heuristic search. heuristic 1.from greek heuriskein, “to find”. 2.of or relating to a usually...

HEURISTIC SEARCH

Heuristic

1. From Greek heuriskein, “to find”.

2. Of or relating to a usually speculative formulation serving as a guide

in the investigation or solution of a problem.

3. Computer Science: Relating to or using a problem-solving technique

in which the most appropriate solution of several found by alternative

methods is selected at successive stages of a program for use in the

next step of the program.

4. the study of methods and rules of discovery and invention.

5. In state space search, heuristics are formalized as rules for

choosing those branches in a state space that are most likely to

lead to an acceptable problem solution.

• Al problem solvers employ heuristics in two situations:-

– A problem may not have an exact solution, because

of inherent ambiguities in the problem statement or

available data.

• Medical diagnosis is an example of this. A given

set of symptoms may have several possible

causes. Doctors use heuristics to chose the most

likely diagnosis and formulate a plan of treatment.

INTRODUCTION

• Al problem solvers employ heuristics in two situations:-

– A problem may not have an exact solution, because

of inherent ambiguities in the problem statement or

available data.

• Vision is another example of an inherently inexact

problem. Visual scenes are often ambiguous,

allowing multiple interpretations of the

connectedness, extent and orientation of objects.

Optical illusions exemplify these ambiguities.

Vision systems use heuristics to select the most

likely of several possible interpretations of a given

scene.

INTRODUCTION

• Al problem solvers employ heuristics in two situations:-

– A problem may have an exact solution, but the

computational cost of finding it may be prohibitive. In

many problems, state space growth is

combinatorially explosive, with the number of

possible states increasing exponentially or factorially

with the depth of the search.

– Heuristics handle above problem by guiding the

search along the most “promising” path through the

space. By eliminating unpromising states and their

descendants from consideration, a heuristic

algorithm can defeat this combinatorial explosion

and find an acceptable solution.

INTRODUCTION

• Heuristics and the design of algorithms to implement

heuristic search have been an important part of artificial

intelligence research.

• Game playing and theorem proving require heuristics to

reduce search space to simplify the solution finding.

• It is not feasible to examine every inference that can be

made in search space of reasoning or every possible

move in a board game to reach to a solution. In this case,

heuristic search provides a practical answer.

INTRODUCTION

• Heuristics are fallible.

– A heuristics is only an informed guess of the next step to be taken in solving a problem. It is often based on experience or intuition.

– Heuristics use limited information, such as the descriptions of the states currently on the open list, they are seldom able to predict the exact behavior of the state space farther along in the search.

– A heuristics can lead a search algorithm to a sub optimal solution or fail to find any solution at all.

INTRODUCTION

Tic-Tac-Toe

• The combinatorics for exhaustive search are high.

– Each of the nine first moves has eight possible responses.

– which in turn have seven continuing moves, and so on.

– As simple analysis puts exhaustive search at 9 x 8 x7x … or 9!.

• Symmetry reduction can decrease the search space, there are really only three initial moves.

– To a corner.

– To the center of a side.

– To the center of the grid.

Tic-Tac-Toe

Symmetry reductions on the second level of states further reduce the

number of possible paths through the space. In following figure the

search space is smaller then the original space but It is still factorial in its

growth.

Tic-Tac-Toe

• Heuristics

– Algorithm analysis the moves in which X has the most

winning lines.

– Algorithm then selects and moves to the state with the

highest heuristic value i.e. X takes the center of the grid.

Figure 4.2: The “most wins” heuristic applied to the first children in tic-tac-toe.

Tic-Tac-Toe• Heuristics

– Other alternatives and their descendants are eliminated. – Approximately two-thirds of the space is pruned away with

the first move .– After the first move, the opponent can choose either of two

alternative moves.

Tic-Tac-Toe

• Heuristics

– After the move of opponent,

The “max win lines” heuristics

can be applied to the resulting

state of the game.

– As search continues, each

move evaluates the children

of a single node: exhaustive

search is not required.

– Figure shows the reduced

search after three steps in the

games. States are marked

with their heuristics values.

