Adversarial Search and Game-Playing .ppt

C H A P T E R 5
C M P T 3 1 0 : S u m m e r 2 0 1 1
O l i v e r S c h u l t e
Adversarial Search and Game-
Playing

Environment Type Discussed In this Lecture
 Turn-taking: Semi-dynamic
 Deterministic and non-deterministic
CMPT 310 - Blind Search
2
Fully
Observable
Multi-agent
Sequential
yes
yes
Discrete
Discrete
yes
Game
Tree
Search
yes
no
Continuous Action Games
Game Matrices
no
yes

Adversarial Search
 Examine the problems that arise when we try to
plan ahead in a world where other agents are
planning against us.
 A good example is in board games.
 Adversarial games, while much studied in AI, are a
small part of game theory in economics.

Typical AI assumptions
 Two agents whose actions alternate
 Utility values for each agent are the opposite of the
other
 creates the adversarial situation
 Fully observable environments
 In game theory terms: Zero-sum games of perfect
information.
 We’ll relax these assumptions later.

Search versus Games
 Search – no adversary
 Solution is (heuristic) method for finding goal
 Heuristic techniques can find optimal solution
 Evaluation function: estimate of cost from start to goal through given node
 Examples: path planning, scheduling activities
 Games – adversary
 Solution is strategy (strategy specifies move for every possible opponent
reply).
 Optimality depends on opponent. Why?
 Time limits force an approximate solution
 Evaluation function: evaluate “goodness” of game position
 Examples: chess, checkers, Othello, backgammon

Types of Games
deterministic Chance moves
Perfect
information
Chess, checkers,
go, othello
Backgammon,
monopoly
Imperfect
information
(Initial Chance
Moves)
Bridge, Skat Poker, scrabble,
blackjack
• Theorem of Nobel Laureate Harsanyi: Every game with
chance moves during the game has an equivalent representation
with initial chance moves only.
• A deep result, but computationally it is more tractable to
consider chance moves as the game goes along.
• This is basically the same as the issue of full observability +
nondeterminism vs. partial observability + determinism.
• on-line
backgam
mon
• on-line
chess
• tic-tac-
toe

Game Setup
 Two players: MAX and MIN
 MAX moves first and they take turns until the game is over
 Winner gets award, loser gets penalty.
 Games as search:
 Initial state: e.g. board configuration of chess
 Successor function: list of (move,state) pairs specifying legal moves.
 Terminal test: Is the game finished?
 Utility function: Gives numerical value of terminal states. E.g. win (+1), lose
(-1) and draw (0) in tic-tac-toe or chess

Size of search trees
 b = branching factor
 d = number of moves by both players
 Search tree is O(bd)
 Chess
 b ~ 35
 D ~100
- search tree is ~ 10 154 (!!)
- completely impractical to search this
 Game-playing emphasizes being able to make optimal decisions in a finite amount of
time
 Somewhat realistic as a model of a real-world agent
 Even if games themselves are artificial

Partial Game Tree for Tic-Tac-Toe

Game tree (2-player, deterministic, turns)
How do we search this tree to find the optimal move?

Minimax strategy: Look ahead and reason backwards
 Find the optimal strategy for MAX assuming an
infallible MIN opponent
 Need to compute this all the down the tree
 Game Tree Search Demo
 Assumption: Both players play optimally!
 Given a game tree, the optimal strategy can be
determined by using the minimax value of each
node.
 Zermelo 1912.

Two-Ply Game Tree
The minimax decision
Minimax maximizes the utility for the worst-case outcome for max

Pseudocode for Minimax Algorithm
function MINIMAX-DECISION(state) returns an action
inputs: state, current state in game
vMAX-VALUE(state)
return the action in SUCCESSORS(state) with value v
function MIN-VALUE(state) returns a utility value
if TERMINAL-TEST(state) then return UTILITY(state)
v  ∞
for a,s in SUCCESSORS(state) do
v  MIN(v,MAX-VALUE(s))
return v
function MAX-VALUE(state) returns a utility value
if TERMINAL-TEST(state) then return UTILITY(state)
v  -∞
for a,s in SUCCESSORS(state) do
v  MAX(v,MIN-VALUE(s))
return v

Example of Algorithm Execution
MAX to move

Minimax Algorithm
 Complete depth-first exploration of the game tree
 Assumptions:
 Max depth = d, b legal moves at each point
 E.g., Chess: d ~ 100, b ~35
Criterion Minimax
Time O(bd)
Space O(bd)



