AIFA: Reasoning Under

Uncertainty
12/03/2024

Koustav Rudra
 Probability Distributions
• A probability distribution is a description of how likely a random variable is to take on each of its
possible states
• Notation: 𝑃(𝑋) is the probability distribution over the random variable 𝑋
• Associate a probability with each value

• Unobserved random variables have distributions

• A distribution is a TABLE of probabilities of values
Axioms of Probability
• The probability of an event 𝐴 in the given sample space 𝒮, denoted as 𝑃 𝐴 , must satisfy the following
properties:

• Non-negativity
• For any event A ∈ 𝒮, 𝑃(𝐴) ≥ 0

• All possible outcomes

• Probability of the entire sample space is 1, 𝑃(𝒮) = 1

• Additivity of disjoint events

• For all events A1, 𝐴2∈ 𝒮 that are mutually exclusive (A1 ∩ 𝐴2 = ∅), the probability that both events
happen is equal to the sum of their individual probabilities, 𝑃(A1 ∨ A2) = 𝑃(𝐴1) +𝑃(A2)
Joint Distributions
• A joint distribution over a set of random variables: 𝑋1, 𝑋2, … , 𝑋𝑛

• Specifies a real number for each assignment (or outcome):

• 𝑃(𝑋1 = 𝑥1, 𝑋2 = 𝑥2, … , 𝑋𝑛 = 𝑥𝑛)
• 𝑃(𝑥1, 𝑥2, … , 𝑥𝑛) T W P
Hot Sun 0.4
• Must satisfy Hot Run 0.1
• 𝑃(𝑥1, 𝑥2, … , 𝑥𝑛) ≥ 0
• ∑𝑥1,𝑥2,…,𝑥𝑛 P(𝑥1,𝑥2,…,𝑥𝑛)=1 Cold Sun 0.2
Cold Run 0.3
• Size of distribution if n variables with domain sizes d?
Probabilistic Models
• A probabilistic model is a joint distribution over a set of random variables

• Probabilistic models:
T W P
• Random variables with domains
• Assignments are called outcomes Hot Sun 0.4
• Joint distributions: say whether
Hot Run 0.1
• assignments (outcomes) are likely
• Normalized: sum to 1.0 Cold Sun 0.2
• Ideally: only certain variables directly interact Cold Run 0.3
Events
• An event is a set E of outcomes
• 𝑃 𝐸 = ∑𝑥1,𝑥2,…,𝑥𝑛∈" P(𝑥1,𝑥2,…,𝑥𝑛)
T W P
• From a joint distribution, we can calculate the probability of any event Hot Sun 0.4
• Probability that it’s hot AND sunny?
Hot Run 0.1
• Probability that it’s hot?
• Probability that it’s hot OR sunny? Cold Sun 0.2
Cold Run 0.3
Marginal Probability Distribution

• Marginal probability distribution is the probability distribution of a single variable

• It is calculated based on the joint probability distribution 𝑃(𝑋,𝑌) using the sum rule:

• 𝑃 𝑋 = 𝑥 = ∑# 𝑃(𝑋 = 𝑥, 𝑌 = 𝑦)
Marginal Distributions
• Marginal distributions are sub-tables which eliminate variables
• Marginalization (summing out): Combine collapsed rows by adding

T P
T W P 𝑃 𝑡 = $ 𝑃(𝑡, 𝑠) = 𝑃(𝑇)
! Hot 0.5
Hot Sun 0.4 Cold 0.5

Hot Run 0.1

𝑃 𝑠 = $ 𝑃(𝑡, 𝑠) = 𝑃(𝑊) W P
Cold Sun 0.2 " Sun 0.6
Cold Run 0.3 Rain 0.4

𝑃 𝑋1 = 𝑥1 = $ 𝑃(𝑋1 = 𝑥1, 𝑋2 = 𝑥2)

#$
Conditional Probabilities
$(&,()
• 𝑃 𝑎𝑏 = T W P
$(()
$(*+,,-+.)
• 𝑃 𝑊=𝑠𝑇=𝑐 = Hot Sun 0.4
$(-+.)
/.1 Hot Run 0.1
• 𝑃 𝑊=𝑠𝑇=𝑐 =
/.2
Cold Sun 0.2
Cold Run 0.3
Conditional Distributions
• Conditional distributions are probability distributions over some variables given fixed values of others
 Normalization Trick

