Local & Adversarial Search

CSD 15-780: Graduate Artificial Intelligence

1
 Local search algorithms
 Sometimes the path to the goal is irrelevant:
 8-queens problem, job-shop scheduling
 circuit design, computer configuration
 automatic programming, automatic graph drawing
 Optimization problems may have no obvious
“goal test” or “path cost”.
 Local search algorithms can solve such
problems by keeping in memory just one
current state (or perhaps a few).

2
Advantages of local search
1. Very simple to implement.
2. Very little memory is needed.
3. Can often find reasonable solutions in
very large state spaces for which
systematic algorithms are not suitable.

3
Hill-climbing search

4
Problems with hill-climbing
 Can get stuck at a local maximum.
 Cannot climb along a narrow ridge when each
possible step goes down.
 Unable to find its way off a plateau.
Solutions:
 Stochastic hill-climbing – select using weighted

random choice
 First-Choice hill-climbing – randomly generate

neighbors until one better

different initial states.

6
Simulated annealing search

7
Beam Search
 Like hill-climbing but instead of tracking just
one best state, it tracks k best states.
 Start with k states and generate successors
 If solution in successors, return it.
 Otherwise, select k best states selected from
all successors.
 Like hill-climbing, there are stochastic forms
of beam search.

8
Genetic Algorithms
 Similar to stochastic beam search,
except that successors are drawn from
two parents instead of one.
 General idea is to find a solution by
iteratively selecting fittest individuals
from a population and breeding them
until either a threshold on iterations or
fitness is hit.
9
Genetic algorithms cont.
 An individual state is represented by a
sequence of “genes”.
 The selection strategy is randomized
with probability of selection
proportional to “fitness”.
 Individuals selected for reproduction
are randomly paired, certain genes are
crossed-over, and some are mutated.
10
Genetic algorithms cont.

11
Genetic Algorithm

12
Genetic algorithms cont.
 Genetic algorithms have been applied to a
wide range of problems.
 Results are sometimes very good and
sometimes very poor.
 The technique is relatively easy to apply and
in many cases it is beneficial to see if it
works before thinking about another
approach.

13
Adversarial Search
 The minimax algorithm
 Alpha-Beta pruning
 Games with chance nodes
 Games versus real-world competitive
situations

14
Adversarial Search
 An AI favorite
 Competitive multi-agent environments
modeled as games

15
From single-agent to two-players
 Actions no longer have predictable
outcomes
 Uncertainty regarding opponent and/or
outcome of actions
 Competitive situation
 Much larger state-space
 Time limits
 Still assume perfect information
16
Formalizing the search problem
 Initial state = initial game/board position
and player
 Successors = operators = all legal moves
 Terminal state test (not “goal”-test) = a
state in which the game ends
 Utility function = payoff function = reward
 Game tree = a graph representing all the
possible game scenarios

17
Partial game tree for Tic-Tac-Toe

18
What are we searching for?
 Construct a “strategy” or “contingent
plan” rather than a “path”
 Must take into account all possible
moves by the opponent
 Representation of a strategy
 Optimal strategy = leads to the highest
possible guaranteed payoff

19
The minimax algorithm
 Generate the whole tree
 Label the terminal states with the payoff
function
 Work backwards from the leaves,
labeling each state with the best
outcome possible for that player
 Construct a strategy by selecting the
the best moves for “Max”

20
Minimax algorithm cont.
 Labeling process leads to the “minimax
decision” that guarantees maximum
payoff, assuming that the opponent is
rational
 Labeling can be implemented using
depth-first search using linear space

21
Illustration of minimax
MAX 3

MIN
3 2 2

3 12 8 2 4 6 14 5 2

22
But seriously...
 Can’t search all the way to leaves
 Use Cutoff-Test function;
generate a partial tree whose leaves
meet the cutoff-test
 Apply heuristic to each leaf
 Assume that the heuristic represents
payoffs, and back up using minimax

23
What’s in an evaluation function?
 Evaluation function assigns each state
to a category, and imposes an ordering
on the categories
 Some claim that the evaluation function
should measure P(winning)...

