Related
Problem:
Given 100 stones, two players alternate to take stones out. One can take any number from 1 to 15; however, one cannot take any number that was already taken. If in the end of the game, there is k stones left, but 1 through k have all been previously taken, one can take k stones. The one who takes the last stone wins. How can the first player always win?
My Idea
Use recursion (or dynamic programming). Base case 1, where player 1 has a winning strategy.
Reducing: for n stones left, if palyer 1 takes m1 stones, he has to ensure that for all options player 2 has (m2), he has a winning strategy. Thus the problem is reduced to (n - m1 - m2).
Follow Up Question:
If one uses DP, the potential number of tables to be filled is large (2^15), since the available options left depend on the history, which has 2^15 possibilities.
How can you optimize?
Assuming that the set of numbers remaining can be represented as R, the highest number remaining after your selection can be represented by RH and the lowest number remaining can be RL, the trick is to use your second-to-last move to raise the number to <100-RH, but >100-RH-RL. That forces your opponent to take a number that will put you in winning range.
The final range of winning, with the total number that you create with your second-to-last move, is:
N < 100-RH
N > 100-RH-RL
By observation I noted that RH can be as high as 15 and as low as 8. RL can be as low as 1 and as high as 13. From this range I evaluated the equations.
N < 100-[8:15] => N < [92:85]
N > 100-[8:15]-[1:13] => N > [92:85] - [1:13] => N > [91:72]
Other considerations can narrow this gap. RL, for instance, is only 13 in an edge circumstance that always results in a loss for Player A, so the true range is between 72 and 91. There is a similar issue with RH and the low end of it, so the final ranges and calculations are:
N < 100-[9:15] => N < [91:85]
N > 100-[9:15]-[1:12] => N > [91:85] - [1:12] => N > [90:73]
[90:73] < N < [91:85]
Before this, however, the possibilities explode. Remember, this is AFTER you choose your second-to-last number, not before. At this point they are forced to choose a number that will allow you to win.
Note that 90 is not a valid choice to win with, even though it might exist. Thus, the maximum it can be is 89. The real range of N is:
[88:73] < N < [90:85]
It is, however, possible to calculate the range of the number that you're using to put your opponent in a no-win situation. In the situation you find yourself in, the lowest number or the highest number might be the one you chose, so if RHc is the highest number you can pick and RLc is the lowest number you can pick, then
RHc = [9:15]
RLc = [1:12]
With this information, I can begin constructing a relative algorithm starting from the end of the game.
N*p* - RHp - RLp < Np < N*p* - RHp, where p = iteration and *p* = iteration + 1
RHp = [8+p:15]
RLp = [1:13-p]
p = -1 is your winning move
p = 0 is your opponent's helpless move
p = 1 is your set-up move
Np is the sum of that round.
Thus, solving the algorithm for your set-up move, p=1, you get:
N*p* - [9:15] - [1:12] < Np < N*p* - [9:15]
100 <= N*p* <= 114
I'm still working out the math for this, so expect adjustments. If you see an error, please let me know and I'll adjust appropriately.
Here is a simple, brute force Python code:
# stoneCount: number of stones to start the game with
# possibleMoves: which numbers of stones may be removed? (*sorted* list of integers)
# return value: signals if winning can be forced by first player;
# if True, the winning move is attached
def isWinningPosition(stoneCount, possibleMoves):
if stoneCount == 0:
return False
if len(possibleMoves) == 0:
raise ValueError("no moves left")
if stoneCount in possibleMoves or stoneCount < possibleMoves[0]:
return True,stoneCount
for move in possibleMoves:
if move > stoneCount:
break
remainingMoves = [m for m in possibleMoves if m != move]
winning = isWinningPosition(stoneCount - move, remainingMoves)
if winning == False:
return True,move
return False
For the given problem size this function returns in less than 20 seconds on an Intel i7:
>>> isWinningPosition(100, range(1,16))
False
(So the first play cannot force a win in this situation. Whatever move he makes, it will result in a winning position for the second player.)