Tic-Tac-Toe

• COMPLAXITY

– It is difficult to compute the exact number of states that must

be examined. However, a crude upper bound can be

computed by assuming a maximum of nine moves in a

game and eight children per move.

– In reality:

• The number of states will be smaller, as board fills and

reduces our options.

• In addition opponent is responsible for half the moves.

• Even this crude upper bound of 8 x 9 or 72 states is an

improvement of four orders of magnitude over 9!.

ALGORITHM FOR HEURISTICS SEARCH

Implementing Best-First-Search

• Best first search uses lists to maintain states:

– OPEN to keep track of the current fringe of the search.

– CLOSED to record states already visited.

• Algorithm orders the states on OPEN according to some

heuristics estimate of their “closeness” to a goal.

• Each iteration of the loop consider the most “promising” state

on the OPEN list.

• Example algorithm sorts and rearrange OPEN in precedence of

lowest heuristics value.

BEST-FIRST-SEARCH

BE

ST

-FIR

ST

-SE

AR

CH

Figure: Heuristic search of a hypothetical state space with OPEN and CLOSED states highlighted.

BEST-FIRST-SEARCH

• At each iteration best-first –search removes the first element from the OPEN list. If it meets the goals conditions, the algorithm returns the solution path that led to the goal.

• Each state retains ancestor information to determine if it had previously been reached by a shorter path and to allow the algorithm to return the final solution path.

• If the first element on OPEN is not a goal, the algorithm generate its descendants.

• If a child state is already on OPEN or CLOSED, the algorithm checks to make sure that the state records the shorter of the two partial solution paths.

• Duplicate states are not retained.

• By updating the ancestor history of nodes on OPEN and CLOSED when they are rediscovered, the algorithm is more likely to find a shorter path to a goal.

BEST-FIRST-SEARCH

• Best-first-search applies a heuristics evaluation to the states on

OPEN, and the list is sorted accordingly to the heuristic values of

those states. This brings the best states to the front of OPEN .

• Because these estimates are heuristic in nature, the next state to

be examined may be from any level of the state space. When

OPEN is maintained as a sorted list, it is often referred a priority

queue.

• Figure 4.4 shows a hypothetical state space with heuristic

evaluations attached to some of its states.

• The goals of best –first search is to find the goals state by looking

at as few states as possible: the more informed the heuristic, the

fewer states are processed in finding the goal.

BEST-FIRST-SEARCH

• The best-first search algorithm always selects the most

promising state on OPEN for further expansion.

• It does not abandon all the other states but maintains them on

OPEN .

• In the event a heuristic leads the search down a path that

proves incorrect, the algorithm will eventually retrieve some

previously generated next best state from OPEN and shift its

focus to another part of the space.

• In the example after the children of state B were found to have

poor heuristic evaluations, the search shifted its focus to state

C.

• In best-first-search the OPEN list allows backtracking from

paths that fail to produce a goal.

BEST-FIRST-SEARCH

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS• We will evaluate performance of various heuristics for

solving the 8-puzzle.

• Figure shows a start and goal state for the 8-puzzle, along with the first three states generated in the search.

• Heuristic 1

– The simplest heuristic, counts the tiles out of place in each state when it is compared with the goal.

– The state that had fewest tiles out of place is probably closer to the desired goal and would be the best to examine next.

– However, this heuristic does not use all of the information available in a board configuration, because it does not take into account the distance the tiles must be moved.

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

• Heuristic 2

– A “better” heuristic would sum all the distances

by which the tiles are out of place, one for each

square a tile must be moved to reach its

position in the goal state.

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

• Heuristic 3

– Above two heuristics fail to take into account the

problem of tile reversals.

– If two tiles are next to each other and the goal

requires their being in opposite locations, it takes

more than two moves to put them back to place,

as the tiles must “go around” each other.

– A heuristic that takes this into account multiplies

a small number (2 , for example) times each

direct title reversal.

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

• The “sum of distances” heuristic provides a

more accurate estimate of the work to be done

than the “number of titles out of place” heuristic.

