void LayerNet::gen_init (
TrainingSet *tptr , // Training set to use
struct LearnParams *lptr // User's general learning parameters
)
{
int i, istart, individual, best_individual, generation, n_cross ;
int first_child, parent1, parent2, improved, crosspt, nchoices, *choices ;
int initpop, popsize, gens, climb, nvars, chromsize, split, ind ;
float pcross, pmutate, error, besterror, *errors, *fitness, worst ;
float fquit, favor_best, fitfac, maxerr, minerr, avgerr, overinit ;
SingularValueDecomp *sptr ;
struct GenInitParams *gptr ; // User's genetic initialization parameters
char *pool1, *pool2, *oldpop, *newpop, *popptr, *temppop, *best ;
char msg[80] ;
gptr = lptr->gp ;
popsize = gptr->pool ;
gens = gptr->gens ;
climb = gptr->climb ;
overinit = gptr->overinit ;
pcross = gptr->pcross ;
pmutate = gptr->pmutate ;
fquit = lptr->quit_err ;
favor_best = 3.1 ;
fitfac = -20.0 ;
/*
--------------------------------------------------------------------------------
Do all scratch memory allocation.
--------------------------------------------------------------------------------
*/
/*
Allocate the singular value decomposition object for REGRESS.
*/
if (nhid2 == 0) // One hidden layer
nvars = nhid1 + 1 ;
else // Two hidden layers
nvars = nhid2 + 1 ;
MEMTEXT ( "GEN_INIT: new SingularValueDecomp" ) ;
sptr = new SingularValueDecomp ( tptr->ntrain , nvars , 1 ) ;
if ((sptr == NULL) || ! sptr->ok) {
memory_message("for genetic initialization. Try ANNEAL NOREGRESS.");
neterr = 1.0 ; // Flag failure to LayerNet::learn which called us
if (sptr != NULL)
delete sptr ;
return ;
}
chromsize = nhid1 * (nin+1) ; // Length of an individual's chromosome
if (nhid2) // is the number of hidden weights
chromsize += nhid2 * (nhid1+1) ;
errors = fitness = NULL ;
choices = NULL ;
pool1 = pool2 = NULL ;
MEMTEXT ( "GEN_INIT: errors, fitness, choices, best, pool1,pool2");
if (((errors = (float*) MALLOC ( popsize * sizeof(float))) == NULL)
|| ((fitness = (float*) MALLOC ( popsize * sizeof(float))) == NULL)
|| ((best = (char*) MALLOC( chromsize )) == NULL)
|| ((choices = (int*) MALLOC ( popsize * sizeof(int))) == NULL)
|| ((pool1 = (char*) MALLOC( popsize * chromsize )) == NULL)
|| ((pool2 = (char*) MALLOC( popsize * chromsize )) == NULL)) {
if (errors != NULL)
FREE ( errors ) ;
if (fitness != NULL)
FREE ( fitness ) ;
if (choices != NULL)
FREE ( choices ) ;
if (pool1 != NULL)
FREE ( pool1 ) ;
if (pool2 != NULL)
FREE ( pool2 ) ;
delete sptr ;
memory_message("for genetic initialization. Try ANNEAL NOREGRESS." ) ;
neterr = 1.0 ; // Flag failure to LayerNet::learn which called us
return ;
}
/*
Generate initial population pool.
We also preserve the best weights across all generations,
as this is what we will ultimately return to the user.
Its mean square error is besterror.
*/
besterror = 1.e30 ; // For saving best (across all individuals and gens)
maxerr = avgerr = 0.0 ; // For progress display only
best_individual = 0 ; // Safety only
initpop = popsize * overinit ; // Overinitialization of initial population
progress_message ( "\nGenerating initial population" ) ;
for (ind=0 ; ind<initpop ; ind++) { // Try overinitialization times
if (ind<popsize) // If still in pop size limit
individual = ind ; // just use next avail space
else { // Else we search entire pop
worst = -1. ; // for the worst member
for (i=0 ; i<popsize ; i++) { // which we will then replace
if (errors[i] > worst) {
worst = errors[i] ;
individual = i ;
}
}
avgerr -= worst ; // Exclude discards from average
}
popptr = pool1 + individual * chromsize ; // Build init pop in pool1
rand_ind ( popptr , chromsize ) ; // Randomly generate individual
decode ( popptr , nin , nhid1 , nhid2 , // Convert genotype (chromosome)
hid1_coefs , hid2_coefs ); // to phenotype (weights)
error = regress ( tptr , sptr ) ; // Evaluate network error
errors[individual] = error ; // and keep all errors
if (error < besterror) { // Keep track of best
besterror = error ; // as it is returned to user
best_individual = individual ; // This is its index in pool1
}
if (error > maxerr) // Max and average error are
maxerr = error ; // for progress display only
avgerr += error ;
if (error <= fquit)
break ;
progress_message ( "." ) ;
}
sprintf (msg , "\nInitial pop: Min err=%7.4lf Max=%7.4lf Avg=%7.4lf",
100. * besterror, 100. * maxerr, 100.0 * avgerr / (float) popsize);
progress_message ( msg ) ;
/*
The initial population has been built in pool1.