24
Evaluating states in
chess
 “material” evaluation
 Count the pieces for each side, giving
each a weight (queen=9, rook=5,
knight/bishop=3, pawn=1)
 What properties do we care about in the
evaluation function?
 Only the ordering matters

25
 Evaluating states in
backgammon
Possible goals (features):
 Hit your opponent's blots
 Reduce the number of blots that are in danger
 Build points to block your opponent
 Remove men from board
 Get out of opponent's home
 Don't build high points
 Spread the men at home positions

26
Learning evaluation functions
 Learning the weights of chess pieces...
can use anything from linear regression
to hill-climbing.
 The harder question is picking the
primitive features to use.

27
Problems with minimax
 Uniform depth limit
 Horizon problem:
over-rates sequences of moves that
“stall” some bad outcome
 Does not take into account possible
“deviations” from guaranteed value
 Does not factor search cost into the
process

28
 Minimax may be inappropriate…
MAX

MIN
99 100

99 1000 1000 1000 100 101 102 100

29
Reducing search cost
 In chess, can only search
full-width tree to about 4 levels
 The trick is to “prune” certain subtrees
 Fortunately, best move is provably
insensitive to certain subtrees

30
Alpha-Beta pruning
 Goal: compute the minimax value of a
game tree with minimal exploration.
 Along current search path, record best
choice for Max (alpha), and best choice
for Min (beta).
 If any new state is known to be worse
than alpha or beta, it can be pruned.
 Simple example of “meta-reasoning”

31
 Illustration of Alpha-Beta
MAX
11

11 10 MIN

11 48 10 MAX

11 9 48 10 10 MIN

X X X X X X X X
41 11 9 37 52 48 20 30 10 27 10 37 50 36 25 3

32
Implementation of Alpha-Beta
function Alpha (state, , )
if Cutoff (state) then return Value(state)
for each s in Successors(state) do
  Max(, Beta (s, , ))
if then return 
end
return 

33
Implementation cont.
function Beta (state, , )
if Cutoff (state) then return Value(state)
for each s in Successors(state) do
  Min(, Alpha (s, , ))
if then return 
end
return 

34
Effectiveness of Alpha-Beta
 Depends on ordering of successors.
 With perfect ordering, can search twice
as deep in a given amount of time (i.e.,
effective branching factor is SQRT(b)).
 While perfect ordering cannot be
achieved, simple heuristics are very
effective.

35
What about time limits?
 Iterative deepening
(minimax to depths 1, 2, 3, ...)
 Can even use iterative deepening
results to improve top-level ordering

36
Games with an element of chance
 Add chance nodes to the game tree
 Use the expecti-max or expecti-minimax
algorithm
 One problem: evaluation function is now
scale dependent (not just ordering!)
 There is even an alpha-beta trick for this
case

37
38
Evaluation is scale dependent

39
State-of-the-art programs
Chess: Deep Blue [Campbell, Hsu, and Tan; 1997]
 Defeated Gary Kasparov in a 6-game match.
 Used parallel computer with 32 PowerPCs
and 512 custom VLSI chess processors.
 Could search 100 bilion positions per move,
reaching depth 14.
 Used alpha-beta with improvements,
following “interesting” lines more deeply.
 Extensive use of libraries of openings and
endgames.
40
 State-of-the-art programs
 Checkers: [Samuel, 1952]
 Expert-level performance using a 1KHz CPU with
10,000 words of memory.
 One of the early example of machine learning.
 Checkers: Chinook [Schaeffer, 1992]
 Won the 1992 U.S. Open and first to challenge for a
world championship.
 Lost in match against Tinsley (World champion for over
40 years who had lost only in 3 games before match).
 Became world champion in 1994.
 Used alpha-beta search combined with a database of
all 444 bilion positions with 8 pieces or less on board.41
State-of-the-art programs
Backgammon: TD-Gammon [Tesauro, 1992]
 Ranked among the top three players in the
world.
 Combined Samuel’s RL method with neural
network techniques to develop a remarkably
good heuristic evaluator.
 Used expecti-minimax search to depth 2 or 3.

42
State-of-the-art programs
Bridge: GIB [Ginsburg, 1999]
 Won computer bridge championship; finished 12th in
a field of 35 at the 1998 world championship.
 Examine how each choice works for a random
sample of the up to 10 million possible arrangements
of the hidden cards.
 Used explanation-based generalization to compute
and cache general rules for optimal play in various
classes of situations.

43
Lots of theoretical problems...
 Minimax only valid on whole tree
 P(win) is not well defined
 Correlated errors
 Perfect play assumption
 No planning

44