1

Informed Search and Informed Search and ApplicationsApplications

Reading: Recommended paper:``Focused Crawling: a new approach to topic-specific Web resource discover”,Soumen Chakrabarti, Martin van den Berg, Byron Domhttp://www.cs.berkeley.edu/~soumen/doc/ww

w1999f/pdf/www1999f.pdf

Next class: chapter 5

The book has been placed on reserve. Be sure you

2

Online SearchOnline Search

� Agent operates by interleaving computation and action

� No time for thinking

� The agent only knows� Actions (s)� The step-cost function c(s,a,s’)� Goal-text (s)

� Cannot access the successors of a state without trying all actions

3

What properties of search are What properties of search are desirable?desirable?� Will A* work?

� Expand nodes in a local order� Depth first� Variant of greedy search

� Difference from offline search:� Agent must physically backtrack� Record states to which agent can backtrack and has not

yet explored

4

Online DFS Online DFS -- setupsetup

� Inputs: s’, a percept that identifies the current state

Static: � result, a table indexed by action and state,

initially empty� unexplored: a table that lists, for each visited

state, the actions not yet tried� unbacktracked: a table that lists, for each

visited state, the backtracks not yet tried� s,a: the previous state and action, initially null

5

Online DFS Online DFS –– the algorithmthe algorithm� If Goal-test(s’) then return stop� If s’ is a new state then unexplored[s’] � actions(s’)� If s is not null then do

� result[a,s]�s’� Add s to the front of unbacktracked[s’]

� If unexplored[s’] is empty� Then return stop� Else a� action b such that

result[b,s’]=pop(unbacktracked[s’])� Else a�pop(unexplored[s’])� s� s’� Return a

6

Local Search AlgorithmsLocal Search Algorithms

� Operate using a single current state

� Move only to neighbors of the state

� Paths followed by search are not retained

� Iterative improvement� Keep a single current state and try to improve it

7

Advantages to local searchAdvantages to local search

� Use very little memory – usually a constant amount

� Can often find reasonable solutions in large or infinite state spaces (e.g., continuous)

� Unsuitable for systematic search

� Useful for pure optimatization problems� Find the best state according to an objective function� Traveling salesman

8

9

10

Steepest Ascent

11

12

Problems for hill climbingProblems for hill climbing

When the higher the heuristic function the better: maxima (objective fns); when the lower the function the better: minima (cost fns)

� Local maxima: A local maximum is a peak that is higher than each of its neighboring states, but lower than the global maximum

� Ridges: a sequence of local maxima� Plateaux: an area of the state space landscape

where the evaluation function is flat

13

Some solutionsSome solutions� Stochastic hill-climbing

� Chose at random from among the uphill moves

� First-choice hill climbing� Generates successors randomly until one is generated

that is better than current state

� Random-restart hill climbing� Keep restarting from randomly generated initial states,

stopping when goal is found

� Simulated annealing� Generate a random move. Accept if improvement.

Otherwise accept with continually decreasing probability.

� Local beam search� Keep track of k states rather than just 1

14

Back to Online SearchBack to Online Search

� Would hill-climbing be appropriate?

15

Learning Real Time A* Learning Real Time A* (LRTA*)(LRTA*)

� Augment hill-climbing with memory

� Store current best estimate of cost from node to goal: H(s)

� Initially, H(s) = h(s)

� Update H(s) through experience

� Estimated cost to reach the goal through neighbor s’� H(s) = c(s,a,s’)+ H(s’)

16

LRTA* agentLRTA* agent

� Inputs: s’, a percept that identifies the current state

� Static: result, a table indexed by action and state, initially empty

� H, a table of cost estimates indexed by state, initially empty

� S,a, the previous state and action, initially null

17

LRTA*(LRTA*(s’s’))

� If Goal-test(s’) then return stop� If s’ is a new state (not in H), then H[s’]� h(s’)� Unless s is null

� Result[a,s]� s’� H[s] � min LRTA*-Cost(s,b,result[b,s],H)

over all b from Actions(s)

� a� an action b in Actions(s’) that minimizes LRTA*-Cost(s’,b,result[b,s’],H)

� s� s’� Return a

18

LRTA*LRTA*--Cost(Cost(s,a,s’,Hs,a,s’,H) )

� Returns a cost-estimate

� If s’ is undefined then return h(s)� Else return c(s,a,s’) + H[s’]

19

RealReal--World Application for AI World Application for AI SearchSearchThe Problem

� World Wide Web is a vast resource� Google reports indexing 3.3 billion web pages as of 1/2004� About 600GB of text changes every month

� A search engine requires continuous crawls to index all pages

� Inktomi:� Cluster of 100s of Sun Sparc stations� 75GB RAM each� 1 TB disk� Crawls greater than 10 million pages/day

20

Alternative: focused crawlAlternative: focused crawl

� Start with a user specified topic hierarchy (like Yahoo, but user specific)

� Used for training the crawler to distinguish relevant pages from irrelevant pages

� Classifier: a learned function that predicts relevance

� Resource discovery: Starting from a node in the hierarchy the crawler branches out to find other relevant pages

� Must determine which outgoing links are good ones

� Simultanously, the crawler runs a topic distillation algorithm to identify hubs

� Pages with large numbers of links to relevant documents

21

Why is this an AI Search Why is this an AI Search Problem?Problem?� What is the search space?

� What is the goal test?

� What is the heuristic function?

22

System ArchitectureSystem Architecture

� Classifier: makes relevance judgments on pages crawled to decide on link expansion

� Distiller: determiens a measure of centrality of crawled pages to determine viist priorities

� Crawler: dynamically reconfigurable priority controls governed by classifier and distiller

23

DistillationDistillation

� A good strategy for the crawler is to identify hubs: pages that are almost exclusively collection of links to authoritative resources that are relevant to the topic

� Authorities: a web page with many incoming links, particularly from high-prestige, relevant pages

24

Identifying hubs and Identifying hubs and authoritiesauthorities� Each node has two scores, iteratively

determined� a(v): number of incoming edges from relevant

nodes� h(v): number of outgoing edge to relevant

nodes

� Weight these scores by the relevance scores of the pages they point to (a probability between 0 and 1)