Tutorial 3 Solution
Tutorial 3 Solution
1. Consider the a no-data decision problem with the following loss matrix:
a1 a2 a3 a4 a5
1 3 5 6 4 4
2 4 1 3 6 4
Solution
(a) a1 or a5 ( preferably a1 why ? )
(b) The minimax action ~p p a1 (1 p) a2 ( why? )
with p 4 / 5 .
(c)
Regret a1 a2 a3 a4 a5
1 0 2 3 1 1
2 3 0 2 5 3
2. Determine the minimax loss and minimax regret pure actions for the decision
problem with losses as follows:
a1 a2 a3 a4
1 1 2 0 1
2 2 2 3 0
3 4 3 1 2
Solution
Pure minimax loss action : a 4
Regret Table :
Regret a1 a2 a3 a4
1 0 3 1 2
2 4 0 5 2
3 3 2 0 1
Pure minimax regret action : a 4
0 1 2 3 4 5
a0 0 2 4 6 8 10
a1 5 4 3 2 1 0
Let ~
p p a0 (1 p) a1 .
Then L(0, ~ p ) 5(1 p) , L(1, ~
p ) 2 p 4(1 p) , L(2, p) 4 p 3(1 p)
L(3, ~
p ) 6 p 2(1 p) , L(4, ~p ) 8 p (1 p) , L(5, ~
p ) 10 p
1 2
Answer : mixed minimax action : a 0 a 2
3 3
Leong YK & Wong WY 3
4. Consider a decision problem with two actions and five states of nature. The
loss table is given as follows:
1 2 3 4 5
a1 5 3 2 0 4
a2 1 0 4 3 2
Solution
(a) L(1, ~ p ) 3 p , L(3 , ~
p ) 5 p (1 p) , L( 2 , ~ p ) 2 p 4(1 p)
p ) 3(1 p) , L(5 , ~
L( 4 , ~ p ) 4 p 2(1 p)
4 2 p , 0 p 1 / 2
(b) max L( , ~ p)
1 4 p , 1 / 2 p 1
(c) p 1 / 2
(d) 3
5. A lot of six items is to be accepted ( a1 ) or rejected ( a2 ). The state of nature is
characterized by the number of defective items among the six. Assuming
losses equal to the number of good items in lots that are rejected, and equal
twice the number of defectives in the lots that are accepted. Determined the
minimax pure and mixed actions.
Solution
It is clear that action a2 ( rejection) is the minimax pure action.
Let ~ p p a1 (1 p) a2 be a mixed action. By writing out the
expected losses L( , ~ p ) under various states of nature, it is found that the
minimax mixed action is
~ 1 2
p* a1 a2
3 3
Leong YK & Wong WY 4
Solution
(a) For any integer k , 1 / 2 L( n , a0 ) < L( n , ak ) 1 if n k
Thus no action a k ( k 1 ) dominates a0
(b) For any prior probability function of
(i ) wi , wi 0 , i 1, 2 , ; i wi 1
n
Choose n large enough so that wi 1 / 2 . Then
i 1
1
L( , an ) L( , a0 )
2
(c) No! For any prior distribution , L( , am ) L( , an ) for m n .
1 2 3 4
a1 0 1 3 1
a2 2 0 4 2
a3 3 2 0 0
1 2 3 4
a1 4 0 0 1
a2 6 1 1 2
a3 7 1 3 0
Show that against any prior distribution of , the Bayes action when the
loss table is T1 are the same as the Bayes action when the loss table is T2.
Leong YK & Wong WY 6
9. Consider the no-data decision problem with the following loss table
a1 a2 a3 a4
1 2 5 3 1
2 3 1 2 5
a1 a2 a3 a4
1 5 8 6 4
2 2 0 1 4
Show that the Bayes loss action and the Bayes regret action remain
unchanged as found in part (a) or part (b). Can you explain why this
can happen ?
Leong YK & Wong WY 7
Solution
(a) The Bayes losses of the actions are
L( , a1 ) 2w 3(1 w) , L( , a2 ) 5w (1 w) ,
L( , a3 ) 3w 2(1 w) , L( , a4 ) w 5(1 w)
Thus the Bayes actions against the prior distribution is
a 2 , 0 w 1/ 3
a , 1/ 3 w 1/ 2
a 3
a1 , 1/ 2 w 2 / 3
a4 , 2/3 w 1
(b) The regret table is given by
a1 a2 a3 a4
1 1 4 2 0
2 2 0 1 4
(c) One can verify that the Bayes (loss or regret) action remains unchanged.