A compendium containing the course material for MM4 through MM10 can be purchased from Lisbeth Palmelund, room F224, price 100 DKK.
Title: Causal Probabilistic Networks
Literature: Medical Decision Support Systems Based on Causal Probabilistic Networks,
Steen Andreassen, chapter 1 pp. 714, chapter 2 pp. 1525.
Exercise 1:
Calculate P (DçS), from the tables below for P(D) and
P (SçD)
P (SçD) 
D 

Yes 
No 

S 
Yes 
0,90 
0,01 
No 
0,10 
0,99 
P (D) 

Yes 
0,001 
No 
0,999 
Exercise 2: Calculate P (A), P (B), P (AçB) and P (BçA) from table 2.14 below.
Reference: An introduction to Baysian Networks by Finn V. Jensen
Exercise 3: In the graphs below determine which variables are dconnected to A.
Exercise 4:
a) Construct the graph (nodes and links) for a Causal Probabilistic Network (CPN) that represents a model of whether or not a car engine will start. Use the same variables as MM1 exercises“Artificial Intelligence and the design of expert systems (see figure)
b) Produce plausible conditional probability tables for the CPN.
c) Discuss the choice of a priori probabilities for founding nodes (nodes without parents).
d) Find two different ways to represent in the CPN that there may be other reasons why the engine will not turn over.
DECISION SUPPORT SYSTEMS
Minimodule 5
Title:
Propagation of evidence in
Causal Probabilistic Nets
Literature: Steen
Andreassen: Medical Decision Support Systems Based on Causal Probabilistic Networks, pp. 25 –
32.
Exercises:
Exercise 1
1.1
Could Quark (Fig. 2.6) have been triangulated differently?
1.2 If so, how would the cliques have looked?
1.3 Draw the junctiontree.
Exercise 2
2.1
Define causal and diagnostic reasoning.
2.2 What does Bayes theorem state about the relation between causal and
diagnostic reasoning.
Exercise 3
3.1
Use Hugin to build the “Car won’t start” example. Assume that the prior
probabilities:
P(tankfull = “no”) = 0.5%
P(carb_OK = “no”) = 0.1%
P(start_motorOK = “no”) = 0.1%
P(batteryOK = “no”) = 1.0%
P(spark_plugsOK = "no") = 1.0%
Also assume that there is a 1% probability that the engine will not turn over
for “other reasons”.
3.2 What is the probability that your car will not start?
3.3 One morning you find that your car will not start. How
does that affect your belief in: a) tankfull, b) carb_OK, c)
start_motorOK, d) batteryOK and e) lights_work ?
3.4 You note that the lights work. How does that affect your
belief in:
a) tankfull , b) carb_OK, c) start_motorOK, d) batteryOK and e)
lights_work ?
3.5 Next you note that the engine actually turns when you
turn the starter key. How does that affect your belief in:
a) tankfull , b) carb_OK, c) start_motorOK, d) batteryOK and e)
lights_work ?
3.6 You have some gas in your spare tank. You put it into
the cars gas tank. How does that affect your belief in:
a) tankfull , b) carb_OK, c) start_motorOK, d) batteryOK and
e) lights_work ?
3.7 You take out the spark plugs and inspect them. They look
just fine. How does that affect your belief in a) tankfull , b) carb_OK,
c) start_motorOK, d) batteryOK and e) lights_work ?
DECISION SUPPORT SYSTEMS
Title: Knowledge Acquisition and Representation
Literature:
Steen Andreassen: Medical Decision Support Systems Based on Causal Probabilistic Networks, pp. 33 – 59.
1. Give some reasons why the introduction
of additional nodes in a CPN can give substantial savings in the number of
conditional probabilities, as for example in Fig. 3.9.
2. How can the knowledge in
MUNIN be checked?
3.a Verify (without Hugin) that the
initialisation of Quark (Fig. 2.11) is correct.
3.b What is the marginal probability of FLU and THROAT
INF after propagation of FEVER = yes?
3.c Enter and propagate the evidence SORE THROAT = no,
under the condition that FEVER = yes. What is the marginal probability of FLU
and THROAT INF after propagation of FEVER = yes and SORE THROAT = no?
Title: Decisions in CPNs (Decision theory)
Literature:
Steen Andreassen: Medical Decision Support Systems Based on Causal Probabilistic Networks, pp. 123 – 138.
The CPN drawn below represents a situation, where a person may have a Disease. A Treatment is available for that Disease, and the Treatment will influence the probability that the disease is present after the Treatment (DiseaseAfter). The Treatment also has a SideEffect. A Test can be performed. When the Test is performed, a TestResult becomes available. The TestResult depends on whether or not the patient has the Disease. The Test has some complications, that may or may not happen. Utilities are associated with SideEffects, DiseaseAfter and with Complications. Download Hugin Lite (http://www.hugin.com/Products_Services/Products/Demo/Lite/). Download the CPN and make yourself familiar with the conditional probabilities and with the utilities in the CPN.
Exercises:
1) Set Test= no. Use Hugin to determine the utility of Treatment = yes and Treatment = no, respectively
2) How large is the contribution of SideEffects to the utility, in the situation Test= no. and Treatment = yes.
3) Now we will explore if it is worthwhile to perform the Test, before deciding on Treatment.
a. To explore this set Test = yes, and read the probability of TestResult.
b. Further, set TestResult to yes and read the utilities of Treatment = yes and no.
c. Set TestResult to no and read the utilities of Treatment = yes and no.
d. Use the results collected above to construct a decision tree similar to the one in Fig. 6.6 in Medical Decision Support Systems Based on Causal Probabilistic Networks.
e. Is it worthwhile to perform the Test, before deciding on Treatment?
Title: An example of a decision support system: The Diabetes Insulin Advisory System (DIAS)
Literature:
Steen Andreassen: Medical Decision Support Systems Based on Causal Probabilistic Networks, pp. 5968, 111, 140144.
Orsini Federici et al. Evaluation of a decision support system (DIAS) for the adjustment of the insulin doses in type 1 diabetic patients.
A demonstration of DIAS will be given in the lecture room.
Title: Learning in CPN’s
Literature:
A Tutorial on Learning With Bayesian Networks, David Heckerman, Chapter 1 – 6 pp. 123.
A note on updating of conditional probabilities in Bayesian networks, Kristian G. Olesen and Steen Andreassen, pp. 17.
Title: An example of a decision support system: Antibiotic treatment of
severe infections  TREAT
Literature: Steen Andreassen et al. 2004: A probabilistic network for fusion of data
and knowledge in clinical microbiology
Exercises: A demonstration of Treat will be given in the lecture room.