Transcript Randomized algorithms

Randomized Algorithms
Pasi Fränti
1.10.2014
Treasure island
Treasure worth 20.000 awaits
?
5000
5000
Map for sale:
3000
5000
?
DAA
expedition
To buy or not to buy
Buy the map:
20000 – 5000 – 3000
= 12.000
Take a change:
20000 – 5000
= 15.000
20000 – 5000 – 5000
= 10.000
To buy or not to buy
Buy the map:
20000 – 5000 – 3000
= 12.000
Take a change:
20000 – 5000
= 15.000
20000 – 5000 – 5000
= 10.000
Expected result:
0.5 ∙ 15000 + 0.5 ∙ 10000 = 12.500
Three type of randomization
1. Las Vegas
-
Output is always correct result
Result is not always found
Probability of success p
2. Monte Carlo
-
Result is always found
Result can be inaccurate (or even false!)
Probability of success p
3. Sherwood
-
Balancing the worst case behavior
Las Vegas
Eating philosophizes
Las Vegas
Input: Binary vector A[1, n]
Output: Index of any 1-bit from A
LV(A, n)
REPEAT
k ← RAND(1, n);
UNTIL A[k]=1;
RETURN k
8-Queens puzzle
INPUT:
OUTPUT:
Eight chess queens and an 8×8 chessboard
Setup where no queens attack each other
8-Queens brute force
Brute force
• Try all positions
• Mark illegal squares
• Backtrack if dead-end
• 114 setups in total
Where next…?
8
5
Random
4
• Select positions randomly
• If dead-end, start over
…
Randomized
• Select k rows randomly
• Rest rows by Brute Force
8-Queens(k)
Pseudo code
FOR i=1 TO k DO
// k Queens randomly
r  Random[1,8];
IF Board[i,r]=TAKEN THEN RETURN Fail;
ELSE ConquerSquare(i,r);
FOR i=k+1 TO 8 DO // Rest by Brute Force
r1; foundNO;
WHILE (r≤8) AND (NOT found) DO
IF Board[i,r] NOT TAKEN THEN
ConquerSquare(i,r); foundYES;
IF NOT found THEN RETURN Fail;
ConquerSquare(i,j)
Board[i,j]  QUEEN;
FOR z=i+1 TO 8 DO
Board[z,j]
Board[z,j-(z-i)]
Board[z,j+(z-i)]
 TAKEN;
 TAKEN;
 TAKEN;
Probability of success
s = processing time in case of success
e = processing time in case of failure
p = probability of success
q = 1-p = probability of failure
t  ps  q  e  t   ps  qe  qt
 t  qt  ps  qe
 t  t  pt  ps  qe
 t  s
q
p
e
Example:
s=e=1, p=1/6

t=1+5/1∙1=6
Experiments with varying k
K
0
1
2
3
4
5
6
7
8
S
114
39.6
22.5
13.5
10.3
9.3
9.1
9
9
E
36.7
15.1
8.8
7.3
7
7
7
T
114
39.6
25.2
29.0
35.1
46.9
53.5
56.0
56.0
P
100%
100%
88%
49%
26%
16%
14%
13%
13%
Fastest
expected
time
Swap-based clustering
O n e cen tro id , b u t
tw o clu sters .
T w o cen tro id s , b u t
o n ly o n e clu ster .
Clustering by Random Swap
P. Fränti and J. Kivijärvi, "Randomised local search algorithm for the clustering
problem", Pattern Analysis and Applications, 3 (4), 358-369, 2000.
RandomSwap(X) → C, P
C ← SelectRandomRepresentatives(X);
P ← OptimalPartition(X, C);
Select
random
REPEAT T times
neighbor
(Cnew, j) ← RandomSwap(X, C);
Pnew ← LocalRepartition(X, Cnew, P, j);
Cnew, Pnew ← Kmeans(X, Cnew, Pnew);
IF f(Cnew, Pnew) < f(C, P) THEN
(C, P) ← Cnew, Pnew;
Accept
RETURN (C, P);
only if it
improves
Clustering by Random Swap
1. Random swap:
c j  xi
j  random (1, M ), i  random (1, N )
2. Re-partition vectors from old cluster:
p i  arg m in d  x i , c k
1 k  M

2
 i
pi  j
3. Create new cluster:
p i  arg m in d  x i , c k 
k  j  k  pi
2
 i  1, N 
Choices for swap
O(M)
clusters
to be
removed

O(M)
clusters
where to
add
=
O(M2)
different
choices
in total
S w ap is m ad e fro m
cen tro id rich area to
cen tro id p oo r area.
Probability for successful Swap
Select a proper centroid for removal:
– M clusters in total: premoval=1/M.
Select a proper new location:
– N choices: padd=1/N
– M of them significantly different: padd=1/M
In total:
– M2 significantly different swaps.
– Probability of each is pswap=1/M2
– Open question: how many of these are good
– Theorem: α are good for add and removal.
Clustering by Random Swap
Probability of not finding good swap:


q   1 
M

2
2




T
Iterated T times
Estimated number of iterations:
2


log q  T  log  1 
2
M

 T 
log q
2


log  1 
2
M








Bounds for the iterations
Upper limit:
T 

ln q
ln 1  α / M
2
2


- ln q
2
α /M
2
 -ln q 
Lower limit similarly; resulting in:
2

M
T    - ln q  2
α





M
α
2
2
Total time complexity
Time complexity of single step (t):
t = O(αN)
Number of iterations needed (T):
2

M
T    - ln q 
2
α

Total time:
T  N , M   - ln q 
M
α
2
2




 - ln  q   NM
  N   
α

2




Monte Carlo
Monte Carlo
Input: A bit vector A[1, n], iterations I
Output: An index of any 1 bit from A
LV(A, n, I)
i ← 0;
DO
k ← RAND(1, n);
i ← i + 1;
WHILE (A[k]≠1 AND i ≤ I)
RETURN k
Monte Carlo
Potential problems to be considered:
• Detecting prime numbers
• Calculating integral of a function
Sherwood
Selection of pivot element
Something about Quicksort and Selection:
• Practical example of re-sorting
• Median selection
1
N-1
1
N-2
1
N-3
…
O(N2)
Simulated dynamic linked list
1. Sorted array
- Search efficient: O(logN)
- Insert and Delete slow: O(N)
2. Dynamically linked list
- Insert and Delete fast: O(1)
- Search inefficient: O(N)
Simulated dynamic linked list
Example
Head
Linked list:
1
2
4
7
5
Simulated by array:
Head=4
i
1
2
3
4
Value
2
4
15 1
5 21 7
Next
2
5
6
7
1
5
6
0
7
3
15
21
Simulated dynamic linked list
Divide-and-conquer with randomization
SEARCH (A, x)
i := A.HEAD;
max := A[i].VALUE;
Value searched
N random
breakpoints
FOR k:=1 TO N DO
Biggest breakpoint ≤ x
j:=RANDOM(1, N);
y:=A[j].VALUE;
IF (max<y) AND (y≤x) THEN
i:=j; max:=y;
Full search from
RETURN LinearSearch(A, x, i);
breakpoint i
Analysis of the search
•
•
•
•
Divide into N segments
Each segment has N/N = N elements
Linear search within one segment.
Expected time complexity = N + N = O(N)
max
N
N
(on average)
search for
Experiment with students
Data (N=100) consists of numbers from 1..100:
1
2
3
4
Select N breaking points:
99
100
Searching for…
42
Empty space for notes