Of course, there is a lot of room for run time optimization. In the above implementation many situations are reached and recomputed again and again (e.g. when the first play takes one stone and the second player takes two stones this will put the first player into the same situation as when the number of stones taken by each player are reversed). So the first (major) improvement is to memorize already computed situations. Then one could go for more efficient data structures (e.g. encoding the list of possible moves as bit pattern).
There are N sticks placed in a straight line. Bob is planning to take few of these sticks. But whatever number of sticks he is going to take, he will take no two successive sticks.(i.e. if he is taking a stick i, he will not take i-1 and i+1 sticks.)
So given N, we need to calculate how many different set of sticks he could select. He need to take at least stick.
Example : Let N=3 then answer is 4.
The 4 sets are: (1, 3), (1), (2), and (3)
Main problem is that I want solution better than simple recursion. Can their be any formula for it? As am not able to crack it
It's almost identical to Fibonacci. The final solution is actually fibonacci(N)-1, but let's explain it in terms of actual sticks.
To begin with we disregard from the fact that he needs to pick up at least 1 stick. The solution in this case looks as follows:
If N = 0, there is 1 solution (the solution where he picks up 0 sticks)
If N = 1, there are 2 solutions (pick up the stick, or don't)
Otherwise he can choose to either
pick up the first stick and recurse on N-2 (since the second stick needs to be discarded), or
leave the first stick and recurse on N-1
After this computation is finished, we remove 1 from the result to avoid counting the case where he picks up 0 sticks in total.
Final solution in pseudo code:
int numSticks(int N) {
return N == 0 ? 1
: N == 1 ? 2
: numSticks(N-2) + numSticks(N-1);
}
solution = numSticks(X) - 1;
As you can see numSticks is actually Fibonacci, which can be solved efficiently using for instance memoization.
Let the number of sticks taken by Bob be r.
The problem has a bijection to the number of binary vectors with exactly r 1's, and no two adjacent 1's.
This is solveable by first placing the r 1's , and you are left with exactly n-r 0's to place between them and in the sides. However, you must place r-1 0's between the 1's, so you are left with exactly n-r-(r-1) = n-2r+1 "free" 0's.
The number of ways to arrange such vectors is now given as:
(1) = Choose(n-2r+1 + (r+1) -1 , n-2r+1) = Choose(n-r+1, n-2r+1)
Formula (1) is deriving from number of ways of choosing n-2r+1
elements from r+1 distinct possibilities with replacements
Since we solved it for a specific value of r, and you are interested in all r>=1, you need to sum for each 1<=r<=n
So, the solution of the problem is given by the close formula:
(2) = Sum{ Choose(n-r+1, n-2r+1) | for each 1<=r<=n }
Disclaimer:
(A close variant of the problem with fixed r was given as HW in the course I am TAing this semester, main difference is the need to sum the various values of r.
I'm implementing an m,n,k-game, a generalized version of tic-tac-toe, where m is the number of rows, n is the number of columns and k is the number of pieces that a player needs to put in a row to win. I have implemented a check for a win, but I haven't figured out a satisfactory way to check before the board is full of pieces, if no player can win the game. In other words, there might be empty slots on the board, but they cannot be filled in such a way that one player would win.
My question is, how to check this efficiently? The following algorithm is the best that I can think of. It checks for two conditions:
A. Go over all board positions in all 4 directions (top to bottom, right to left, and both diagonal directions). If say k = 5, and 4 (= k-1) consecutive empty slots are found, stop checking and report "no tie". This doesn't take into account for example the following situation:
OX----XO (Example 1)
where a) there are 4 empty consecutive slots (-) somewhere between two X's, b) next it is O's turn, c) there are less than four other empty positions on the board and no player can win by putting pieces to those, and d) it is not possible to win in any other direction than horizontally in the shown slots either. Now we know that it is a tie because O will eventually block the last winning possibility, but erroneously it is not reported yet because there are four consecutive empty slots. That would be ok (but not great). Checking this condition gives a good speed-up at the beginning when the checking algorithm usually finds such a case early, but it gets slower as more pieces are put on the board.
