Massachusetts Institute of Technology

16.412J/6.834J Intelligent Embedded Systems

Problem Set #3 Solutions

Problem 1 – Propositional Sentences and Models

(This is problem 7.5 in the new Russell and Norvig text.)

Consider a vocabulary with only four propositions: A, B, C, and D. How many

models are there for the following sentences?

a.  A  B    C  D 
b. A  B
c. A  B  C

a. The truth table showing which models are true is as follows:

A B C D  A  B   C  D
F F F F F
F F F T F
F F T F F
F F T T T
F T F F F
F T F T F
F T T F F
F T T T T
T F F F F
T F F T F
T F T F F
T F T T T
T T F F T
T T F T T
T T T F T
T T T T T
Therefore, there are 7 models for which proposition a. is true.
b. The truth table showing which models are true is as follows:

A B C D A B
F F F F F
F F F T F
F F T F F
F F T T F
F T F F T
F T F T T
F T T F T
F T T T T
T F F F T
T F F T T
T F T F T
T F T T T
T T F F T
T T F T T
T T T F T
T T T T T

Therefore, there are 12 models for which proposition b. is true.

c. First, note that the truth table for “iff” is as follows:

A B A B
F F T
F T F
T F F
T T T
Therefore, the truth table showing which models are true is as follows:

A B C D A BC
F F F F T
F F F T T
F F T F F
F F T T F
F T F F F
F T F T F
F T T F F
F T T T F
T F F F F
T F F T F
T F T F F
T F T T F
T T F F F
T T F T F
T T T F T
T T T T T

Therefore, there are 4 models for which proposition c. is true.

Problem 2 – Whodunit; Propositional Encoding and

Proof
Inspector Poirot has been called in to investigate a murder at Colonel Mustard’s
mansion. After some questioning, Poirot has ascertained the following:

Mrs. Green was murdered in the study.

The murderer was in a room adjacent to the study just before the murder.
The butler had been serving tea to Colonel Mustard in the dining room, but had
left just before the murder.
Miss Plum was in a room adjacent to the study just before the murder.
The dining room is adjacent to the study.
There are three other rooms on the same floor as the study, two of which are
adjacent to the study.
Miss Plum, the butler, and Colonel Mustard were in different rooms just before
the murder.
The butler always leaves the dining room via the kitchen.
Prove that the butler didn’t do it. First, encode the above information into a set of
propositional sentences. Then, generate a proof using rules of inferencing.

Solution

Propositional Encoding

Mustard _ murderer  Mustard _ adjacent R1

Butler _ murderer  Butler _ adjacent R2

Plum _ murderer  Plum _ adjacent R3

Mustard _ dining R4

Dining _ adjacent R5

Mustard _ dining  Dining _ adjacent  Mustard _ adjacent R6

Butler _ left _ dining R7

Plum _ adjacent R8
 Dining _ a djacen t  Kitchen _ a djacen t  Livin
 Dining _ a djacen t  Livin g _ adja cen t   Kitch
 Kitch en _ a djacen t  Livin g _ adja cent   Dinin

R9
 M us ta rd _ d in in g  Plu
 M us ta rd _ d in in g  Bu
 M us ta rd _ l i vi ng  Plu
 M u s ta rd _ l i vi ng  Bu t
 M u s ta rd _ ki t ch en  Bu
  M ust a rd _ ki t ch en  P

R10

Butler _ left _ dining  Butler _ kitchen R11

The following clauses, based on common sense, is also required.

 Mus tard _ d ining   Mus tard _ livin g   Mu
  Mus tard _ dining  Mus tard _ livin g   Mu
  Mus tard _ dining   Mus tard _ liv ing  Mu

R12

 P lum _ din in g  P lum _ living   Plum _ kitch

  Plum _ dining  P lum _ living   Plum _ kitch
  Plum _ dining  Plu m _ livin g  Plum _ kitch

R13
 bu tler _ d in ing   butler _ livin g   bu tler _ ki
  b utler _ din ing  butler _ livin g   bu tler _ ki
  b utler _ din ing   bu tler _ living  bu tler _ ki

R14

Proof

By And-Elimination, R12 is equivalent to

 Mus ta rd _ dining   Mu stard _ livin g   Mus tard _ kitc hen  
  Mus ta rd _ dining  
  Mus ta rd _ dining 