T W P • SELECT the joint • Normalize the

probabilities 𝑃(𝑐, 𝑊) selection 𝑃(𝑊|𝑇 = 𝑐)
Hot Sun 0.4 matching the
evidence
T W P W P
Hot Run 0.1
Cold Sun 0.2 Sun 0.2
Cold Sun 0.2 Cold Rain 0.3 Rain 0.3
Cold Run 0.3
Product Rule

• Sometimes have conditional distributions but want the joint

• 𝑃 𝑦 𝑃 𝑥 𝑦 = 𝑃(𝑥, 𝑦)
Bayesian Network
Conditional Independence
• Two events 𝐴 and 𝐵 are conditionally independent given another event 𝐶 with 𝑃(𝐶) > 0:
• 𝑃 𝐴 ∧ 𝐵 𝐶 = 𝑃 𝐴|𝐶 𝑃(𝐵|𝐶)

$(3∧5|7)
• 𝑃 𝐴 𝐵∧𝐶 =
$(5|7)
$(3|7)∧$(5|7)
• 𝑃 𝐴 𝐵∧𝐶 =
$(5|7)
• 𝑃 𝐴 𝐵 ∧ 𝐶 = 𝑃(𝐴|𝐶)
Conditional Independence and Chain Rule
Trivial decomposition:
P(Traffic, Rain, Umbrella)=
P(Rain)P(Traffic|Rain)P(Umbrella|Rain,Traffic)

With assumption of conditional independence:

P(Traffic, Rain, Umbrella)=
P(Rain)P(Traffic|Rain)P(Umbrella|Rain)

Bayes nets / graphical models help us express conditional independence assumptions

Conditional Independence
• 𝑃 𝑥8 , … , 𝑥9 = 𝑃 𝑥9 𝑥9:8 , … , 𝑥8 𝑃(𝑥9:8 , … , 𝑥8 )
• 𝑃 𝑥8 , … , 𝑥9 = 𝑃 𝑥9 𝑥9:8 , … , 𝑥8 𝑃 𝑥9:8 𝑥9:1 , … , 𝑥8 … 𝑃 𝑥1 𝑥8 𝑃(𝑥8 )
• 𝑃 𝑥8 , … , 𝑥9 = ∏9;+8 𝑃(𝑥; |𝑥;:8 , … , 𝑥8 )

• The belief network represents conditional independence:

• 𝑃 𝑥; 𝑥; , … , 𝑥8 = 𝑃(𝑥; |𝑃𝑎𝑟𝑒𝑛𝑡𝑠 𝑥; )
Bayes Nets
• A Bayes’ net is an efficient encoding of a probabilistic model of a domain

• Questions we can ask:

• Inference: given a fixed BN, what is P(X |e)?
• Representation: given a BN graph, what kinds of distributions can it encode?
• Modeling: what BN is most appropriate for a given domain?
Bayes Net Semantics

• A set of nodes, one per variable X

• A directed, acyclic graph
• A conditional distribution for each node
• A collection of distributions over X, one for each combination of
parents’ values
• 𝑃(𝑋|𝑎8 , 𝑎1 , … , 𝑎9 )
• CPT: conditional probability table
• Description of a noisy “causal” process

A Bayes net = Topology (graph) + Local Conditional Probabilities

Probabilities in BN
• Bayes’ nets implicitly encode joint distributions
• As a product of local conditional distributions
• 𝑃 𝑥; 𝑥;:8 , 𝑥;:1 , … , 𝑥8 = 𝑃(𝑥; |𝑃𝑎𝑟𝑒𝑛𝑡𝑠 𝑥; )

• 𝑃𝑎𝑟𝑒𝑛𝑡𝑠 𝑥; : minimal set of predecessors of 𝑋𝑖 in the total ordering such that other predecessors
are conditionally independent of 𝑋𝑖 given 𝑃arent(𝑋𝑖)
Bayesian Network
• What is the issue with joint probability distribution?
• Become intractably large as the number of variables grows
• Specifying probabilities for atomic events is really difficult

• How does Bayesian Network help?

