hmm.h

经典numerical receip 配套代码

struct HMM {
	MatDoub a, b;
	VecInt obs;
	Int fbdone;
	Int mstat, nobs, ksym;
	Int lrnrm;
	MatDoub alpha, beta, pstate;
	VecInt arnrm, brnrm;
	Doub BIG, BIGI, lhood;
	HMM(MatDoub_I &aa, MatDoub_I &bb, VecInt_I &obs);
	void forwardbackward();
	void baumwelch();
	Doub loglikelihood() {return log(lhood)+lrnrm*log(BIGI);}
};

HMM::HMM(MatDoub_I &aa, MatDoub_I &bb, VecInt_I &obss) :
	a(aa), b(bb), obs(obss), fbdone(0),
	mstat(a.nrows()), nobs(obs.size()), ksym(b.ncols()),
	alpha(nobs,mstat), beta(nobs,mstat), pstate(nobs,mstat),
	arnrm(nobs), brnrm(nobs), BIG(1.e20), BIGI(1./BIG)  {
	Int i,j,k;
	Doub sum;
	if (a.ncols() != mstat) throw("transition matrix not square");
	if (b.nrows() != mstat) throw("symbol prob matrix wrong size");
	for (i=0; i<nobs; i++) {
		if (obs[i] < 0 || obs[i] >= ksym) throw("bad data in obs");
	}
	for (i=0; i<mstat; i++) {
		sum = 0.;
		for (j=0; j<mstat; j++) sum += a[i][j];
		if (abs(sum - 1.) > 0.01) throw("transition matrix not normalized");
		for (j=0; j<mstat; j++) a[i][j] /= sum;
	}
	for (i=0; i<mstat; i++) {
		sum = 0.;
		for (k=0; k<ksym; k++) sum += b[i][k];
		if (abs(sum - 1.) > 0.01) throw("symbol prob matrix not normalized");
		for (k=0; k<ksym; k++) b[i][k] /= sum;
	}
}
void HMM::forwardbackward() {
	Int i,j,t;
	Doub sum,asum,bsum;
	for (i=0; i<mstat; i++) alpha[0][i] = b[i][obs[0]];
	arnrm[0] = 0;
	for (t=1; t<nobs; t++) {
		asum = 0;
		for (j=0; j<mstat; j++) {
			sum = 0.;
			for (i=0; i<mstat; i++) sum += alpha[t-1][i]*a[i][j]*b[j][obs[t]];
			alpha[t][j] = sum;
			asum += sum;
		}
		arnrm[t] = arnrm[t-1];
		if (asum < BIGI) {
			++arnrm[t];
			for (j=0; j<mstat; j++) alpha[t][j] *= BIG;
		}
	}
	for (i=0; i<mstat; i++) beta[nobs-1][i] = 1.;
	brnrm[nobs-1] = 0;
	for (t=nobs-2; t>=0; t--) {
		bsum = 0.;
		for (i=0; i<mstat; i++) {
			sum = 0.;
			for (j=0; j<mstat; j++) sum += a[i][j]*b[j][obs[t+1]]*beta[t+1][j];
			beta[t][i] = sum;
			bsum += sum;
		}
		brnrm[t] = brnrm[t+1];
		if (bsum < BIGI) {
			++brnrm[t];
			for (j=0; j<mstat; j++) beta[t][j] *= BIG;
		}		
	}
	lhood = 0.;
	for (i=0; i<mstat; i++) lhood += alpha[0][i]*beta[0][i];
	lrnrm = arnrm[0] + brnrm[0];
	while (lhood < BIGI) {lhood *= BIG; lrnrm++;}
	for (t=0; t<nobs; t++) {
		sum = 0.;
		for (i=0; i<mstat; i++) sum += (pstate[t][i] = alpha[t][i]*beta[t][i]);
		// sum = lhood*pow(BIGI, lrnrm - arnrm[t] - brnrm[t]);
		for (i=0; i<mstat; i++) pstate[t][i] /= sum;
	}
	fbdone = 1;
}
void HMM::baumwelch() {
	Int i,j,k,t;
	Doub num,denom,term;
	MatDoub bnew(mstat,ksym);
	Doub powtab[10];
	for (i=0; i<10; i++) powtab[i] = pow(BIGI,i-6);
	if (fbdone != 1) throw("must do forwardbackward first");	
	for (i=0; i<mstat; i++) {
		denom = 0.;
		for (k=0; k<ksym; k++) bnew[i][k] = 0.;
		for (t=0; t<nobs-1; t++) {
			term = (alpha[t][i]*beta[t][i]/lhood)
				* powtab[arnrm[t] + brnrm[t] - lrnrm + 6];
			denom += term;
			bnew[i][obs[t]] += term;
		}
		for (j=0; j<mstat; j++) {
			num = 0.;
			for (t=0; t<nobs-1; t++) {
				num += alpha[t][i]*b[j][obs[t+1]]*beta[t+1][j]
					* powtab[arnrm[t] + brnrm[t+1] - lrnrm + 6]/lhood;
			}
			a[i][j] *= (num/denom);
		}
		for (k=0; k<ksym; k++) bnew[i][k] /= denom;
	}
	b = bnew;
	fbdone = 0;
}