Copy its best member to 'best' in case it never gets beat (unlikely
but possible!).
Also, we will need best if the climb option is true.
*/
popptr = pool1 + best_individual * chromsize ; // Point to best
memcpy ( best , popptr , chromsize ) ; // and save it
/*
This is the main generation loop. There are two areas for population pool
storage: pool1 and pool2. At any given time, oldpop will be set to one of
them, and newpop to the other. This avoids a lot of copying.
*/
oldpop = pool1 ; // This is the initial population
newpop = pool2 ; // The next generation is created here
for (generation=0 ; generation<gens ; generation++) {
if (error <= fquit) // We may have satisfied this in init pop
break ; // So we test at start of generation loop
error_to_fitness ( popsize , favor_best , fitfac , errors , fitness ) ;
fitness_to_choices ( popsize , fitness , choices ) ;
nchoices = popsize ; // Will count down as choices array emptied
n_cross = pcross * popsize ; // Number crossing over
first_child = 1 ; // Generating first of parent's 2 children?
improved = 0 ; // Flags if we beat best
if (climb) { // If we are to hill climb
memcpy ( newpop , best , chromsize ) ; // start with best
errors[0] = besterror ; // Record its error
istart = 1 ; // and start children past it
}
else
istart = 0 ;
/*
Generate the children
*/
maxerr = avgerr = 0.0 ; // For progress display only
minerr = 1.0 ; // Ditto
for (individual=istart ; individual<popsize ; individual++) {
popptr = newpop + individual * chromsize ; // Will put this child here
if (first_child) // If this is the first of 2 children, pick parents
pick_parents ( &nchoices , choices , &parent1 , &parent2 ) ;
if (n_cross-- > 0) // Do crossovers first
reproduce ( oldpop + parent1 * chromsize , oldpop + parent2 * chromsize ,
first_child , chromsize , popptr , &crosspt , &split ) ;
else if (first_child) // No more crossovers, so just copy parent
memcpy ( popptr , oldpop + parent1 * chromsize , chromsize ) ;
else
memcpy ( popptr , oldpop + parent2 * chromsize , chromsize );
if (pmutate > 0.0)
mutate ( popptr , chromsize , pmutate ) ;
decode ( popptr , nin , nhid1 , nhid2 , hid1_coefs , hid2_coefs ) ;
error = regress ( tptr , sptr ) ; // Evaluate child's error
errors[individual] = error ; // and keep each
if (error < besterror) { // Keep track of best
besterror = error ; // It will be returned to user
best_individual = individual ; // This is its index in newpop
improved = 1 ; // Flag so we copy it later
}
if (error > maxerr) // Min, max and average error
maxerr = error ; // for progress display only
if (error < minerr)
minerr = error ;
avgerr += error ;
if (error <= fquit)
break ;
first_child = ! first_child ;
} // For all genes in population
/*
We finished generating all children. If we improved (one of these
children beat the best so far) then copy that child to the best.
Swap oldpop and newpop for the next generation.
*/
if (improved) {
popptr = newpop + best_individual * chromsize ; // Point to best
memcpy ( best , popptr , chromsize ) ; // and save it
}
temppop = oldpop ; // Switch old and new pops for next generation
oldpop = newpop ;
newpop = temppop ;
sprintf(msg, "\nGeneration %3d: Min err=%7.4lf Max=%7.4lf Avg=%7.4lf",
generation+1, 100. * minerr, 100. * maxerr,
100.0 * avgerr / (float) popsize ) ;
progress_message ( msg ) ;
}
/*
We are all done.
*/
decode ( best , nin , nhid1 , nhid2 , hid1_coefs , hid2_coefs ) ;
besterror = regress ( tptr , sptr ) ; // Evaluate network error
MEMTEXT ( "GEN_INIT: errors, fitness, choices, best, pool1,pool2");
FREE ( errors ) ;
FREE ( fitness ) ;
FREE ( choices ) ;
FREE ( best ) ;
FREE ( pool1 ) ;
FREE ( pool2 ) ;
MEMTEXT ( "GEN_INIT: delete sptr" ) ;
delete sptr ;
}
double genetic (
double (*func)( int n , double *x ) , // Function to be minimized
int ngens , // Number of complete generations
int popsize , // Number of individuals in population
int climb , // Do we hill climb via elitism?
