cardinality & sorting networks. cardinality constraint appears in many practical problems:...

38
Cardinality & Sorting Networks

Post on 22-Dec-2015

224 views

Category:

Documents


0 download

TRANSCRIPT

  • Slide 1
  • Cardinality & Sorting Networks
  • Slide 2
  • Cardinality constraint Appears in many practical problems: scheduling, timetabling etc. Also takes place in the Max-Sat problem. Has the form: where the symbol is one of, variables are boolean.
  • Slide 3
  • Cardinality constraint As an example, the car sequencing problem (from last week) defines the number of cars to be manufactured, per model. If we agree that a Boolean variable is true iff the ith car of the resulting sequence belongs to some model M, then the constraint enforces the existence of k cars from that model to be in the sequence (where the sequence size is n).
  • Slide 4
  • Cardinality constraint 01 5 2 43 3 4 2 5 1 2 3 4 5 6 7 8 9 10 0 0 1 0 0 0 0 0 0 1 Each Boolean variable below the sequence is 1 iff a yellow model appears in the corresponding position. car sequence instance:
  • Slide 5
  • Cardinality constraint In the Max-Sat problem we seek an assignment which satisfies the maximal number of clauses in a propositional formula One approach to solve the problem is to add fresh blocking variables to each clause, giving. We then search for a minimal k such that is satisfied.
  • Slide 6
  • Cardinality constraint There are several ways to encode a cardinality constraint as propositional formula. In this presentation we consider encoding which is based on sorting networks.
  • Slide 7
  • Sorting networks A (Boolean) sorting network is a circuit that receives n Boolean inputs and permutes them to obtain the sorted outputs. The network is composed of wired comparators. Each comparator has two inputs and two outputs. The upper output,, receives the maximal input value, where the lower output,, receives the minimal input value. The network computation is performed in parallel.
  • Slide 8
  • The 0-1 principle Theorem: If a sorting network sorts every sequence of 0's and 1's, then it sorts every arbitrary sequence of values.
  • Slide 9
  • Sorting networks 0110 10 10 1 1 0 0 0 0 1 1 1 0 0 1
  • Slide 10
  • 0101 Is it a sorting network ? 1010 No its not!
  • Slide 11
  • Sorting network encoding A comparator can be encoded into CNF by the following 6 clauses: Given a set of comparators of any sorting network, it is straightforward to construct a CNF which is a conjunction of the encoded comparators.
  • Slide 12
  • Sorting network encoding For any assignment on the sequence input, a complete satisfying assignment of the CNF, yields a sorted output assignment. We can use this property to check whether a given network is a sorting network. This could be done by adding constraint which is satisfied only if the output is not sorted. A satisfying assignment for the CNF and the above constraint means that the network does not sort.
  • Slide 13
  • Cardinality constraint encoding The cardinality constraint, over the input sequence is obtained by setting the kth largest output to 0. It implies that all outputs from position k are zero. Hence, there are less than k ones amongst the input values. If the propositional formula describes the relation between the input and the output of a sorting network, we search a satisfying assignment for
  • Slide 14
  • Sorter(8) Cardinality constraint encoding
  • Slide 15
  • Sorter(8) Cardinality constraint encoding
  • Slide 16
  • Sorter(8) 0 0 0 0 0 Cardinality constraint encoding
  • Slide 17
  • The odd-even sorting network Batchers odd-even network is a classic sorting network. It was devised back in 68. It uses the divide and conquer design. The approach is similar to the merge-sort algorithm: for sorting a list of 2n inputs, partition the list into two sub lists, with n values each. Recursively sort these two lists, and finally merge them. Networks size:, depth:.
  • Slide 18
  • The odd-even sorting network Sorter(2n)
  • Slide 19
  • Sorter(n) Merger(2n) The odd-even sorting network
  • Slide 20
  • Sorter(n/2) Merger(2n) Sorter(n/2) Merger(n) The odd-even sorting network
  • Slide 21
  • The odd-even merger The odd even merger uses the divide and conquer design, as well. It is assumed that its two input sequences: and are already sorted. The procedure divides each input into odd and even sequences, namely: and for the as. and for the bs. The even and the odd sequences are merged recursively. The merged sequences are combined by comparing each in the result outcome
  • Slide 22
  • The odd-even merger Merger(n)
  • Slide 23
  • The odd-even merger Merger(n/2)
  • Slide 24
  • The odd-even merger Merger(n/2) 111110 1 0 111 1
  • Slide 25
  • The odd-even merger Merger(n/2) 1100 0110
  • Slide 26
  • The odd-even merger Merger(n/2) 0110 111 1 1110 111 0
  • Slide 27
  • Unit Propagation If a set of propositional clauses contains a unit clause l, the other clauses are simplified. This process is called unit propagation. For a comparator : If the cardinality constraint parameter k is known a priory, we can use the above process in order to design a simplified network.
  • Slide 28
  • Odd-Even sorting network properties If there are p input variables that are set to 1, then by unit propagation the first p output variables are set to 1, as well. If there are p input variables that are set to 1, and the output variable in the p+1 position is set to 0, then by unit propagation the reminder of the input variables are set to 0.
  • Slide 29
  • Sorter(8) Odd-Even sorting network properties
  • Slide 30
  • Sorter(8) Odd-Even sorting network properties 1 1
  • Slide 31
  • Sorter(8) Odd-Even sorting network properties
  • Slide 32
  • Sorter(8) Odd-Even sorting network properties
  • Slide 33
  • The pairwise sorting network Devised in 94 by Ian Parberry Takes a different form of the odd-even sorting network.
  • Slide 34
  • The pairwise sorting network Has the same size, depth and properties of the odd- even sorting network. Better unit propagation. The simplified CNF of the corresponding cardinality network is significantly smaller. Simple recursive definition of the corresponding cardinality network when k is known a priory. Experiments have shown better performance for structured problems.
  • Slide 35
  • Cardinality (0,1) matrix constraint A cardinality (0,1) matrix constraint is a constraint C defined on a Matrix M=x[i,j] of boolean variables. Every row i is associated with two positive integers lr[i] and ur[i], such that lr[i] ur[i]. Every column j is associated with two positive integers lc[j] and uc[j], such that lc[j] uc[j]. Each row and column of variables in the matrix M is constrained by:
  • Slide 36
  • Cardinality (0,1) matrix constraint The following Boolean martix is a solution instance for the cardinality (0,1) matrix, where each sum of row and column is bounded by 3 from bottom, and by 4 from top.
  • Slide 37
  • Exercise You are required to solve n cardinality (0,1) matrix constraint instances. The boolean matrix size for each problem instance is nn, choose significant size n. For each problem instance i [1..n] the sum of each row and column is exactly i. Use a sorting network for the cardinality encoding. Measure the solving time of each problem instance. Explain your results.
  • Slide 38