WEEK 4 BN PART 2

Question 1

Do we need to draw arcs between all nodes to make states dependent on each other?

Answer 1

NO

Question 2

Temporal Representation

Answer 2

How a state changes over time, using a series of time slices.
Assume that CPT for a given node are going to be the same across each time slice. You can specify different CPT but that is harder.

Question 3

Dynamic Bayesian Network (DBN)

Answer 3

Consists of 3 aspects: 
1. Transition Model- how states evolve over time 
2. Observation Model 
3. Belief at time t0
Example: Knowledge vs Solution 
Bottom/ Blue Nodes is the solution
K above is the knowledge 
Ex. Blue node observed, you know Kasia has knowledge because of the work she has produced, updating your knowledge at each time slice. 

Don’t consider grandparents: only considering the one parent before.
Modelling forgetting??

Question 4

Probability Theory

Answer 4

Uncertainty comes from incomplete knowledge

Question 5

Utility Theory

Answer 5

Provides a way to represent a reason about an agents preferences.
U is assigned as a value to each outcome expressing the desirability of that outcome for the agent

Question 6

Expected Utility (EU)

Answer 6

Chosing the action that maximizes that expected utility
Probability of each outcome multiply by the utility of each outcome added together to get the expected utility of each decision

MULTIPLY EACH PROBABILITY BY THE UTILITY AND ADD THEM TOGETHER
We pick the decision with the highest utility
EU= Combining probabilities with utilities

Question 7

Utility Function

Answer 7

Higher means a better outcome

Utility function is specific to the agent

Question 8

Decision Network (DN)

Answer 8

Chance Nodes: Random variables, same as in BN.
Decision Nodes: Do not have an associated CPT.

Utility Nodes: Have no children. Values in utility table specify agents preferences.

Question 9

Multi- Attribute Decisions

Answer 9

Having more than one parent affecting the decision

Memorization Type Question

Question 10

Naive Bayes Classifier

Answer 10

Very simple BN lets us obtain the probability of a hypothesis given evidence (attributes)
Can be used for classification
Need to train the network, gives CPT values
Then test using data –> clamp the evidence nodes not the hypothesis node
Ex. Play tennis? –> Outlook, temp, humidity, wind