R15

This can be resolved with R4 to get

 Mustard _ dining  Mustard _ living  Mustard _ kitchen R16

Using And-Elimination again, this yields

Mustard _ living
R17
and

Mustard _ kitchen
R18
Resolving these with R10, and using And-Elimination again yields
 Plum _ living  Butler _ kitchen  
 Butler _ living  Plum _ kitchen 

R19

Resolving R5 with R9 and using And-Elimination yields

 Kitchen _ adjacent  Living _ adjacent  
 Living _ adjacent  Kitchen _ adjacent 

R20

Using Modus Ponens with R7 and R11 yields

Butler _ kitchen R21

Resolving with R14 yields

 butler _ dining  butler _ living  butler _ kitchen R22

Using And-Elimination yields

butler _ dining R23

and
butler _ living R24

Resolving R24 with R20 and using And-Elimination yields

Plum _ living R25

If Plum_living and plum_adjacent, then living_adjacent

This means that not kitchen_adjacent

Butler_kitchen and not kitchen_adjacent implies not butler adjacent

Proof.

Problem 3 – Cowboys and Snakes

Consider a game, similar to wumpus world (and especially to minesweeper), that
is played on a rectangular grid. The grid contains one or more cowboys and one
or more snakes. Any square can be occupied by a rattle snake. Any square can
be occupied by a cowboy, but instant death results if a cowboy moves onto a
square occupied by a rattle snake. A Cowboy can’t see where the snakes are,
but he can hear them. He can even hear how many there are in squares directly
or diagonally adjacent to the one he is on.

a. Let Si,j be true iff square [i,j] contains a snake. Suppose that [1,1] is the
lower-left corner square of the grid ([2,1] is the square directly to the left, [1,2] is
the square directly above, etc.). Write the assertion that there are exactly two
snakes adjacent to [1,1] as a sentence involving some Si,j propositions.

The most intuitively straight-forward way to write this is as a disjunction of the

three possible configurations:

 S12  S 22  S 21    S12  S 22  S 21    S12  S 22  S 21  3.1

Unfortunately, this is a disjunction of conjunctions, and CNF form requires a
conjunction of disjunctions. Distribution can be used to convert this. Combining
the first two terms by distribution yields:

 S12  S12    S12  S 22    S12  S 21  

 S 22  S12    S 22  S 22    S 22  S 21  
 S 21  S12    S 21  S 22    S 21  S 21 
3.2

Now, the first term is just S12 , and because this is in CNF form, any clause
containing S12 can be eliminated. Also, tautologies can be eliminated.
Therefore, this whole expression reduces to

S12   S 22  S 21    S 22  S 21  3.3

This makes sense; it expresses an xor relation if S12 is true. Now, combining
this with the third term of 3.1 yields

 S12  S12    S12  S 22  S 21    S12  S 22  S 21  

 S 22  S12    S 22  S 21    S 22  S 22  S 21  
 S 21  S12    S 21  S 22    S 21  S 21  S 22 
3.4

Removing tautologies yields

 S12  S 22  S 21    S12  S 22  S 21  
 S 22  S12    S 22  S 21  
 S 21  S12    S 21  S 22 
3.5

The first term can be removed due to the fourth term, and the last term can be
removed because it is identical to the fourth. This results in

 S12  S 22  S 21  
 S 22  S12    S 22  S 21  
 S 21  S12 
3.6

This makes intuitive sense. The first term ensures that at most two of the three
propositions are true. The last three terms ensure that at least two of the three
propositions are true.
b. Generalize this by explaining how to construct a CNF sentence asserting that
k of n neighbors contain snakes.

Such sentences could be constructed using the distribution procedure of part a.

However, it is easier to use intuition gained from the result of part a. In particular,
such a sentence would have clauses that restrict the minimum to be k, and
another set of clauses that restrict the maximum to be k.

Thus, the clauses that restrict the minimum to be k would be combinations of the
n neighbor propositions taken (n – k + 1) at a time. Similarly, the clauses that
restrict the maximum to be k would be combinations of the n neighbor negation
propositions taken (n – (n – k) + 1) = (k + 1) at a time.

In part a, n = 3 and k = 2. This yields clauses of the form given in 3.6. If n = 4

and k = 2, and the neighbor propositions are A, B, C, and D, then the clauses
are:
 A  B  C    A  C  D   B  C  D 
 A  B  C    A  C  D    B  C  D 

3.7

c. Suppose that we introduce a new player, an escaped bull, to the game.

