Adversarial Search
In which we examine the problems that arise when we try to plan
ahead in a world where other agents are planning against us.
Dr.shahnawaz qureshi
Adversarial Search
Multi-agent environment:
any given agent needs to consider the actions
of other agents and how they affect its own
welfare
introduce possible contingencies into the
agent’s problem-solving process
cooperative vs. competitive
Adversarial search problems: agents have
conflicting goals -- games
2
Games vs. Search Problems
"Unpredictable"
opponent
specifying a move
for every possible
opponent reply
Time limits
unlikely to find
goal, must
approximate
3
AI and Games
In AI, “games” have special
format:
deterministic, turn-taking, 2-
player, zero-sum games of
perfect information
Zero-sum describes a situation in
which a participant’s gain or loss
is exactly balanced by the losses
or gains of the other
participant(s)
Or, the total payoff to all players Go! 围棋
is the same for every instance of
the game (constant sum)
4
Game Problem Formulation
A game with 2 players (MAX and MIN, MAX moves
first, turn-taking) can be defined as a search
problem with:
initial state: board position
player: player to move
successor function: a list of legal (move, state) pairs
goal test: whether the game is over – terminal states
utility function: gives a numeric value for the terminal
states (win, loss, draw)
Game tree = initial state + legal moves
5
Game Tree (2-player, deterministic)
Utility value
6
Optimal Strategies
MAX must find a contingent strategy, specifying MAX’s move in:
the initial state
the states resulting from every possible response by MIN
E.g., 2-ply game (the tree is one move deep, consisting of two half-
moves, each of which is called a ply):
MAX 3
MIN 3 2 2
MAX
3 12 8 2 4 6 14 5 2
7
Minimax Value
Perfect play for deterministic game, assume both
players play optimally
Idea: choose move to position with highest
minimax value = best achievable payoff against
best play
MINIMAX VALUE ( n)
Utility ( n) if n is a terminal state
max s Successors ( n ) MINIMAX ( s ) if n is a MAX node
min s Successors ( n ) MINIMAX ( s ) if n is a MIN node
8
A Partial Game Tree for Tic-Tac-Toe
O’s turn (MIN)
X’s turn (MAX)
by Yosen Lin 9
Minimax Algorithm
10
Analysis of Minimax
Optimal play for MAX assumes that MIN also plays
optimally, what if MIN does not play optimally?
A complete depth-first search?
Yes
Time complexity?
O(bm)
Space complexity?
O(bm) (depth-first exploration)
For chess, b ≈ 35, m ≈ 100 for "reasonable"
games
exact solution completely infeasible
CS 420: Artificial Intelligence 12
Optimal Decisions for Multiplayer Games
Extend minimax idea to multiplayer
A (1,2,6)
B (1,2,6) (1,5,2)
C (1,2,6) (6,1,2) (1,5,2) (5,4,5)
A
(1,2,6) (4,2,3) (6,1,2) (7,4,1) (5,1,1) (1,5,2) (7,7,1) (5,4,5)
13
Interesting Thoughts
Multiplayer games usually involve alliances,
which can be a natural consequence of
optimal strategies
If the game is non zero-sum, collaboration
can also occur
For example, a terminal state with utilities <Va
= 1000, Vb = 1000>
The optimal strategy is for both players to do
everything possible to reach this state
14
α-β Pruning
The number of game states with minimax
search is exponential in the # of moves
Is it possible to compute the correct
minimax decision without looking at every
node in the game tree?
