COMMON AI ARCHITECTURES

	HARD-CODED/EMBEDDED (IMPLICIT) LOGIC PROGRAM
		CONTROL PROGRAM

	RULE-BASED REASONING PROGRAM
		PROLOG PROGRAM

	CASE-BASED REASONING PROGRAM

	LEARNING PROGRAM
		NEURAL NET PROGRAM

	INDUCTIVELY REASONING PROGRAM
	PROBABILISTICALLY REASONING PROGRAM
	EXPECTED UTILITY & PROBABILISTICALLY REASONING PROGRAM

	EXPLICITLY LOGICALLY-DEDUCING PROGRAM
	FUZZY LOGIC PROGRAM

	CONSTRAINT SATISFACTION PROGRAM

	PLANNING PROGRAM

	OPTIMIZING PROGRAM
	HEURISTICALLY OPTIMIZING PROGRAM
		HEURISTICALLY OPTIMIZING THROUGH TREE-SEARCH PROGRAM
		HEURISTICALLY OPTIMIZING THROUGH GRADIENT-ASCENT PROGRAM
		HEURISTICALLY OPTIMIZING THROUGH GRADIENT-ASCENT+SIMULATED ANNEALING PROGRAM

	GENETIC PROGRAM

	GRAMMAR/GENERATIVE PROGRAM

PREFERENCE

	a > b, "a is preferred to b"

UTILITY

	what makes it possible to measure the intensity of preference?
	u(a) a real number?  why not an integer?  what is it that makes something extensive?
	are there other scales that are unique up to affine transformation?

	lotteries.
	
	independence.  substitutivity.  totality.

	is u a function of one variable?
	is it convex?  strictly monotonic?

	how is u elicited/specified?

PROBABILITY

	for A = {a,b,c}, mutually exclusive & exhaustive,
	if you tell me p(i), i in A, i can fill in p(s) s in 2^A, also do conditional p.

	what if you tell me some p's, e.g., p({a,b}) = .5, what is p({b,c})?
	or p(foo|bar) = .9, p(bar|moo) = .9, what is p(foo|moo)?

	use maximum entropy?
	what if p(foo|bar,moo) = p(foo|bar)?

	statistical inference from data:  maximimum likelihood estimators?
	
HEURISTIC OPTIMIZATION

	an objective function f(x), f(x,y), or f(x1, ..., xn).
	what is max f?  what is argmax f?

	how long does f take to evaluate?  .000001 sec?  or 10 hrs?
	f could be a simulation with parameters

	x1, ..., xn could be booleans

	1.  random evaluation
		for (i=1; i<=1M; i++) { x = rand(); savebest(f(x)) }

	2.  gradient ascent
		x = rand()
		for (i=1; i<=1M; i++)
		  foreach (x' in {x+delta, x-delta})
		    x = savebest(f(x')) 

	3.  gradient ascent with monte carlo starts
		for (j=1; j<=1M; j++) {
		  x = rand()
		  for (i=1; i<=1M; i++)
		    foreach (x' in {x+delta, x-delta})
		      x = savebest(f(x')) 
		}

	4.  simulated annealing
		x = rand()
		for (i=1; i<=1M; i++)
		  if (rand() < 1/e^-k(i))
		    foreach (x' in {x+delta, x-delta})
		      x = savebest(f(x')) 
		  else
		    x' = rand({x+delta, x-delta})

	5.  hybrid simulated annealing
		for (d=1; d<=n; d++) x(d) = rand()
		for (i=1; i<=1M; i++)
		  if (rand() < 1/e^-k(i)) {
		    S = Union(d'=1..n)({x+delta(d'), x-delta(d')})
		    S' = randomsample(i,S)
		    foreach (x' in S')
		      x = savebest(f(x')) 
		  } else {
		    d' = rand(1..n)
		    x' = rand( {x+delta(d'), x-delta(d')} )
		  }

	how much better is 5 than 4?
	(one recent paper:  much better!)