B. If this k-1-consecutive-empty-slots-condition isn't met, the algorithm would check the slots again consecutively in all 4 directions. Suppose we are currently checking from left to right. If at some point an X is encountered and it was preceded by an O or - (empty slot) or a board border, then start counting the number of consecutive X's and empty slots, counting in this first encountered X. If one can count to 5, then one knows it is possible for X to win, and "no tie" is reported. If an O preceded by an X is encountered before 5 consecutive X's, then X cannot win in those 5 slots from left to right starting from where we started counting. For example:
X-XXO (Example 2)
12345
Here we started checking at position 1, counted to 4, and encountered an O. In this case, one would continue from the encountered O in the same way, trying to find 5 consecutive O's or empty slots this time. In another case when counting X's or empty slots, an O preceded by one or more empty slots is encountered, before counting to 5. For example:
X-X-O (Example 3)
12345
In this case we would again continue from the O at position 5, but add to the new counter (of consecutive O's or empty slots) the number of consecutive empty slots that preceded O, here 1, so that we wouldn't miss for example this possible winning position:
X-X-O---X (Example 4)
In this way, in the worst case, one would have to go through all positions 4 times (4 directions, and of course diagonals whose length is less than k can be skipped), giving running time O(mn).
The best way I could think of was doing these two described checks, A and B, in one pass. If the checking algorithm gets through all positions in all directions without reporting "no tie", it reports a tie.
Knowing that you can check a win just by checking in the vicinity of the last piece that was added with running time O(k), I was wondering if there were quicker ways to do an early check for a tie. Doesn't have to be asymptotically quicker. I'm currently keeping the pieces in a two-dimensional array. Is there maybe a data structure that would allow an efficient check? One approach: what is the highest threshold of moves that one can wait the players to make before running any checks for a tie at all?
There are many related questions at Stack Overflow, for example this, but all discussions I could find either only pointed out the obvious tie condition, where the number of moves made is equal to the size of the board (or they checked if the board is full), or handled only the special case where the board is square: m = n. For example this answer claims to do the check for a tie in constant time, but only works when m = n = k. I'm interested in reporting the tie as early as possible and for general m,n and k. Also if the algorithm works for more than two players, that would be neat.
I would reduce the problem of determining a tie to the easier sub-problem:
Can player X still win?
If the answer is 'no' for all players, it is a tie.
To find out whether Player X can win:
fill all blank spaces with virtual 'X'-pieces
are there k 'X'-pieces in a row anywhere?
if there are not --> Player X cannot win. return false.
if there are, find the row of k stones with the least amount of virtual pieces. Count the number of virtual pieces in it.
count the number of moves player X has left, alternating with all other players, until the board is completely full.
if the number of moves is less than the amount of virtual pieces required to win, player X cannot win. return false.
otherwise, player X can still win. return true.
(This algorithm will report a possible win for player X even in cases where the only winning moves for X would have another player win first, but that is ok, since that would not be a tie either)
If, as you said, you can check a win just by checking in the vicinity of the last piece that was added with running time O(k), then I think you can run the above algorithm in O(k * Number_of_empty_spots): Add all virtual X-Piece, note any winning combinations in the vicinity of the added pieces.
The number of empty slots can be large, but as long as there is at least one empty row of size k and player X has still k moves left until the board is filled, you can be sure that player X can still win, so you do not need to run the full check.
This should work with any number of players.
Actually the constant time solution you referenced only works when k = m = n as well. If k is smaller then I don't see any way to adapt the solution to get constant time, basically because there are multiple locations on each row/column/diagonal where a winning consecutive k 0's or 1's may occur.
