Nondeterministic Finite Automata (NFA) : (Include Lecture 5 and 6)
Nondeterministic Finite Automata (NFA) : (Include Lecture 5 and 6)
Chapter 3
Nondeterministic Finite Automata (NFA)
(include lecture 5 and 6)
Nondeterminism
An important notions(or abstraction) in computer science refer to situations in which the next state of a computation is not uniquely determined by the current state. Ex: find a program to compute max(x,y): pr1: case x y => print x; y x => print y endcase; Then which branch will be executed when x = y ? ==> don't care nondeterminism Pr2: do-one-of { {if x < y fail; print x}, {if y < x fail, print y} }. ==>The program is powerful in that it will never choose branches that finally lead to fail -- an unrealistic model. ==> don't know nondeterminism.
Transparency No. 3-2
nondeterminism (cont'd)
a nondeterministic sorting algorithm: nondet-sort(A, n) 1. for i = 1 to n do 2. nondeterministically let k := one of {i, ..., n} ; 3. exchange A[i] and A[k] 4. endfor 5 for i = 1 to n-1 do if A[i] > A[i+1] then fail; 6. return(A). Notes: 1. Step 2 is magic in that it may produce many possible outcomes. However all incorrect results will be filtered out at step 5. 2. The program run in time NTIME O(n) cf: O(n lg n) is required for all sequential machines.
Transparency No. 3-3
nondeterminism (cont'd)
Causes of nodeterminism in real life: incomplete information about the state external forces affecting the course of the computation ex: the behavior of a process in a distributed system Nondeterministic programs cannot be executed directly but can be simulated by real machine. Nondeterminism can be used as a tool for the specification of problem solutions. an important tool in the design of efficient algorithms There are many problems with efficient nondeterministic algorithm but no known efficient deterministic one. the open problem NP = P ? How to make DFAs become nondeterministic ? ==> allow multiple transitions for each state-input-symbol pair ==> modify the transition function d. Transparency No. 3-4
A NFA is a five-tuple N = (Q,S,d,S,F) where everything is the same as in a DFA, except: S Q is a set of starting states, instead of a single state. d is the transition function d: Q x S -> 2Q. For each state p and symbol a, d(p,a) is the set of all states that N is allowed to move from p in one step under input symbol a. diagrammatic notation: p --a--> q a
Note: d(p,a) can be the empty set
The extended transition function D (multi-step version of d) for NFA can be defined analogously to that of DFAs: D: 2QxS* -> 2Q is defined inductively as follows: 1. Basis: D(A, e) = ____ for every set of states A (6.1) 2. Ind. case: D(A, xa) = ____ for every x S* and a S (6.2) Note: Intuitively q D(A,x) means q can be reached from some state A after scanning input string x.
Transparency No. 3-5
Note: Like DFAs, the extended transition function D on a NFA N is uniquely determined by N. pf: left as an exercise. N = (Q,S,d,S,F) : a NFA; x: any string over S; D: the extended transition function of N. 1. x is said to be accepted by N if D(S,x) F {} i.e., x is accepted if there is an accept state q F such that q is reachable from a start state under input string x (i.e., q D(S,x)) 2. The set (or language) accepted by N, denoted L(N), is the set of all strings accepted by N. i.e., L(N) =def {x S* | N accepts x }. 3. Two finite automata (FAs, no matter deterministic or nondeterministic) M and N are said to be equivalent if L(M) = L(N). Transparency No. 3-6
Equivalence of FAs
Note: under such definition, every DFA M = (Q,S,d,s,F) is equivalent to an NFA N = (Q,S,d',{s},F) where d'(p,a) = {d(p,a)} for every state p and input a.