Suppose that the bull has been chased by two cowboys into the lower-left square
of the board, as shown in the following diagram.

C1

C2

B S

Thus, the bull is in square [1,1]. Cowboy C1 is in [1,3], and cowboy C2 is in [3,2].
A snake is known to be in [3,1].

The bull is trying to escape, and the cowboys are trying to prevent this. The bull
is hemmed in (by mountains to the south and west), so it cannot move off the
board. It has to try to go to the upper right, and it will be successful if it is able to
reach one of the squares on the top row or right column ([1,3], [2,3], [3,3], [3,2],
or [3,1]).
The bull can only move to squares directly adjacent to the one it is on; it cannot
move diagonally. It cannot move onto a square occupied by a snake or a
cowboy. Assume that the cowboys and snakes cannot move from the squares
they are currently in.

Suppose that cowboy C1 and C2 each have exactly 1 snake adjacent to them.
Write a propositional sentence, in CNF form, that expresses the constraints of
this situation, and expresses the constraints that allow the bull to be unblocked.

A proposition that provides an unblocked path is as follows (Uij means that

square [i,j] is unblocked):

 U 12   U 13   U 22   U 23  U 32  
 U 21   U 31   U 22   U 23  U 32 
3.8

In CNF form, this is

U 12  U 21  
U 21  U13  U 22  U 23    U 21  U 13  U 22  U 32  
U 12  U 31  U 22  U 23    U 12  U 31  U 22  U 32  
3.9
U 13  U 31  U 22  U 23    U 13  U 31  U 22  U 32 

This should be conjoined with a set of clauses that express that a square is
unblocked if it doesn’t have a snake or cowboy:

 S11  C11  U 11  
 S12  C12  U 12  
 S13  C13  U 13  
 S 21  C 21  U 21  
 S 22  C 22  U 22  
 S 23  C 23  U 23  
 S 31  C31  U 31  
 S 32  C32  U 32  
 S 33  C33  U 33 
3.10

(Sij indicates presence of a snake, Cij indicates presence of a cowboy).

This should be conjoined with the known positions of cowboys and snakes:

C13  C32  S 31 3.11

Finally, this should be conjoined with a constraint representing 1 adjacent snake
for cowboy C1:
 S 23  S 22  S12  
 S 23  S 22    S 23  S12    S 22  S12  3.12

d. Implement WALKSAT and use it to find models that satisfy the propositional
sentence written in part c. Use the programming language of your choice for this
implementation. Please include comments in your code. Feel free to discuss the
algorithms with other students; however, please do not share code.

Provide results of running your code.

Email code (both source and executable) to the TA as a zip file. Also, include
any data files needed to reproduce results.

e. Suppose that the situation changes and that cowboy C1 now has 2 snakes
adjacent to him. Write a propositional sentence expressing this new situation.
Run your WALKSAT implementation using this new sentence. Explain what
happens.

The WALKSAT algorithm does not finish because the problem is infeasible.
Problem 4 – Model-based Diagnosis
Consider the mixing process shown in the following diagram.

A B

filterA filterB

f1 f2

LS1

Mixing
Tank

f3
v1

Two fluids, A and B, are passed through filters, filterA and filterB, and then flow
into the mixing tank. The mixture flows out via valve v1 which is always open.
The level sensor, LS1, indicates whether the level in the mixing tank is within its
nominal range.

More precisely, here are the CNF clauses for this model.

G  Tank   F1  F 2  F 3  LN 4.1

G  Tank   F1  LN 4.2
G  Tank   F 2  LN 4.3
G  Tank   F 3  LN 4.4
G  FilterA   F1 4.5
G  FilterB   F 2 4.6
G V 1  F 3 4.7
G  LS1  LN  LS1NOM 4.8
G  LS1  LN  LS1NOM 4.9

Here, F1, F2, and F3 are the flows, and LN is the proposition that the tank level is
in its nominal range. LS1NOM is the proposition that the tank level sensor
indicates that the tank is in its nominal range. G(<component>) is a proposition
used to indicate that the component is good.

Now, suppose you make the observation that the level sensor indicates that the
tank level is not in its nominal range. This is expressed as

LS1NOM
4.10

A. For this part, assume there is a single fault, and use constraint suspension to
find a diagnosis.

Generate the list of suspected faulty components. To do this, assume that all
components are working correctly. Perform unit propagation using the
observable stated above and the clauses describing the behavior of the full
model.