However, maintaining auxiliary information for each row/column/diagonal can give a speed up. For each row/column/diagonal, you can store the start and end locations for consecutive occurrences of 1's and blanks as possible winning positions for player 1, and similarly store start and end locations of consecutive occurrences of 0's and blanks as possible winning positions for player 0. Note that for a given row/column/diagonal, intervals for player 0 and 1 may overlap if they contain blanks. For each row/column/diagonal, store the intervals for player 1 in sorted order in a self-balancing binary tree (Note you can do this because the intervals are disjoint). Similarly store the intervals for player 0 sorted in a tree. When a player makes a move, find the row/column/diagonals that contain the move location and update the intervals containing the move in the appropriate row column and diagonal trees for the player that did not make the move. For the player that did not make a move, this will split an interval (if it exists) into smaller intervals that you can replace the old interval with and then rebalance the tree. If an interval ever gets to length less than k you can delete it. If a tree ever becomes empty then it is impossible for that player to win in that row/column/diagonal. You can maintain a counter of how many rows/columns/diagonals are impossible to win for each player, and if the counter ever reaches the total number of rows/columns/diagonals for both players then you know you have a tie. The total running time for this is O(log(n/k) + log(m/k)) to check for a tie per move, with O(mn/k) extra space.
You can similarly maintain trees that store consecutive intervals of 1's (without spaces) and update the trees in O(log n + log m) time when a move is made, basically searching for the positions before and after the move in your tree and updating the interval(s) found and merging two intervals if two intervals (before and after) are found. Then you report a win if an interval is ever created/updated and obtains length greater than or equal to k. Similarly for player 0. Total time to check for a win is O(log n + log m) which may be better than O(k) depending on how large k is. Extra space is O(mn).
Let's look at one row (or column or diagonal, it doesn't matter) and count the number of winning lines of length k ("k-line") it's possible to make, at each place in the row, for player X. This solution will keep track of that number over the course of the game, checking fulfillment of the winning condition on each move as well as detecting a tie.
1 2 3... k k k k... 3 2 1
There is one k-line including an X in the leftmost slot, two with the second slot from the left, and so on. If an opposing player, O or otherwise, plays in this row, we can reduce the k-line possibility counts for player X in O(k) time at the time of the move. (The logic for this step should be straightforward after doing an example, needing no other data structure, but any method involving checking each of the k rows of k from will do. Going left to right, only k operations on the counts is needed.) An enemy piece should set the possibility count to -1.
Then, a detectably tied game is one where no cell has a non-zero k-line possibility count for any player. It's easy to check this by keeping track of the index of the first non-zero cell. Maintaining the structure amounts to O(k*players) work on each move. The number of empty slots is less than those filled, for positions that might be tied, so the other answers are good for checking a position in isolation. However, at least for reasonably small numbers of players, this problem is intimately linked with checking the winning condition in the first place, which at minimum you must do, O(k), on every move. Depending on your game engine there may be a better structure that is rich enough to find good moves as well as detect ties. But the possibility counting structure has the nice property that you can check for a win whilst updating it.
If space isn't an issue, I had this idea:
For each player maintain a structure sized (2mn + (1 - k)(m + n) + 2(m - k + 1)(n - k + 1) + 2(sum 1 to (m - k))) where each value represents if one of another player's moves are in one distinct k-sized interval. For example for a 8-8-4 game, one element in the structure could represent row 1, cell 0 to 3; another row 1, cell 1 to 4; etc.
In addition, one variable per player will represent how many elements in their structure are still unset. Only one move is required to set an element, showing that that k-interval can no longer be used to win.
An update of between O(k) and O(4k) time per player seems needed per move. A tie is detected when the number of players exceeds the number of different elements unset.
Using bitsets, the number of bytes needed for each player's structure would be the structure size divided by 8. Notice that when k=m=n, the structure size is 4*k and update time O(4). Less than half a megabyte per player would be needed for a 1000,1000,5 game.
Below is a JavaScript example.
