SEARCH


	Generate and Test		Simplest/RES
	
	Generate a candidate solution and test to see
	if it solves the problem, 
	Repeat until you solve the problem	

	Very little information: guaranteed to succeed
	


	Hill Climbing

	Generate a candidate solution by modyfying 
	a previous solution and and test to see
	if the new solution solves the problem
	Repeat until you solve the problem

	Local Search

	More information, Makes the assumption that.....
	Ie: local information


	Population of Hill-Climbers

	Do this in parallel with as many resources 
	as you can lay your hands on. Choose the first
	one that solves your problem

	Parallel

	Lots of local searches
	Same information
	Ie: local information



	Population of information exchanging Hill-Climbers
	Genetic Algorithms

	Same information
	But exchanging information means
	it is a more global search

	Parallel




	Hard Problems:		Computational Complexity

	Binary Search		log N

	Linear search		N

	Bubble  Sort		N^2

	Scheduling: 		NP-complete	a^N


	
	Hard Problems:		Poorly-defined

	Robotics		How do we manage to catch 
				a ball, Navigation, Competition

	Seismology		Predicting subsurface structure

	User Interfaces		Predict next command

	Medicine		Will this patient live?

	Design			Design a bridge/truss, 
				Analog-Circuit, Antennas

	Chemical equilibrium

	Control			Process control....



Search as a solution to hard problems

Basic Strategy: Generate and Test
Generate a potential solution and test to see if the 
potential solution solves the problem

Information to guide generation of potential solutions?

	No information means RES is our only recourse 
	and we wait. 
		Lots of EXPLORATION of the search space
		No information to EXPLOIT
		Lot of time

	Lots of information means we use information to 
	directly solve problem. 
		Little EXPLORATION of the search space
		EXPLOIT information
		Little Time

	If we can compare two solutions and tell which is 
	better, we can use a variety of search algorithms
		tradeoff EXPLORATION and EXPLOITATION


	Increasing Exploration means
		Better chance of finding good solution
		But Increasing Time


	Increasing (if we can) Exploitation means 
		Less Time
		But can get stuck in local optima which means
                    less chance of finding good solution


The amount of Information available about a problem
influences our choice of an appropriate search algorithm
and how we TUNE the search algorithm

TUNING:

	Balancing EXPLOITATION against EXPLORATION

Assumptions, Search Bias, and Information:

RES:
	Search Bias: None
	Assumptions: Tester will tell us if/when we reach solution
	Information: Solution is in search space


Hillclimbing: 
	Search Bias: +Follow gradient -- climb the hill
	Assumptions: +Following gradient will lead to solution
	Information: +Gradient information is NOT misleading
		
GAs and other, similar algorithms:
	Search Bias: +Follow gradient -- but hedge your bets
		      sometimes go against gradient
	Assumptions: +Following gradient will USUALLY lead 
		     to solution
		     +Exchanging information USUALLY helps
	Information: +Gradient information is USUALLY a good 
		     idea



	Questions:

	1) What is the  encoding? 
		How do we generate candidate solutions?
		What are the Operators ?
		What is the search bias ?
		
	2) Where is the balance between 
		speed (converging quickly) (EXPLOITATION)
			and
		coverage (do we cover the  (EXPLORATION)
			  entire gamut of 
			  possible solutions?)