Problem: Does the converse hold as well ? i.e. For every NFA N there is a DFA M s.t. L(M) = L(N). Ans: ____
Ex: Find a NFA accepting A = { x {0,1}* | the fifth symbol from the right is 1 } = {010000, 11111,...}. Sol: 1. (in diagram form) 0,1 1 0,1 0,1 0,1 0,1
2: tabular form:
Note: there are many possible computations on the input string: 010101, some of which reach the (only) final state (accepted or successful computation), some of which do not (fail). Since there exists an accepted computation, by definition, the string is accepted by the machine - 0 - 1 - 0 - 1 - 0 - 1 - (fail)
0,1 1
0,1
0,1
0,1
0,1
- 0 - 1 - 0 - 1 - 0 - 1 - (accept)
Transparency No. 3-9
Lem 6.1: D(A,xy) = D(D(A,x),y). pf: by induciton on |y|: 1. |y| = 0 => D(A,xe) = D(A,x) = D(D(A,x),e) -- (6.1). 2. y = zc => D(A,xzc) = Uq D(A,xz) d(q,c) -- (6.2) = U q D(D(A,x),z) d(q,c) -- ind. hyp. = D(D(A,x),zc) -- (6.2) Lem 6.2 D commutes with set union: i.e., D (Ui I Ai,x) = Ui I D(Ai,x). in particular, D(A,x) = Up A D({p},x) pf: by ind. on |x|. Let B = U i I Ai 1. |x|= 0 => D (U i I Ai, e) = Ui I Ai = Ui I D(Ai, e) -- (6.1) 2. x = ya => D (U i I Ai, ya) = U p D(B,y) d(p,a) -- (6.2) = UpUi I D(Ai,y) d(p,a) -- ind. hyp. = UiIUp D(Ai,x) d(P,a) -- set theory = U i I D(Ai,ya) (6.2)
Transparency No. 3-10
N = (QN,S,dN,SN,FN) : a NFA. M = (QM,S,dM,sM,FM) (denoted 2N): a DFA where QM = 2 QN dM(A,a) = DN(A,a) ( = q A dN(q,a) ) for every A QN. sM = SN and FM = {A QN | A FN {}}. note: States of M are subsets of states of N. Lem 6.3: for any A QN. and x in S*, DM(A,x) = DN(A,x). pf: by ind on |x|. if x = e => DM(A,e) = A = DN(A,e). --(def) if x = ya =>DM(A,ya) = dM(DM(A,y),a) -- (def) = dM(DN(A,y),a) -ind. hyp. = DN(DN(A,y),a) -- def of dM =DN(A, ya) -- lem 6.1 Theorem 6.4: M and N accept the same set. pf: x L(M) iff DM(sM,x)FM iff DN(SN,x) FN {} iff x L(N).
Transparency No. 3-11
1. NFA N accepting A = { x {0,1}* | the second symbol from the right is 1 } = {x1a | x {0,1}* and a {0,1} }. sol: 0,1 1 p q 0,1 r
0 1 --------------------------------------------------------{} {} {} -> {p} {p} {p,q} {q} {r} {r} {r}F {} {} {p,q} {p,r} {p,q,r} {p,r}F {p} {p,q} {q,r}F {r} {r} {p,q,r}F {p,r} {p,q,r}
2. DFA M equivalent to N is given as : 3. some states of M are redundant in the sense that they are never reachable from the start state and hence can be removed from the machine w/o affecting the languages accepted.