• Explore independence and conditional independence relationships among variables
• To greatly reduce number of probabilities to be specified to define full joint distribution
Bayesian Network
• A set of random variables makes up the nodes of the network
• Variables may be discrete or continuous

• A set of directed links or arrows connects pairs of nodes

• Arrows represent probabilistic dependence among variables

• An arrow from X →Y indicates X is parent of Y

• Each node 𝑋% ℎ𝑎𝑠 𝑎 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 𝑃(𝑋% |𝑃𝑎𝑟𝑒𝑛𝑡𝑠 𝑋% )

• Quantifies the effect of the parents on the node

• The graph has no directed cycles (DAG)

Example1: Traffic
Example2: Traffic
Causality?
• When Bayes’ nets reflect the true causal patterns:
• Often simpler (nodes have fewer parents)
• Often easier to think about and to elicit from experts

• BNs need not actually be causal

• Sometimes no causal net exists over the domain (especially if variables are missing)
• End up with arrows that reflect correlation, not causation

• What do the arrows really mean?

• Topology may happen to encode causal structure
• Topology really encodes conditional independence
Example3: Home Alarm Network
• Burglar alarm at home
• Fairly reliable at detecting a burglary
• Responds at times to minor earthquakes

• Two neighbors, on hearing alarm, calls police

• John always calls when he hears the alarm
• But sometimes confuses the telephone ringing with the alarm and calls then too

• Mary likes loud music

• But sometimes misses the alarm altogether
Belief Network
P(B) P(E) Earthquake
Burglary 0.001 0.002

Alarm
A P(J) A P(M)
t 0.90 B E P(A) t 0.70
f 0.05 t t 0.95 f 0.01
t f 0.94
f t 0.29
JohnCalls f f 0.001 MaryCalls
Joint probability distribution
• 𝑃 𝑥! , … , 𝑥" = ∏"#$! 𝑃(𝑥# |𝑃𝑎𝑟𝑒𝑛𝑡𝑠(𝑥# ))

• 𝑃(𝐽 ∧ 𝑀 ∧ 𝐴 ∧ ~𝐵 ∧ ~𝐸)
• 𝑃 𝐽𝐴 ∗
• 𝑃 𝑀𝐴 ∗
• 𝑃 𝐴 ~𝐵 ∧ ~𝐸 ∗
• 𝑃 ~𝐵 ∗
• P(~𝐸)

• 𝑃 𝐽 ∧ 𝑀 ∧ 𝐴 ∧ ~𝐵 ∧ ~𝐸 = 0.9×0.7×0.001×0.999×0.998 = 0.00062

• 𝑃 𝐽 =?
Conditional Independence
• 𝑃 𝐽, 𝑀, 𝐴, 𝐵, 𝐸 = 𝑃 𝐽 𝑀, 𝐴, 𝐵, 𝐸 𝑃(𝑀, 𝐴, 𝐵, 𝐸)

• 𝑃 𝐽, 𝑀, 𝐴, 𝐵, 𝐸 = 𝑃 𝐽 𝐴 𝑃 𝑀 𝐴, 𝐵, 𝐸 𝑃(𝐴, 𝐵, 𝐸)

• 𝑃 𝐽, 𝑀, 𝐴, 𝐵, 𝐸 = 𝑃 𝐽 𝐴 𝑃 𝑀 𝐴 𝑃 𝐴 𝐵, 𝐸 𝑃(𝐵, 𝐸)

• 𝑃 𝐽, 𝑀, 𝐴, 𝐵, 𝐸 = 𝑃 𝐽 𝐴 𝑃 𝑀 𝐴 𝑃 𝐴 𝐵, 𝐸 𝑃(𝐵)𝑃(𝐸)

How does ordering matter?

 Conditional Independence
• Earthquake, Burglary, Alarm, JohnCalls, MaryCalls
• P(E|B,A,J,M)
MaryCalls JohnCalls

May have to define conditional

probability of confusing set of events

Alarm

Burglary Earthquake
 Conditional Independence
• Alarm, Burglary, Earthquake, JohnCalls, MaryCalls

MaryCalls JohnCalls

May have to construct

large probability tables

Earthquake

Burglary Alarm
Incremental Network Construction
• Choose the set of relevant variables 𝑋; , that describe the domain

• Choose an ordering for the variables [important step]

• While there are variables left:

• Pick a variable X and add a node for it
• Set Parents(X) to some minimal set of existing nodes such that the conditional independence
property is satisfied
• Define conditional probability table for X

Why do we construct Bayes Network?

To answer queries related to joint probability distribution

Size of a Bayes Net
• How big is a joint distribution over N • Both give you the power to calculate
Boolean variables?
• 2<
• BNs: Huge space savings!

• How big is an N-node net if nodes have up to

k parents? • Also easier to elicit local CPTs
• 𝑂(𝑁 ∗ 2= )
• Also faster to answer queries (coming)
Thank You