• Title reversal heuristic gives out an evaluation of

‘0’ since non of these states have any direct tile

reversals.

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

• Devising A Good Heuristic

– Good heuristics are difficult to devise. Judgment and intuition help, but the final measure of a heuristic must be its actual performance on problem instances.

– Each heuristic proposed above ignores some critical information and needs improvement.

– An improved version is discussed next.

• Devising A Good Heuristic

– The distance from the starting state to its

descendants can be measured by maintaining a

depth count for each state. This count is ‘0’ for the

beginning state and may be incremented by ‘1’ for

each level of the search.

– It records the actual number of moves that have

been used to go from the starting state in the search

to each descendant.

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

• Devising A Good Heuristic

– Let evaluation function ‘f(n)’, be the sum of two components:

f(n) = g(n) + h(n)

– Where:-

• g(n) measures the actual length of the path from start state to any state n.

• h(n) is a heuristic estimate of the distance from the state n to a goal.

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

IMPLEMENTING HEURISTIC EVALUATION FUNCTIONS

h(n) = 5 h(n) = 3 h(n) = 5

BEST-FIRST SEARCH GRAPH

• Each state is labeled with a letter and its

heuristic weight, f(n) = g(n) + h(n).

• The number at the top

of each state indicates

the order in which it

was taken off the

OPEN list.

Successive stages of OPEN and CLOSED that generates the graph are:

• The g(n) component of the evaluation function gives

the search more of a breadth-first flavor.

• This prevent it from being misled by an erroneous

evaluation, If a heuristic continuously returns “good”

evaluations for states along a path that fails to reach

a goal, the g value will grow to dominate h and force

search back to a shorter solution path.

• This guarantees that the algorithm will not become

permanently lost, descending an infinite branch.

HEURISTIC SEARCH AND EXPERT SYSTEMS

• Simple games such as the 8-puzzle are ideal vehicles for exploring

the design and behavior of heuristic search algorithms.

• More realistic problems greatly complicate the implementation and

analysis of heuristic search by requiring multiple heuristics to deal

with different situations in the problems space.

• A single heuristic may not apply to each state in these domains.

• Instead, situation specific problem-solving heuristics are encoded to

deal with various states.

• Each solution step incorporates its own heuristic that determines

when, it should be applied.

EXAMPLE: THE FINANCIAL ADVISOR • An individual with adequate savings and income should invest in stocks.

saving_account(adequate)income(adequate)investment(stocks).

• The individual may prefer the added security of a combination strategy.

saving_account(adequate)income(adequate)investment(combination).

• The individual may prefer placing all investment money in savings.

saving_account(adequate)income(adequate)investment(savings).

• Thus the rule is a heuristic in nature. The problem solver should be programmed to account for this uncertainty. Additional factors may also be taken into account to make the rule more informed and capable of finer distinctions.

– Age of the investor.

– Long term prospects for security.

– Advancement in investors profession.

EXAMPLE: AN EXPERT SYSTEMTHE FINANCIAL ADVISOR

• Now how the financial advisor is going to weigh the above heuristic to decide the best option or to find out a solution which is closer to the best solution.

• One way is to attach a confidence measure with the conclusion of each rule. Attach a real number between –1 to 1 as:-

• 1 corresponding to certainty ‘True’.

• -1 to define a value of ‘False’.

• Values in between reflect varying confidence levels.

• This way the preceding rules would be reflected as under:

saving_account(adequate)income(adequate)investment(stocks).

with confidence = 0.8

saving_account(adequate)income(adequate)investment(combination).

with confidence = 0.5

saving_account(adequate)income(adequate)investment(savings).

with confidence = 0.1

ADMISSIBILITY MEASURESADMISSIBILITY MEASURES

• A search algorithm is admissible if it is guaranteed to find a minimal path to solution, whenever such a path exists.

• Breadth-first search is an admissible search strategy .

• In determining the properties of admissible heuristics, we define an evaluation function f*.

f*(n) = g*(n) + h*(n)

Let: f(n) = g(n) + h(n)

• g(n) measures the depth(cost) at which state n has been found in the graph.

