Satisfiability and SAT Solvers CS 270 Math Foundations of CS Jeremy Johnson
2 Conjunctive Normal Form s x 0 x 1 f
Satisfiability A formula is satisfiable if there is an assignment to the variables that make the formula true A formula is unsatisfiable if all assignments to variables eval to false A formula is falsifiable if there is an assignment to the variables that make the formula false A formula is valid if all assignments to variables eval to true (a valid formula is a theorem or tautology)
Satisfiability Checking to see if a formula f is satisfiable can be done by searching a truth table for a true entry Exponential in the number of variables Does not appear to be a polynomial time algorithm (satisfiability is NP-complete) There are efficient satisfiability checkers that work well on many practical problems Checking whether f is satisfiable can be done by checking if f is not valid An assignment that evaluates to false provides a counter example to validity
DNF vs CNF It is easy to determine if a boolean expression in DNF is satisfiable but difficult to determine if it is valid It is easy to determine if a boolean expression in CNF is valid but difficult to determine if it is satisfiable It is possible to convert any boolean expression to DNF or CNF; however, there can be exponential blowup
SAT Solvers Input expected in CNF Using DIMACS format One clause per line delimited by 0 Variables encoded by integers, not variable encoded by negating integer We will use MiniSAT (minisat.se)
MiniSAT Example (x1 | -x5 | x4) & (-x1 | x5 | x3 | x4) & (-x3 | x4). DIMACS format (c = comment, “p cnf” = SAT problem in CNF) c SAT problem in CNF with 5 variables and 3 clauses p cnf
MiniSAT Example (x1 | -x5 | x4) & (-x1 | x5 | x3 | x4) & (-x3 | x4). This is MiniSat 2.0 beta ============================[ Problem Statistics ]================== | | | Number of variables: 5 | | Number of clauses: 3 | | Parsing time: 0.00 s | …. SATISFIABLE v
Avionics Application Aircraft controlled by (real time) software applications (navigation, control, obstacle detection, obstacle avoidance …) Applications run on computers in different cabinets 500 apps 20 cabinets Apps 1, 2 and 3 must run in separate cabinets Problem: Find assignment of apps to cabinets that satisfies constraints
Corresponding SAT problem
Constaints in CNF
DIMACS Format
Avionics Example
p cnf c clauses for valid map forall a exists c AC^c_a
Avionics Example c constaints ~AC^c_1 + ~AC^c_2 and ~AC^c_1 + ~AC^c_ c constraint ~AC^c_2 + ~AC^c_
Avionics Example Programs]$./MiniSat_v1.14_linux aircraft assignment ==================================[MINISAT]=================================== | Conflicts | ORIGINAL | LEARNT | Progress | | | Clauses Literals | Limit Clauses Literals Lit/Cl | | ============================================================================== | 0 | | nan | % | ============================================================================== restarts : 1 conflicts : 0 (nan /sec) decisions : 39 (inf /sec) propagations : 50 (inf /sec) conflict literals : 0 ( nan % deleted) Memory used : 1.67 MB CPU time : 0 s SATISFIABLE
Avionics Assignment SAT True indicator variables: 3 = 5*0 + 3 => AC(1,3) 7 = 5*1 + 2 => AC(2,2) 11 = 5*2 + 1 => AC(3,1) 16 = 5*3+1 => AC(4,1) 21 = 5*4+1 => AC(5,1) 26 = 5*5=1 => AC(6,1) 31 = 5*6+1 => AC(7,1) 36 = 5*7+1 => AC(8,1) 41 = 5*8 + 1 => AC(9,1) 46 = 5*9+1 => AC(10,1)
N-Queens Problem Given an N x N chess board Find a placement of N queens such that no two queens can take each other
N Queens
Backtrack
N Queens
Backtrack
N Queens Backtrack
N Queens
Solution Found
Recursive Solution to N-Queens Define Queens(board, current, size) Input: board a size x size chess board with placement of current queens in positions without conflict only using the first current columns Output: true if board is a conflict free placement of size queens if (current = size) then return true for row = 0 to size-1 do position := (row,column+1) if ConflictFree(board,position) Update(board,position) done := Queens(board,column+1,size) if done = true return true return false
N-Queens as a SAT Problem Introduce variables B ij for 0 ≤ i,j < N B ij = T if queen at position (i,j) F otherwise Constraints Exactly one queen per row Row i = B ij, j=0…N-1 Exactly one queen per column Column j = B ij, i=0…N-1 At most one queen on diagonal Diagonal k- = B ij, i-j = k = -N+1…,N-1 Diagonal k+ = B ij, i+j = k = 0…,2N
4-Queens SAT input Exactly one queen in row i B i0 B i1 B i2 B i3 B i0 B i1 B i2 B i3 B i1 B i2 B i3 B i2 B i3
4-Queens SAT input Exactly one queen in column j B 0j B 1j B 2j B 3j B 0j B 1j B 2j B 3j B 1j B 2j B 3j B 2j B 3j
4-Queens SAT input At most one queen in diagonal k- B 20 B 31 … B 00 B 11 B 22 B 33 B 11 B 22 B 33 B 22 B 33 … B 02 B 13
4-Queens SAT input At most one queen in diagonal k+ B 01 B 10 … B 30 B 21 B 12 B 03 B 21 B 12 B 03 B 12 B 03 … B 32 B 23
DPLL Algorithm Tries to incrementally build a satisfying assignment A: V {T,F} (partial assignment) for a formula in CNF A is grown by either Deducing a truth value for a literal Whenever all literals except one are F then the remaining literal must be T (unit propagation) Guessing a truth value Backtrack when guess (leads to inconsistency) is wrong
DPLL Example OperationAssignFormula
DPLL Example OperationAssignFormula Deduce1
DPLL Example OperationAssignFormula Deduce1
DPLL Example OperationAssignFormula Deduce1 Guess
DPLL Example OperationAssignFormula Deduce1 Guess Deduce Inconsistency
DPLL Example OperationAssignFormula Deduce 11 Guess 3 Deduce 4 Undo 3 Backtrack
DPLL Example OperationAssignFormula Deduce 11 Guess 3 Deduce 4 Undo 3 Assignment found