var m = 1000, n = 1000, k = 5, numberOfPlayers = 2
, numberOfHorizontalKIs = m * Math.max(n - k + 1,0)
, numberOfverticalKIs = n * Math.max(m - k + 1,0)
, horizontalVerticalKIArraySize = Math.ceil((numberOfHorizontalKIs + numberOfverticalKIs)/31)
, horizontalAndVerticalKIs = Array(horizontalVerticalKIArraySize)
, numberOfUnsetKIs = horizontalAndVerticalKIs
, upToM = Math.max(0,m - k) // southwest diagonals up to position m
, upToMSum = upToM * (upToM + 1) / 2
, numberOfSouthwestKIs = 2 * upToMSum //sum is multiplied by 2 to account for bottom-right-corner diagonals
+ Math.max(0,n - m + 1) * (m - k + 1)
, diagonalKIArraySize = Math.ceil(2 * numberOfSouthwestKIs/31)
, diagonalKIs = Array(diagonalKIArraySize)
, numberOfUnsetKIs = 2 * numberOfSouthwestKIs + numberOfHorizontalKIs + numberOfverticalKIs
function checkTie(move){
var row = move[0], column = move[1]
//horizontal and vertical
for (var rotate=0; rotate<2; rotate++){
var offset = Math.max(k - n + column, 0)
column -= offset
var index = rotate * numberOfHorizontalKIs + (n - k + 1) * row + column
, count = 0
while (column >= 0 && count < k - offset){
var KIArrayIndex = Math.floor(index / 31)
, bitToSet = 1 << index % 31
if (!(horizontalAndVerticalKIs[KIArrayIndex] & bitToSet)){
horizontalAndVerticalKIs[KIArrayIndex] |= bitToSet
numberOfUnsetKIs--
}
index--
column--
count++
}
//rotate board to log vertical KIs
var mTmp = m
m = n
n = mTmp
row = move[1]
column = move[0]
count = 0
}
//rotate board back
mTmp = m
m = n
n = mTmp
// diagonals
for (var rotate=0; rotate<2; rotate++){
var diagonalTopColumn = column + row
if (diagonalTopColumn < k - 1 || diagonalTopColumn >= n + m - k){
continue
} else {
var offset = Math.max(k - m + row, 0)
row -= offset
column += offset
var dBeforeM = Math.min (diagonalTopColumn - k + 1,m - k)
, dAfterM = n + m - k - diagonalTopColumn
, index = dBeforeM * (dBeforeM + 1) / 2
+ (m - k + 1) * Math.max (Math.min(diagonalTopColumn,n) - m + 1,0)
+ (diagonalTopColumn < n ? 0 : upToMSum - dAfterM * (dAfterM + 1) / 2)
+ (diagonalTopColumn < n ? row : n - 1 - column)
+ rotate * numberOfSouthwestKIs
, count = 0
while (row >= 0 && column < n && count < k - offset){
var KIArrayIndex = Math.floor(index / 31)
, bitToSet = 1 << index % 31
if (!(diagonalKIs[KIArrayIndex] & bitToSet)){
diagonalKIs[KIArrayIndex] |= bitToSet
numberOfUnsetKIs--
}
index--
row--
column++
count++
}
}
//mirror board
column = n - 1 - column
}
if (numberOfUnsetKIs < 1){
return "This player cannot win."
} else {
return "No tie."
}
}
For an application I'm working on, I need to sample a small set of values from a very large data set, on the order of few hundred taken from about 60 trillion (and growing).
Usually I use the technique of seeing if a uniform random number r (0..1) is less than S/T, where S is the number of sample items I still need, and T is the number of items in the set that I haven't considered yet.
However, with this new data, I don't have time to roll the die for each value; there are too many. Instead, I want to generate a random number of entries to "skip", pick the value at the next position, and repeat. That way I can just roll the die and access the list S times. (S is the size of the sample I want.)
I'm hoping there's a straightforward way to do that and create an unbiased sample, along the lines of the S/T test.
To be honest, approximately unbiased would be OK.
This is related (more or less a follow-on) to this persons question:
https://math.stackexchange.com/questions/350041/simple-random-sample-without-replacement
One more side question... the person who showed first showed this to me called it the "mailman's algorithm", but I'm not sure if he was pulling my leg. Is that right?
How about this:
precompute S random numbers from 0 to the size of your dataset.
order your numbers, low to high
store the difference between consecutive numbers as the skip size
iterate though the large dataset using the skip size above.
...The assumption being the order you collect the samples doesn't matter
So I thought about it, and got some help from http://math.stackexchange.com
It boils down to this:
If I picked n items randomly all at once, where would the first one land? That is, min({r_1 ... r_n}). A helpful fellow at math.stackexchange boiled it down to this equation:
x = 1 - (1 - r) ** (1 / n)
that is, the distribution would be 1 minus (1 - r) to the nth power. Then solve for x. Pretty easy.