List the set of propositions assigned true, false, or both using unit propagation.

List conflicts derived from unit propagation. A conflict arises when a proposition
is deduced to be both true and false. The conflict is the list of component modes
used to derive this inconsistency. This is your list of suspects.

Assuming components are working correctly, resolving 4.10 with 4.9 yields

LN 4.11

Again, assuming components are working correctly, 4.5 – 4.7 yields

F1 4.12
F2 4.13
F3 4.14

Resolving these with 4.1 yields

LN 4.15
which is in conflict with 4.11. Since every component appears in the constraints
that lead to this conflict, all components are suspects!

This makes intuitive sense. If the valve is stuck closed, for example, the tank
level will be too high. If the filters are clogged, the tank level will be too low. If
the tank has a leak, the tank level will be too low. Or, everything might be OK
except that the level sensor is faulty.

B. Test each suspect for consistency. For each suspect, suspend its
constraints, by removing the clauses that model its behavior. Then apply
WALKSAT to test consistency (you may do this manually, or you may use the
program you wrote in problem 3). List each suspect that is consistent. If a
suspect is consistent, then list a satisfying assignment for that suspect.

Suspending constraints for any of the components eliminates the conflict. Thus,
each component being not good is a possible diagnosis.

C. Let’s assume now that flow sensors are added to measure flows F1, F2, and
F3.

The following clauses are added to the model to represent these sensors.
G  FS1  F1  FS1FLOW
G  FS1  F1  FS1FLOW
G  FS 2   F 2  FS 2 FLOW
G  FS 2   F 2  FS 2 FLOW
G  FS 3  F 3  FS 3FLOW
G  FS 3  F 3  FS 3FLOW

Also, we will now consider multiple faults. We will use conflict directed A* search
to generate the most likely diagnoses.

Please refer to
“Conflict-directed A* and its Role in Model-based Embedded Systems”
(Williams and Ragno)
for details on this algorithm. In particular, refer to pgs. 12 – 24.

With this algorithm, an assignment of modes to components is called a

candidate. A candidate that is consistent with the model is called a diagnosis.
The problem is that, often, a large percentage of candidates are diagnoses. This
is addressed by using a utility function to convert this from a constraint
satisfaction problem into a constrained optimization problem. The utility function
is based on probabilities. Specifically, if we assume that component failures are
independent, then the utility for a particular assignment is
g  y    Pi  yi 
i

where Pi is the PMF denoting the probability that component i is in each of its
modes.

Assume the following probabilities for failure of a type of component:

filter – 0.5%
tank - .1%
flow sensor – 2%
level sensor – 1%
valve – 0.2%

Suppose that the observables are now

LS1NOM
FS1FLOW

Use conflict-directed A* to generate the 1st 6 leading candidates. How many of

these are diagnoses?
Comment on search efficiency as defined on pg. 19 of the paper.

The first 6 leading candidates are:

1: G(FilterA), G(FilterB), G(Tank), G(LS1), G(V1), G(FS1), G(FS2), G(FS3)

2: G(FilterA), G(FilterB), G(Tank), G(LS1), G(V1), U(FS1), G(FS2), G(FS3)
3: U(FilterA), G(FilterB), G(Tank), G(LS1), G(V1), G(FS1), G(FS2), G(FS3)
4: G(FilterA), G(FilterB), G(Tank), U(LS1), G(V1), U(FS1), G(FS2), G(FS3)
5: G(FilterA), U(FilterB), G(Tank), G(LS1), G(V1), U(FS1), G(FS2), G(FS3)
6: U(FilterA), G(FilterB), G(Tank), G(LS1), G(V1), U(FS1), G(FS2), G(FS3)

The last four of these are valid diagnoses. Search efficiency is therefore 66%.

D. Comment on how this model could be improved to correspond more closely

to physical reality (hint – think about what it means for tank level to not be
nominal). Discuss briefly how the model would be changed to make this
improvement.
A better model would indicate whether tank level is high or low rather than just
indicating that it is nominal. This would allow for deducing leaks and clogs.
Problem 5 – Rough Draft of Advanced Lecture Slides
Work together with other members of your group to produce a rough draft of
slides for your advanced lecture.