Practical problem with minimax search
 Number of game states is exponential in the number of
moves.
 Solution: Do not examine every node
=> pruning
Remove branches that do not influence final decision
 Revisit example …

Alpha-Beta Example
[-∞, +∞]
[-∞,+∞]
Range of possible values
Do DF-search until first leaf

Alpha-Beta Example (continued)
[-∞,3]
[-∞,+∞]

[3,+∞]
[3,3]

[-∞,2]
[3,+∞]
[3,3]
This node is worse
for MAX

[-∞,2]
[3,14]
[3,3] [-∞,14]
,

[−∞,2]
[3,5]
[3,3] [-∞,5]
,

[2,2]
[−∞,2]
[3,3]
[3,3]

[2,2]
[-∞,2]
[3,3]
[3,3]

Alpha-beta Algorithm
 Depth first search – only considers nodes along a single
path at any time
a = highest-value choice that we can guarantee for MAX
so far in the current subtree.
b = lowest-value choice that we can guarantee for MIN so
far in the current subtree.
 update values of a and b during search and prunes
remaining branches as soon as the value is known to be
worse than the current a or b value for MAX or MIN.
 Alpha-beta Demo.

Effectiveness of Alpha-Beta Search
 Worst-Case
 branches are ordered so that no pruning takes place. In this case alpha-beta
gives no improvement over exhaustive search
 Best-Case
 each player’s best move is the left-most alternative (i.e., evaluated first)
 in practice, performance is closer to best rather than worst-case
 In practice often get O(b(d/2)) rather than O(bd)
 this is the same as having a branching factor of sqrt(b),
 since (sqrt(b))d = b(d/2)
 i.e., we have effectively gone from b to square root of b
 e.g., in chess go from b ~ 35 to b ~ 6
 this permits much deeper search in the same amount of time
 Typically twice as deep.

Example
3 4 1 2 7 8 5 6
-which nodes can be pruned?
MAX
MIN
MAX

Final Comments about Alpha-Beta Pruning
 Pruning does not affect final results
 Entire subtrees can be pruned.
 Good move ordering improves effectiveness of
pruning
 Repeated states are again possible.
 Store them in memory = transposition table

Practical Implementation
How do we make these ideas practical in real game trees?
Standard approach:
 cutoff test: (where do we stop descending the tree)
 depth limit
 better: iterative deepening
 cutoff only when no big changes are expected to occur next (quiescence search).
 evaluation function
 When the search is cut off, we evaluate the current state
by estimating its utility using an evaluation function.

Static (Heuristic) Evaluation Functions
 An Evaluation Function:
 estimates how good the current board configuration is for a player.
 Typically, one figures how good it is for the player, and how good it is for the
opponent, and subtracts the opponents score from the players
 Othello: Number of white pieces - Number of black pieces
 Chess: Value of all white pieces - Value of all black pieces
 Typical values from -infinity (loss) to +infinity (win) or [-1, +1].
 If the board evaluation is X for a player, it’s -X for the opponent.
 Many clever ideas about how to use the evaluation function.
 e.g. null move heuristic: let opponent move twice.
 Example:
 Evaluating chess boards,
 Checkers
 Tic-tac-toe

Iterative (Progressive) Deepening
 In real games, there is usually a time limit T on making a
move
 How do we take this into account?
 using alpha-beta we cannot use “partial” results with any confidence
unless the full breadth of the tree has been searched
 So, we could be conservative and set a conservative depth-limit
which guarantees that we will find a move in time < T
 disadvantage is that we may finish early, could do more search
 In practice, iterative deepening search (IDS) is used
 IDS runs depth-first search with an increasing depth-limit
 when the clock runs out we use the solution found at the previous
depth limit

The State of Play
 Checkers:
 Chinook ended 40-year-reign of human world champion
Marion Tinsley in 1994.
 Chess:
 Deep Blue defeated human world champion Garry Kasparov in
a six-game match in 1997.
 Othello:
 human champions refuse to compete against computers: they
are too good.
 Go:
 human champions refuse to compete against computers: they
are too bad b > 300 (!)
 See (e.g.) http://www.cs.ualberta.ca/~games/ for more information