If I generate a uniform random number and plug it in for r, this is distributed the same as min({r_1 ... r_n}) -- the same way that the lowest item would fall. Voila! I've just simulated picking the first item as if I had randomly selected all n.
So I skip over that many items in the list, pick that one, and then....
Repeat until n is 0
That way, if I have a big database (like Mongo), I can skip, find_one, skip, find_one, etc. Until I have all the items I need.
The only problem I'm having is that my implementation favors the first and last element in the list. But I can live with that.
In Python 2.7, my implementation looks like:
def skip(n):
"""
Produce a random number with the same distribution as
min({r_0, ... r_n}) to see where the next smallest one is
"""
r = numpy.random.uniform()
return 1.0 - (1.0 - r) ** (1.0 / n)
def sample(T, n):
"""
Take n items from a list of size T
"""
t = T
i = 0
while t > 0 and n > 0:
s = skip(n) * (t - n + 1)
i += s
yield int(i) % T
i += 1
t -= s + 1
n -= 1
if __name__ == '__main__':
t = [0] * 100
for c in xrange(10000):
for i in sample(len(t), 10):
t[i] += 1 # this is where we would read value i
pprint.pprint(t)
You have a biased random number generator that produces a 1 with a probability p and 0 with a probability (1-p). You do not know the value of p. Using this make an unbiased random number generator which produces 1 with a probability 0.5 and 0 with a probability 0.5.
Note: this problem is an exercise problem from Introduction to Algorithms by Cormen, Leiserson, Rivest, Stein.(clrs)
The events (p)(1-p) and (1-p)(p) are equiprobable. Taking them as 0 and 1 respectively and discarding the other two pairs of results you get an unbiased random generator.
In code this is done as easy as:
int UnbiasedRandom()
{
int x, y;
do
{
x = BiasedRandom();
y = BiasedRandom();
} while (x == y);
return x;
}
The procedure to produce an unbiased coin from a biased one was first attributed to Von Neumann (a guy who has done enormous work in math and many related fields). The procedure is super simple:
Toss the coin twice.
If the results match, start over, forgetting both results.
If the results differ, use the first result, forgetting the second.
The reason this algorithm works is because the probability of getting HT is p(1-p), which is the same as getting TH (1-p)p. Thus two events are equally likely.
I am also reading this book and it asks the expected running time. The probability that two tosses are not equal is z = 2*p*(1-p), therefore the expected running time is 1/z.
The previous example looks encouraging (after all, if you have a biased coin with a bias of p=0.99, you will need to throw your coin approximately 50 times, which is not that many). So you might think that this is an optimal algorithm. Sadly it is not.
Here is how it compares with the Shannon's theoretical bound (image is taken from this answer). It shows that the algorithm is good, but far from optimal.
You can come up with an improvement if you will consider that HHTT will be discarded by this algorithm, but in fact it has the same probability as TTHH. So you can also stop here and return H. The same is with HHHHTTTT and so on. Using these cases improves the expected running time, but are not making it theoretically optimal.
And in the end - python code:
import random
def biased(p):
# create a biased coin
return 1 if random.random() < p else 0
def unbiased_from_biased(p):
n1, n2 = biased(p), biased(p)
while n1 == n2:
n1, n2 = biased(p), biased(p)
return n1
p = random.random()
print p
tosses = [unbiased_from_biased(p) for i in xrange(1000)]
n_1 = sum(tosses)
n_2 = len(tosses) - n_1
print n_1, n_2
It is pretty self-explanatory, and here is an example result:
0.0973181652114
505 495
As you see, nonetheless we had a bias of 0.097, we got approximately the same number of 1 and 0
The trick attributed to von Neumann of getting two bits at a time, having 01 correspond to 0 and 10 to 1, and repeating for 00 or 11 has already come up. The expected value of bits you need to extract to get a single bit using this method is 1/p(1-p), which can get quite large if p is especially small or large, so it is worthwhile to ask whether the method can be improved, especially since it is evident that it throws away a lot of information (all 00 and 11 cases).