• h(n) is the heuristic estimate of the cost from n to a goal.

• f(n) is the estimate of total cost of the path from the start state through node n to the goal state.

Let: f*(n) = g*(n) + h*(n)

• g*(n) is the cost of the shortest path from the start node to node n

• h*(n) returns the actual cost of the shortest path from n to the goal.

• f*(n) is the actual cost of the optimal path from a start node through node n to a goal node.

ADMISSIBILITY MEASURESADMISSIBILITY MEASURES

• In an algorithm A, Ideally function f should be a close estimate of f*.

– The cost of the current path, g(n) to state n, is a reasonable estimate of g*(n).

– Where: g(n) g*(n).

– These are equal only if the graph search has discovered the optimal path to state n.

• Similarly, we compare h*(n) with h(n).

• h(n) is bounded from above by h*(n) i.e, h(n) is always less than or equal to the actual cost of a minimal path h*(n) i.e. h(n) h*(n).

• If algorithm A uses an evaluation function f in which h(n) h*(n), the algorithm is called algorithm A*.

The heuristic of counting the number of tiles out of place is certainly less than

or equal to the number of moves required to move them to their goal position.

Thus this heuristic is admissible and guarantees an optimal solution path.

h*(n)h(n)

•Comparison of search using heuristics with search by breadth-first search.

•Heuristic used is f(n) = g(n) + h(n) where h(n) is tiles out of place.

The portion of the graph searched heuristically is shaded. The optimal solution path is in bold.

nj

ni

ng

• Monotonic heuristic is admissible. Consider any path in the space as a sequence of states S1, S2, …… Sg, where S1 is the start state and Sg is the goal: For the sequence of moves in this arbitrary selected path:

S1 to S2 h(S1) - h(S2) cost (S1, S2) by monotone property

S2 to S3 h(S2) - h(S3) cost (S2, S3) by monotone property

S3 to S4 h(S3) - h(S4) cost (S3, S4) by monotone property

. . . . by monotone property

. . . . by monotone property

Sg-1 to Sg h(Sg-1) - h(Sg) - cost (Sg-1, Sg) by monotone property

• Summing each column and using the monotone property of h(Sg) = 0:

path S1 to Sg h(S1) cost (S1, Sg)

This means that monotone heuristic h is A* and admissible.

• If breadth first search is considered a A* algorithm with heuristic h1, then it must be less than h* (as breadth first search is admissible).

• Also number of tiles out of place with respect to goal state is a heuristic h2. As this is also admissible therefore, h2 is less than h*.

• In this case h1 h2 h*. It follows that the “number of tiles out of place” heuristic is more informed than breadth-first search.

• Both h1 and h2 find the optional path, but h2 evaluates many fewer states in the process.

• Similarly the heuristic that calculates the sum of the direct distances by which all the tiles are out of place is again more informed than the calculation of the number of tiles are out of place with respect to the goal state.

• One can visualize a sequence of search spaces, each smaller than the previous one, converging on the direct optimal path solution.

• If a heuristic h2 is more informed then h1 then the set of states examined by h2 is a subset of those examined by h1.

• In general, then the more informed and A* algorithm, the less of the space it needs to expand to get the optimal solution.

Figure 4.17: Two-ply minimax applied to the opening move of tic-tac-toe, from Nilsson (1971).

Figure 4.18: Two-ply minimax and one of two possible MAX second moves, from Nilsson (1971).

Figure 4.19: Two-ply minimax applied to X’s move near the end of the game, from Nilsson (1971).

Figure 4.20: Alpha-beta pruning applied to state space of Figure 4.15. States without numbers are not evaluated.

Figure 4.21: Number of nodes generated as a function of branching factor, B, for various lengths, L, of solution paths. The relating equation is: T = B(BL - 1)/(B - 1), adapted from Nilsson (1980).

Figure 4.22: Informal plot of cost of searching and cost of computing heuristic evaluation against informedness of heuristic, adapted from Nilsson (1980).

Figure 4.23: The sliding block puzzle.

