Presentations:

	A paper (someone else's work)
		you may need to dig out references
		ask me; go to the library

	20+5 minutes each
	
		expect questions, ask questions

		You are, in effect, teaching the class
		something we don't know

	Transparencies

		I can provide a few for each presenter
		

	Blackboard/Whiteboard

		Use them if you want
		have a spare transparency or two

	After the presentation

		I will critique, privately



Why is it a good idea to give exponentially
increasing trials to above average schemata/schemas?

Two Armed Bandit problem (Slot machine)
Each arm represents a competing schema

Arm1 : 
payoff	U1   variance  s1^2

Arm1 : 
payoff	U2   variance  s2^2

Also 
	U1 > U2

To maximize  payoff
Which arm should be play?



Clearly U1

BUT

	We don't know BEFOREHAND which arm has
higher payoff!

	We can play the machines and sample payoffs

Experiment:

Total of N plays or N trials

allocate n trials to arm1 
allocate n trials to arm2

where	2n < N

allocate N - 2n trials to arm with best payoff

losses:
	1. Losses due to lower payoffs on arm2
	2. Losses due to choosing arm2 after
	   2n trials



To minimize losses
	What is an optimal value of n? 
	n*

and therefore what should be the number of trials 
given to the better arm

Holland calculated this to be

	exponential in n*

The GA comes close
by giving AT LEAST an exponentially increasing 
number of trials to above average schema.



Homework:

Look at /sushil/classes/gas/code/original

	a. Four different selection functions
		Roulette Wheel
		CHC selection
		Scaled roulette wheel
		Scaled CHC
	b. Different Crossovers
		1pt, 2pt, Uniform crossover

Increased selection pressure:
	quicker convergence (perhaps to a local optima)
	

Increased Crossover sites
	More exploration, more disruption

Balance