Approximate Inference Sampling

Approximate Inference Sampling
P(B|+a)
Likelihood weighting
Inconsistent

GIBBS Sampling
Markov chain monte carlo mcmc
+c +s -r -w
+c -s -l -w
+c -s +r -w

P(A) = 0.5, P(B|A)=0.2, P(B|¬A)=0.8
P(¬A)=1-P(A)=0.5
P(A|B)=0.2
P(A|B)=P(B|A)P(A)/P(B)=0.2*0.5/P(B|A)P(A)+P(B|¬A)P(¬A)
=0.1/(0.1+0.8+0.5)=0.2

Simple Bayes Net
P(A)=0.5, Vi:P(Xi|A)=0.2, P(Xi|¬A)=0.6
P(A|X1 X2 ¬X3)=P(¬X3|A)P(A|X1X2)
P(¬A|x1X2¬X3)αP(¬X3hA)P(x2|¬A) P(x1|¬A)P(¬A)

Probabilistic Inference

Probabilistic Inference
-Probability theory
-Bayes network
-Independence
-Inference

Evidence -> Hidden -> Query
P(Q1,Q2…|E1=e1, E2=e2)

Enumeration
P(+b|+j,+m)=P(+b,+j,+m)/P(+j,+m)
P(+b,+j,+m)=ΣeΣaP(+b,+j,+m,e,a)

Pulling out terms
P(+b)ΣeΣaP(e)P(a|+b,e)P(+j|a)P(+m,a)

Maximize independence
O(n), O(2n)

Variable Elimination

R->T->L
P(+l)=ΣrΣt P(r)R(t|r)P(+l|t)

Different type of bayes network

Different type of bayes network
P(S)= 0.7
P(R)= 0.01
P(H|S,R) = 1
P(H|¬S,R) = 0.9
P(H|S,¬R) = 0.7
P(H|¬S,¬R) = 0.1
P(R|S)=0.01
P(R|H1 S)=0.0142
= P(H|R1 S)*P(R|S)/P(H|S) = P(H|R1S)*P(R)/P(H|R1S)P(R)+ P(H|TR1S)P(¬R)
P(R|H) = P(H|R)P(R)/P(H)= 0.97*0.01/0.5245 = 0.0185
P(R|H,¬S) = P(H|R,¬S)P(R|¬S)/P(H|¬S) = 0.009/0.9*0.01+0.1*0.99=0.0083

Computing Bayes Rule

Bays rule
P(A|B) = P(B(A))P(A)/P(B)
P(¬A|B) = P(B|¬A)P(¬A)/P(B)
P(A|B)+P(¬A|B) = 1

P'(A|B) = P(B|A)P(A)
P'(¬A|B) = P(B|¬A)P(¬A)

P(C)=0.01
P(+|c)=0.9
P(-|+c)=0.8
P(¬C)=0.99
P(-|C)=0.1
P(+|+C)=0.2

Conditionally Independent
P(T2|C1 T1) = P(T2|C)

P(T2 =+1 | T1 =t)
= P(+2 | +1 ,C) P(C|+1)+P(+2|+,¬C)P(¬C|+1)
= P(+2|C)*0.043 + P(+2|¬C)*0.957
= 0.2301

Weather

P(D1) P(D1=sunny)=0.9
P(D2=sunny | D1=sunny) = 0.8
P(D2 = rainy|D1=sunny) =  0.2
P(D2 = sunny|D1 = rainy) = 0.6
P(D2 = rainy|D1 = rainy) = 0.4

Bayes Rule
P(A|B) = P(B|A)* P(A)/ P(B)
Posterior = Likelihood * prior / marginal likekihood

P(c|+) = P(+|c)*P(c)/ P(+) = 0.9 * 0.01/ 0.9*0.01 + 0.2*0.99

bays rule
A: not observable P(A)
B: observable P(B|A), P(B|¬A)
Diagnostic reasoning: P(A|B), P(B|¬A)

Probability in AI

Bayes network

altenator broken, fanbelt broken ->
battery dead -> battery flat -> car won’t start
-battery meter, battery age
light, oil light, gas gague
no oil, no gas, fuel line blocked, starter broken

Binary events
Probability
Simple bayes networks
Conditional independence
Bayes networks
D-seperation
Parameter counts

Bayes networks -> diagnostics, prediction, machine learning
Finance, Google, Robotics
particle filters, HMM, MDP + POMDPs, KALMAN filters …

Probabilities is certainty in AI
P(head) = 1/2, P(Tail) = 1/2
P(h, h, h) = 1/8, P(h) = 1/2
P(x1=x2=x3=x4)=0.125,
P({x1,x2,x3,x4} contains >= 3 h) = 5 / 16

Search Comparison

Breakth-frist
Cheapest-first
Depth-first

Greedy best-first search
A* algorithm
f = g + h
g(path) = path cost
h(path) = h(s) = estimated distance to goal

A* finds lowest cost path is:
h(s) < true cost Sliding blocks puzzle (15puzzle) h1 = #misplaced blocks h2 = sum(distances of blocks) a block can move A -> B
if (A adjacent to B)
and (B is blank)
h2 h1
h = max(h1, h2)

Problem-solving works when:
-fully observable
-known
-discrete
-deterministic
-static

AI and Uncertainty

AI as uncertainty management
AI = What to do when you don’t know what to do?
Reasons for uncertainty.
Sensor limits

Definition
-initial state
-action(s) -> {a1, a2, a3 …}
-result(s,a) -> s1
-GoalTest(s) -> T|F
-PATH Cost(s->a -> s ->a ->s)-> n
step cost(s, a, s’) -> n

Tree search

function TREE SEARCH(problem):
	frontier = {[initial]}
	loop:
		if frontier is empty: return FAIL
		path = remove.choice(frontier)
		s = path.end
		if s is a goal: return path
		for a in actions:
			add [path + a -> result(s, a)]
			to frontier

Terminology for AI

1.Fully versus partially observablez

-perception action cycle
Agent, State
(sensors, actuators)

2.Deterministic versus stochastic

3.Discrete versus continuous

4.Benign(no objective) versus adversarial(such as chess, games)

for example:
robot car -> partially observable, stochastic, continuous, adverial