Need to prune away branches that cannot
possibly influence the final decision
15
α-β Pruning Example
[-∞,+∞]
[-∞, 3]
16
α-β Pruning Example
[-∞,+∞]
[-∞, 3]
17
α-β Pruning Example
[3,+∞]
[3, 3]
18
α-β Pruning Example
[3,+∞]
[3, 3] [-∞, 2]
19
α-β Pruning Example
[3, 14]
[3, 3] [-∞, 2] [-∞, 14]
20
α-β Pruning Example
[3, 5]
[3, 3] [-∞, 2] [-∞, 5]
21
α-β Pruning Example
[3, 3]
[3, 3] [-∞, 2] [2, 2]
22
A Simple Formula
MINIMAX VALUE (root )
max(min(3,12,8), min(2, x, y ), min(14,5, 2))
max(3, min(2, x, y ), 2)
max(3, z , 2) where z ≤ 2
3
The value of the root and hence the minimax decision
are independent of the values of pruned leaves x and y
23
Basic Idea of α-β
Consider a node n such that
Player has a choice of
moving to
If Player has a better choice
m either at the parent of n
or at any choice point
further up, then n will never
be reached in actual play
α-β pruning gets it name
from the two parameters
that describe bounds on the
backed-up values
25
Definitions of α and β
α = the value of the best (highest-value) choice
we have found so far at any choice point along the
path for MAX
β = the value of the best (lowest-value) choice we
have found so far at any choice point along the
path for MIN
α-β search updates the values of α and β as it
goes along and prunes the remaining branches at
a node as soon as the value of the current node is
worse than the current α or β for MAX or MIN
respectively
26
The α-β Algorithm
CS 420: Artificial Intelligence 27
Analysis of α-β
Pruning does not affect final result
The effectiveness of alpha-beta pruning is highly dependent
on the order of successors
It might be worthwhile to try to examine first the successors
that are likely to be best
With "perfect ordering," time complexity = O(bm/2)
effective branching factor becomes b
For chess, 6 instead of 35
it can look ahead roughly twice as far as minimax in the same
amount of time
Ordering in chess: captures, threats, forward moves, and then
backward moves
29
Random Ordering?
If successors are examined in random order rather
than best-first, the complexity will be roughly
O(b3m/4)
Adding dynamic move-ordering schemes, such as
trying first the moves that were found to be best
last time, brings us close to the theoretical limit
The best moves are often called killer moves
(killer move heuristic)
30
Dealing with Repeated States
In games, repeated states occur frequently
because of transpositions --
different permutations of the move sequence end up in
the same position
e.g., [a1, b1, a2, b2] vs. [a1, b2, a2, b1]
It’s worthwhile to store the evaluation of this
position in a hash table the first time it is
encountered
similar to the “explored set” in graph-search
Tradeoff:
Transposition table can be too big
Which to keep and which to discard
31
Imperfect, Real-Time Decisions
Minimax generates the entire game search space
Alpha-beta prunes large part of it, but still needs
to search all the way to terminal states
However, moves must be made in reasonable
amount of time
Standard approach: turning non-terminal nodes
into terminal leaves
cutoff test: replaces terminal test, e.g., depth limit
heuristic evaluation function = estimated desirability or
utility of position
32
Evaluation Functions
The performance of a game-playing
program is dependent on the quality of its
evaluation function
Order the terminal states the same way as the
true utility function
Evaluation of nonterminal states correlate with
the actual chance of winning
Computation must not take too long!
33
Playing Chess, Experience is
Important
Most eval. Functions work by calculating various
features of the state
Different features form various categories of
states
Consider the category of two-pawns vs. one pawn
and suppose experience suggests that 72% of the
states lead to a win (+1), 20% to a loss(0), and
8% to a draw (1/2)
Expected value = 0.72*1 + 0.20*0 + 0.08*1/2 = 0.76
Too many categories and too much experience are
required
34
Deterministic Games in Practice
Checkers: Chinook ended 40-year-reign of human world champion
Marion Tinsley in 1994. Used a precomputed endgame database
defining perfect play for all positions involving 8 or fewer pieces on
the board, a total of 444 billion positions.
Chess: Deep Blue defeated human world champion Garry Kasparov
in a six-game match in 1997. Deep Blue searches 200 million
positions per second, uses very sophisticated evaluation, and
undisclosed methods for extending some lines of search up to 40
ply.
Othello: human champions refuse to compete against computers,
who are too good.
Go: human champions refuse to compete against computers, who
are too bad. In go, b > 300, so most programs use pattern
knowledge bases to suggest plausible moves.
38
Summary
Games are fun to work on!
They illustrate several important points
about AI
Perfection is unattainable must
approximate
39
Exercise 1
Prove that: for every game tree, the utility
obtained by MAX using minimax decisions
against a suboptimal MIN will never be
lower than the utility obtained playing
against an optimal MIN.
40
Exercise 2
Consider the following game:
Draw the complete game tree, using the following convention:
State: (Sa, Sb) where Sa and Sb denote the token locations
Identify the terminal state in a square box and assign it a value
Put loop states in double square boxes
Since it’s not clear how to assign values to loop states, annotate
each with a “?”
A B
1 2 3 4
41
Answer to Ex1
Consider a MIN node whose children are
terminal nodes. If MIN plays suboptimally,
then the value of the node is greater than
or equal to the value it would have if MIN
played optimally. Hence, the value of the
MAX node that is the MIN node’s parent
can only be increased. This argument can
be extended by a simple induction all the
way to the root.
42
Game Tree for Ex2
Think about how to get
the values
43