Figure 4.24

Figure 4.25

Alpha-beta Search

Two-player games• The object of a search is to find a path from

the starting position to a goal position

• In a puzzle-type problem, you (the searcher) get to choose every move

• In a two-player competitive game, you alternate choosing moves with the other player

• The other player doesn't want to reach your goal

• Your search technique must be very different

Payoffs

• Each game outcome has a payoff, which we can represent as a number

• By convention, we prefer positive numbers

• In some games, the outcome is either a simple win (+1) or a simple lose (-1)

• In some games, you might also tie, or draw (0)

• In other games, outcomes may be other numbers (say, the amount of money you win at Poker)

Zero-sum games• Most common games are zero-sum: What I win

($12), plus what you win (-$12), equals zero

• There do exist non-zero-sum games

• For simplicity, we consider only zero-sum games

• From our point of view, positive numbers are good, negative numbers are bad

• From our opponents point of view, positive numbers are bad, negative numbers are good

A trivial "game"

• Wouldn't you like to win 50?

• Do you think you will?

• Where should you move?

Minimaxing• Your opponent will

choose smaller numbers

• If you move left, your opponent will choose 3

• If you move right, your opponent will choose -8

• Your choice is 3 or -8• Play will be as shown

Heuristics• In a large game, you don't really know the payoffs• A heuristic function computes, for a given node, your

best guess as to what the payoff will be• The heuristic function uses whatever knowledge you

can build into the program• We make two key assumptions:

– Your opponent uses the same heuristic function– The more moves ahead you look, the better your heuristic

function will work

PBVs• A PBV is a preliminary backed-up value

– Explore down to a given level using depth-first search

– As you reach each lowest-level node, evaluate it using your heuristic function

– Back up values to the next higher node, according to the following rules:

• If your move, bring up the largest value, possibly replacing a smaller value

• If opponent's move, bring up smallest value, possible replacing a larger value

Using PBVs

• Explore left-bottom 8; bring it up

• Explore 5; smaller, so ignore it

• Bring 8 up another level• Explore 2; bring it up• Explore 9; better, so

bring up, replacing 2• Don't replace 8

Bringing up values

• If it's your move, and the next child of this node has a larger value than this node, replace this value

• If it's your opponent's move, and the next child of this node has a smaller value than this node, replace this value

• At your move, never reduce a value

• At your opponent's move, never increase a value

Alpha cutoffs

• The value at your move is 8 (so far)

• If you move right, the value there is 1 (so far)

• Your opponent will never increase the value at this node

• You can ignore the black nodes

Beta cutoffs

• An alpha cutoff occurs where– It is your opponent's turn to move– The node has a lower PBV than some other node you could

choose– Your opponent will never increase the PBV of this node

• A beta cutoff occurs where– It is your turn to move– The node has a higher PBV than some other node the opponent

could choose– You will never decrease the PBV of this node

Using beta cutoffs

• Beta cutoffs are harder to understand, because you have to see things from your opponent's point of view

• Your opponent's alpha cutoff is your beta cutoff• We assume your opponent is rational, and is

using a heuristic function similar to yours• Even if this assumption is incorrect, it's still the

best we can do

The importance of cutoffs

• If you can search to the end of the game, you know exactly how to play

• The further ahead you can search, the better

• If you can prune (ignore) large parts of the tree, you can search deeper on the other parts

• Since the number of nodes at each level grows exponentially, the higher you can prune, the better

• You can save exponential time

Heuristic alpha-beta searching

• The higher in the search tree you can find a cutoff, the better (because of exponential growth)

• To maximize the number of cutoffs you can make:– Apply the heuristic function at each node you come to,

not just at the lowest level– Explore the "best" moves first– "Best" means best for the player whose move it is at

that node

Best game playing strategies

• For any game much more complicated than tic-tac-toe, you have a time limit

• Searching takes time; you need to use heuristics to minimize the number of nodes you search

• But complex heuristics take time, reducing the number of nodes you can search

• Seek a balance between simple (but fast) heuristics, and slow (but good) heuristics

The End