)
score = ABNegaMax(newBoard, maxDepth, depth+1,
-beta,
-max(alpha, bestScore))
score = -score
if ( score > bestScore )
bestScore = score
bestMove = move
# early loop exit (pruning)
if ( bestScore >= beta ) return bestScore, bestMove
return bestScore, bestMove
ABNegaMax(board, player, maxDepth, 0, -INFINITY, INFINITY)

IMGD 4000 (D 10)

19

Move Order

Benefits of pruning depend heavily on order in which

branches (moves) are visited
for example, if branches visited right to left above no pruning
happens!
for chess, on average, pruning reduces 35 branches -> 6
allows search twice as deep!

IMGD 4000 (D 10)

20

10

Move Order
Can we improve branch (move) order?
apply static evaluation function at intermediate
nodes and check best first
logical idea
can improve pruning
but may effectively give up depth of search advantage (in
fixed time interval) due to high cost of function evalution

better idea: use results of previous minimax

searches
negascout algorithm (extra credit, see Millington 8.2.7)

IMGD 4000 (D 10)

21

Chess Notes
Static evaluation function
typically use weighted function
c1*material + c2*mobility + c3*kingSafety + ...

simplest is point value for material

pawn 1, knight 3, bishop 3, castle 3, queen 9

see references in homework instructions

checkmate is worth more than rest combined
what about a draw (stalemate)?
can be good (e.g., if opponent strong)
can be bad (e.g., if opponent weak)
adjust with contempt factor (above or below zero)
IMGD 4000 (D 10)

22

11

Chess Notes
Chess has many forced tactical situations
e.g., exchanges of pieces
minimax may not find these
add cheap test at start of turn to check for
immediate captures

Library of openings and/or closings

Use iterative deepening
search 1-ply deep, check time, search 2nd ply,
check time, etc.

IMGD 4000 (D 10)

23

12