Googling for "von neumann trick biased" produced this paper that develops a better solution for the problem. The idea is that you still take bits two at a time, but if the first two attempts produce only 00s and 11s, you treat a pair of 0s as a single 0 and a pair of 1s as a single 1, and apply von Neumann's trick to these pairs. And if that doesn't work either, keep combining similarly at this level of pairs, and so on.
Further on, the paper develops this into generating multiple unbiased bits from the biased source, essentially using two different ways of generating bits from the bit-pairs, and giving a sketch that this is optimal in the sense that it produces exactly the number of bits that the original sequence had entropy in it.
You need to draw pairs of values from the RNG until you get a sequence of different values, i.e. zero followed by one or one followed by zero. You then take the first value (or last, doesn't matter) of that sequence. (i.e. Repeat as long as the pair drawn is either two zeros or two ones)
The math behind this is simple: a 0 then 1 sequence has the very same probability as a 1 then zero sequence. By always taking the first (or the last) element of this sequence as the output of your new RNG, we get an even chance to get a zero or a one.
Besides the von Neumann procedure given in other answers, there is a whole family of techniques, called randomness extraction (also known as debiasing, deskewing, or whitening), that serve to produce unbiased random bits from random numbers of unknown bias. They include Peres's (1992) iterated von Neumann procedure, as well as an "extractor tree" by Zhou and Bruck (2012). Both methods (and several others) are asymptotically optimal, that is, their efficiency (in terms of output bits per input) approaches the optimal limit as the number of inputs gets large (Pae 2018).
For example, the Peres extractor takes a list of bits (zeros and ones with the same bias) as input and is described as follows:
Create two empty lists named U and V. Then, while two or more bits remain in the input:
If the next two bits are 0/0, append 0 to U and 0 to V.
Otherwise, if those bits are 0/1, append 1 to U, then write a 0.
Otherwise, if those bits are 1/0, append 1 to U, then write a 1.
Otherwise, if those bits are 1/1, append 0 to U and 1 to V.
Run this algorithm recursively, reading from the bits placed in U.
Run this algorithm recursively, reading from the bits placed in V.
This is not to mention procedures that produce unbiased random bits from biased dice or other biased random numbers (not just biased bits); see, e.g., Camion (1974).
I discuss more on randomness extractors in a note on randomness extraction.
REFERENCES:
Peres, Y., "Iterating von Neumann's procedure for extracting random bits", Annals of Statistics 1992,20,1, p. 590-597.
Zhou, H. And Bruck, J., "Streaming algorithms for optimal generation of random bits", arXiv:1209.0730 [cs.IT], 2012.
S. Pae, "Binarization Trees and Random Number Generation", arXiv:1602.06058v2 [cs.DS].
Camion, Paul, "Unbiased die rolling with a biased die", North Carolina State University. Dept. Of Statistics, 1974.
Here's one way, probably not the most efficient. Chew through a bunch of random numbers until you get a sequence of the form [0..., 1, 0..., 1] (where 0... is one or more 0s). Count the number of 0s. If the first sequence is longer, generate a 0, if the second sequence is longer, generate a 1. (If they're the same, try again.)
This is like what HotBits does to generate random numbers from radioactive particle decay:
Since the time of any given decay is random, then the interval between two consecutive decays is also random. What we do, then, is measure a pair of these intervals, and emit a zero or one bit based on the relative length of the two intervals. If we measure the same interval for the two decays, we discard the measurement and try again
HotBits: How It Works
I'm just explaining the already proposed solutions with some running proof. This solution will be unbiased, no matter how many times we change the probability. In a head n tail toss, the exclusivity of consecutive head n tail or tail n head is always unbiased.
import random
def biased_toss(probability):
if random.random() > probability:
return 1
else:
return 0
def unbiased_toss(probability):
x = biased_toss(probability)
y = biased_toss(probability)
while x == y:
x = biased_toss(probability)
y = biased_toss(probability)
else:
return x
# results with contain counts of heads '0' and tails '1'
results = {'0':0, '1':0}
for i in range(1000):
# on every call we are changing the probability
p = random.random()
results[str(unbiased_toss(p))] += 1
# it still return unbiased result
print(results)