Deep Blue
 1957: Herbert Simon
 “within 10 years a computer will beat the world chess champion”
 1997: Deep Blue beats Kasparov
 Parallel machine with 30 processors for “software” and 480
VLSI processors for “hardware search”
 Searched 126 million nodes per second on average
 Generated up to 30 billion positions per move
 Reached depth 14 routinely
 Uses iterative-deepening alpha-beta search with
transpositioning
 Can explore beyond depth-limit for interesting moves

Summary
 Game playing can be effectively modeled as a search problem
 Game trees represent alternate computer/opponent moves
 Evaluation functions estimate the quality of a given board
configuration for the Max player.
 Minimax is a procedure which chooses moves by assuming that
the opponent will always choose the move which is best for them
 Alpha-Beta is a procedure which can prune large parts of the
search tree and allow search to go deeper
 For many well-known games, computer algorithms based on
heuristic search match or out-perform human world experts.

AI Games vs. Economics Game Theory
 Seminal Work on Game Theory: Theory of Games and
Economic Behavior, 1944, by von Neumann and
Morgenstern.
 Agents can be in cooperation as well as in
conflict.
 Agents may move simultaneously/independently.

Example: The Prisoner’s Dilemma
Other Famous Matrix Games:
• Chicken
• Battle of The Sexes
• Coordination

Solving Zero-Sum Games
• Perfect Information: Use Minimax Tree Search.
• Imperfect Information: Extend Minimax Idea
with probabilistic actions.
➩ von Neumann’s Minimax Theorem: there
exists an essentially unique optimal probability
distribution for randomizing an agent’s
behaviour.

Matching Pennies
Heads Tails
Heads 1,-1 -1,1
Tails -1,1 1,-1
• Why should the players randomize?
• What are the best probabilities to use in their actions?

Nonzero Sum Game Trees
 The idea of “look ahead, reason backward” works for
any game tree with perfect information.
 I.e., also in cooperative games.
 In AI, this is called retrograde analysis.
 In game theory, it is called backward induction or
subgame perfect equilibrium.
 Can be extended to many games with imperfect
information (sequential equilibrium).

Backward Induction Example: Hume’s Farmer Problem
1
2 2
2
2
0
3
3
0
1
1
H Not H
H1
notH
1 H2
Not
H2

Summary: Solving Games
Zero-sum Non zero-sum
Perfect Information Minimax, alpha-beta Backward induction,
retrograde analysis
Imperfect Information Probabilistic minimax Nash equilibrium
Nash equilibrium is beyond the scope of this course.

Single Agent vs. 2-Players
 Every single agent problem can be considered as a
special case of a 2-player game. How?
1. Make one of the players the Environment, with a constant
utility function (e.g., always 0).
1. The Environment acts but does not care.
2. An adversarial Environment, with utility function the
negative of agent’s utility.
1. In minimization, Environment’s utility is player’s costs.
2. Worst-Case Analysis.
3. E.g., program correctness: no matter what input user gives,
program gives correct answer.
 So agent design is a subfield of game theory.

Single Agent Design = Game Theory
Von Neumann-Morgenstern Games
Decision Theory = 2-player game, 1st player the
“agent”, 2nd player “environment/nature”
(with constant or adversarial utility function)
Markov Decision Processes
Planning Problems
From
General
To
Special
Case

Example: And-Or Trees
 If an agent’s actions have nondeterministic effects,
we can model worst-case analysis as a zero-sum
game where the environment chooses the effects of
an agent’s actions.
 Minimax Search ≈ And-Or Search.
 Example: The Erratic Vacuum Cleaner.
 When applied to dirty square, vacuum cleans it and sometimes
adjacent square too.
 When applied to clean square, sometimes vacuum makes it
dirty.
 Reflex agent: same action for same location, dirt status.

And-Or Tree for the Erratic Vacuum
• The agent
“moves” at labelled
OR nodes.
• The environment
“moves” at
unlabelled AND
nodes.
•The agent wins if it
reaches a goal
state.
• The environment
“wins” if the agent
goes into a loop.

Summary
 Game Theory is a very general, highly developed
framework for multi-agent interactions.
 Deep results about equivalences of various
environment types.
 See Chapter 17 for more details.

Adversarial Search and Game-Playing .ppt

More Related Content

Similar to Adversarial Search and Game-Playing .ppt

Recently uploaded

Adversarial Search and Game-Playing .ppt