double stddev , // Standard deviation for initial population
double pcross , // Probability of crossover (.6-.9 typical)
double pmutate , // Probability of mutation (.0001 to .001 typical)
double favor_best ,// Factor for favoring best over average (2-3 is good)
double fitfac , // Factor for converting fval to raw fitness (-20 good)
double fquit , // Quit if function reduced to this amount
int nvars , // Number of variables in x
double *x , // Set to starting estimate when called; returns best
double *best, // Work vector nvars long
int *choices , // Work vector popsize long
double *fvals , // Work vector popsize long
double *fitness , // Work vector popsize long
double *pool1, // Work vector nvars * popsize long
double *pool2 ) // Work vector nvars * popsize long
{
int istart, individual, best_individual, generation, n_cross ;
int first_child, parent1, parent2, improved, crosspt, nchoices ;
double fval, bestfval, *oldpop, *newpop, *popptr, *temppop ;
/*
Generate initial population pool.
We also preserve the best point across all generations, as this is what
we will ultimately return to the user. Its objective function value is
bestfval.
*/
bestfval = 1.e30 ; // For saving best (across all individuals)
best_individual = 0 ; // Safety only
for (individual=0 ; individual<popsize ; individual++) {
popptr = pool1 + individual * nvars ; // Build population in pool1
shake ( nvars , x , popptr , stddev ) ; // Randomly perturb about init x
fval = (*func) ( nvars , popptr ) ; // Evaluate function there
fvals[individual] = fval ; // and keep function value of each
if (fval < bestfval) { // Keep track of best
bestfval = fval ; // as it will be returned to user
best_individual = individual ; // This is its index in pool1
}
if (fval <= fquit)
break ;
}
/*
The initial population has been built in pool1.
Copy its best member to 'best' in case it never gets beat (unlikely
but possible!).
*/
popptr = pool1 + best_individual * nvars ; // Point to best
memcpy ( best , popptr , nvars * sizeof(double) ) ; // and save it
/*
This is the main generation loop. There are two areas for population pool
storage: pool1 and pool2. At any given time, oldpop will be set to one of
them, and newpop to the other. This avoids a lot of copying.
*/
oldpop = pool1 ; // This is the initial population
newpop = pool2 ; // The next generation is created here
for (generation=0 ; generation<ngens ; generation++) {
if (fval <= fquit) // We may have satisfied this in initial population
break ; // So we test at start of generation loop
fval_to_fitness ( popsize , favor_best , fitfac , fvals , fitness ) ;
fitness_to_choices ( popsize , fitness , choices ) ;
nchoices = popsize ; // Will count down as choices array emptied
n_cross = pcross * popsize ; // Number crossing over
first_child = 1 ; // Generating first of parent's 2 children?
improved = 0 ; // Flags if we beat best
if (climb) { // If we are to hill climb
memcpy ( newpop , best , nvars * sizeof(double) ) ; // start with best
fvals[0] = bestfval ; // Record its error
istart = 1 ; // and start children past it
}
else
istart = 0 ;
/*
Generate the children
*/
for (individual=istart ; individual<popsize ; individual++) {
popptr = newpop + individual * nvars ; // Will put this child here
if (first_child) // If this is the first of 2 children, pick parents
pick_parents ( &nchoices , choices , &parent1 , &parent2 ) ;
if (n_cross-- > 0) // Do crossovers first
reproduce ( oldpop + parent1 * nvars , oldpop + parent2 * nvars ,
first_child , nvars , popptr , &crosspt ) ;
else if (first_child) // No more crossovers, so just copy parent
memcpy( popptr , oldpop + parent1 * nvars , nvars * sizeof(double));
else
memcpy( popptr , oldpop + parent2 * nvars , nvars * sizeof(double));
if (pmutate > 0.0)
mutate ( popptr , nvars , stddev , pmutate ) ;
fval = (*func) ( nvars , popptr ) ; // Evaluate function for this child
fvals[individual] = fval ; // and keep function value of each
if (fval < bestfval) { // Keep track of best
bestfval = fval ; // It will be returned to user
best_individual = individual ; // This is its index in newpop
improved = 1 ; // Flag so we copy it later
}
if (fval <= fquit)
break ;
first_child = ! first_child ;
} // For all genes in population
/*
We finished generating all children. If we improved (one of these
children beat the best so far) then copy that child to the best.
Swap oldpop and newpop for the next generation.
*/
if (improved) {
popptr = newpop + best_individual * nvars ; // Point to best
memcpy ( best , popptr , nvars * sizeof(double) ) ; // and save it
}
temppop = oldpop ; // Switch old and new pops for next generation
oldpop = newpop ;
newpop = temppop ;
}
/*
We are all done. Copy the best to x, as that is how we return it.
*/
memcpy ( x , best , nvars * sizeof(double) ) ; // Return best
return bestfval ;
}