0 1 --------------------------------------------------------1 0,1 {} {} {} -> {p} {p} {p,q} p q r {q} {r} {r} {r}F {} {} {p} {p} {p,q} 1. Copy the transition table {p,q} {p,r} {p,q,r} 2. add Row(S) /* =def {p,r}F {p} {p,q} {r} {r} SumpS Row(p) to table */ {q,r}F {p,q,r}F {p,r} {p,q,r}
sol:
0,1
3. D={X|X in Row(p).tail } {S} // S is the initial set of states 4. While D != {} do { S1 = D.pop() ; // remove any element from D. add(Row(S1)) to table D = D U Row(S1).tail. } Transparency No. 3-13
e-transition
Another extension of FAs, useful but adds no more power. An e-transition is a transition with label e, a label standing for the empty string e. e p q The FA can take such a transition anytime w/o reading an input symbol. Ex 6.5 : The set accepted by the FA is {b,bb,bbb}. Ex 6.6 : A NFA-e accepting the set {x {a}* | |x| is dividable by 3 or 5 }. e e s t u b e b e b real advantage of e -transition: convenient for specification add no extra power
q
Ex6.5
NFA-e
N = (Q,S,d,S,F) : a NFA-e,where Q, S, S and F are the same as NFA, d : Q x (SU {e}) -> 2Q. The set Eclosure(A) is the set of ref. and transitive closure of the e-transition of A = { q Q | e-path p p1 p2 -pn with p A and pn = q }
Note: Eclosure(A) (abbreviated as EC(A) ) = EC(EC(A)). The multistep version of d is modified as follows: D: 2Q x S* 2Q where, for all A Q , y S*, a A D(A, e) = Eclosure(A) D(A, ya) = U p D(A,y) Eclosure( d(p,a) )
L(N) = { x | D(EC(S), x) F {} } //The language accepted by N
Transparency No. 3-15
E-closure
Eclosure(A) is the set of states reachable from states of A without consuming any input symbols, (i.e., qEclosure(A) iffp A s.t. q D(p, ek) for some k 0 ). Eclosure(A) can be computed as follows: 1. R=F={}; nF=A; //F: frontier; nF: new frontier 2. do { R = R U nF; F = nF; nF={}; 3. For each q F do 4. nF= nF U (d(q,e)- R) 5. }while nF {}; 6. return R Note:1. q D(A, ek) => q R after k-th iteration of the program. 2. We can precompute the matrix T* where T is the e-transition matrix of the NFA. and use the result to get Eclosure(A) for all required As.
Transparency No. 3-16
N = (QN,S,dN,SN,FN) : a NFA-e.where dN : Q x (SU {e}) -> 2Q. M = (QM,S,dM,sM,FM) (denoted 2N): a DFA where QM = { EC(A) | A QN } dM(A,a) = q Ec(A) EC(dN(q,a)) for every A QM. sM = EC(SN) and FM = {A QM | A FN {}}. note: States of M are subsets of states of N. Lem 6.3: for any A QN. and x S*, DM(A,x) = DN(A,x). pf: by ind on |x|. if x = e => DM(A,e) = A =EC(A) = DN(A,e). --(def) if x = ya =>DM(A,ya) = dM(DM(A,y),a) -- (def) = dM(DN(A,y),a) -- ind. hyp. = U q DN(A,y) EC(dN(q,a)) -- def of dM = DN(A, ya) def of DN Theorem 6.4: M and N accept the same set. pf: x L(M) iff DM(sM,x)FM iff DN(EC(SN),x) FN {} iff x L(N).
Transparency No. 3-17
If A and B are regular languages, then so are AB and A*. M = (Q1,S,d1,S1,F1), N=(Q2,S,d2,S2,F2) : two NFAs The machine M N, which firstly executes M and then execute N, can be defined as follows: M N =def (Q, S, d, S, F) where Q = disjoint union of Q1 and Q2, S = S1, F = F2, d = d1 U d2 U { (p, e, q ) | p F1 and q S2 } Lem: 1. x L(M) and y L(N) then xy L(MN) 2. x L(MN) => $ y,z s.t. x = yz and y L(M) and z L(N). Corollary: L(MN) = L(M) L(N)
M* machine
M = (Q1,S,d1,S1,F1) : a NFA The machine M*, which executes M a nondeterministic number of times, can be defined as follows: M* =def (Q, S, d, S, F) where Q = Q U {s,f}, where s and f are two new states Q S = {s}, F = {f}, d = d1 U {(s, e, f)} U {(s,e,p) | p S1 } U {(q,e,s) | q F1 } Theorem: L(M*) = L(M)* e M
e
e
Transparency No. 3-19
M*