1. Bayesian Networks What is the likelihood of X given evidence E? i.e. P(X|E) = ?

2. Issues Representational Power –allows for unknown, uncertain information Inference –Question: What is Probability of X if E is true. –Processing: in general, exponential Acquisition or Learning –network: human input – probabilities: data+ learning

3. Bayesian Network Directed Acyclic Graph Nodes are RV’s Edges denote dependencies Root nodes = nodes without predecessors –prior probability table Non-root nodes –conditional probabilites for all predecessors

4. Bayes Net Example: Structure Burglary Earthquake John Calls Mary Calls Alarm

5. Probabilities Structure dictates what probabilities are needed P(B) =.001 P(-B) =.999 P(E) =.002 P(-E) =.998 etc. P(A|B&E) =.95 P(A|B&-E) =.94 P(A|-B&E) =.29 P(A|-B&-E) =.001 P(JC|A) =.90 P(JC|-A) =.05 P(MC|A) =.70 P(MC|-A) =.01

6. Joint Probability yields all Event = fully specified values for RVs. Prob of event: P(x1,x2,..xn) = P(x1|Parents(X1))*..P(xn|Parents(Xn)) E.g. P(j&m&a&-b&-e) = P(j|a)*P(m|a)*P(a|-b^-e)*P(-b)*P(-e) =.9*.7*.001*.999*..998 =.00062. Do this for all events and then sum as needed. Yields exact probability (assumes table right)

7. Many Questions With 5 boolean variables, joint probability has 2^5 entries, 1 for each event. A query corresponds to the sum of a subset of these entries. Hence 2^2^5 queries possibles. – 4 billion possible queries.

8. Probability Calculation Cost With 5 boolean variables need 2^5 entries. In general 2^n entries with n booleans. For Bayes Net, only need tables for all conditional probabilities and priors. If max k inputs to a node, and n RVs, then need at most n*2^k table entries. Data and computation reduced.

9. Example Computation Method: transform query so matches tables Bold = in a table P(Burglary|Alarm) = P(B|A) = P(A|B)*P(B)/ P(A) P(A|B) = P(A|B,E)*P(E)+P(A|B,~E)*P(~E). Done. Plug and chug.

10. Query Types Diagnostic: from effects to causes –P(Burglary | JohnCalls) Causal: from causes to effects –P(JohnCalls | Burglary) Explaining away: multiple causes for effect –P(Burglary | Alarm and Earthquake) Everything else

11. Approximate Inference Simple Sampling: logic sample Use BayesNetwork as a generative model Eg. generate million or more models, via topological order. Generates examples with appropriate distribution. Now use examples to estimate probabilities.

12. Logic Sampling: simulation Query: P(j&m&a&-b&-e) Topological sort Variables, i.e –Any order that preserves partial order –E.g B, E, A, MC, JC Use prob tables, in order to set values –E.g. p(B = t) =.001 => create a world with B being true once in a thousand times. –Use value of B and E to set A, then MC and JC Yields (1 million).000606 rather than.00062 Generally huge number of simulations for small probabilities.

13. Sampling -> probabilities Generate examples with proper probability density. Use the ordering of the nodes to construct events. Finally count to yield an estimate of the exact probability.

14. Sensitivity Analysis: Confidence of Estimate Given n examples and k are heads. How many examples needed to be 99% certain that k/n is within.01 of the true p. From statistic: Mean = np, Variance = npq For confidence of.99, t = 3.25 (table) 3.25*sqrt(pq/N) N >6,400. But correct probabilities not needed, just correct ordering.

15. Lymphoma Diagnosis PathFinder systems 60 diseases, 130 features I: rule based, performance ok II: used mycin confidence, better III: Do Bayes Net: best IV: Better Bayes Net: (add utility theory) – outperformed experts – solved the combination of expertise problem

16. Summary Bayes nets easier to construct then rule- based expert systems –Years for rules, days for random variables and structure Probability theory provides sound basis for decisions –Correct probabilities still a problem Many diagnostic applications Explanation less clear: use strong influences