source: cpp/frams/genetics/oper_fx.cpp @ 743

// This file is a part of Framsticks SDK.  http://www.framsticks.com/
// Copyright (C) 1999-2015  Maciej Komosinski and Szymon Ulatowski.
// See LICENSE.txt for details.

#include <ctype.h>  //isupper()
#include "oper_fx.h"
#include <common/log.h>
#include <common/nonstd_math.h>
#include <frams/util/rndutil.h>

static double distrib_force[] =   // for '!'
{
        3,             // distribution 0 -__/ +1
        0.001, 0.2,    // "slow" neurons
        0.001, 1,
        1, 1,          // "fast" neurons
};
static double distrib_inertia[] =  // for '='
{
        2,             // distribution 0 |..- +1
        0, 0,          // "fast" neurons
        0.7, 0.98,
};
static double distrib_sigmo[] =  // for '/'
{
        5,             // distribution -999 -..-^-..- +999
        -999, -999,    //"perceptron"
        999, 999,
        -5, -1,        // nonlinear
        1, 5,
        -1, 1,         // ~linear
};


int GenoOperators::roulette(const double *probtab, const int count)
{
        double sum = 0;
        int i;
        for (i = 0; i < count; i++) sum += probtab[i];
        double sel = rnd01*sum;
        for (sum = 0, i = 0; i < count; i++) { sum += probtab[i]; if (sel < sum) return i; }
        return -1;
}

bool GenoOperators::getMinMaxDef(ParamInterface *p, int i, double &mn, double &mx, double &def)
{
        mn = mx = def = 0;
        int defined = 0;
        if (p->type(i)[0] == 'f')
        {
                double _mn = 0, _mx = 1, _def = 0.5;
                defined = p->getMinMaxDouble(i, _mn, _mx, _def);
                if (defined == 1) _mx = _mn + 1.0;
                if (_mx < _mn && defined == 3) _mn = _mx = _def; //only default was defined, let's assume min=max=default
                if (defined < 3) _def = (_mn + _mx) / 2.0;
                mn = _mn; mx = _mx; def = _def;
        }
        if (p->type(i)[0] == 'd')
        {
                paInt _mn = 0, _mx = 1, _def = 0;
                defined = p->getMinMaxInt(i, _mn, _mx, _def);
                if (defined == 1) _mx = _mn + 1;
                if (_mx < _mn && defined == 3) _mn = _mx = _def; //only default was defined, let's assume min=max=default
                if (defined < 3) _def = (_mn + _mx) / 2;
                mn = _mn; mx = _mx; def = _def;
        }
        return defined == 3;
}

int GenoOperators::selectRandomProperty(Neuro* n)
{
        int neuext = n->extraProperties().getPropCount(),
                neucls = n->getClass() == NULL ? 0 : n->getClass()->getProperties().getPropCount();
        if (neuext + neucls == 0) return -1; //no properties in this neuron
        int index = randomN(neuext + neucls);
        if (index >= neuext) index = index - neuext + 100;
        return index;
}

double GenoOperators::mutateNeuProperty(double current, Neuro *n, int i)
{
        if (i == -1) return mutateCreepNoLimit('f', current, -10, 10); //i==-1: mutating weight of neural connection
        Param p;
        if (i >= 100) { i -= 100; p = n->getClass()->getProperties(); }
        else p = n->extraProperties();
        double newval = current;
        /*bool ok=*/getMutatedProperty(p, i, current, newval);
        return newval;
}

bool GenoOperators::mutatePropertyNaive(ParamInterface &p, int i)
{
        double mn, mx, df;
        if (p.type(i)[0] != 'f' && p.type(i)[0] != 'd') return false; //don't know how to mutate
        getMinMaxDef(&p, i, mn, mx, df);

        ExtValue ev;
        p.get(i, ev);
        ev.setDouble(mutateCreep(p.type(i)[0], ev.getDouble(), mn, mx));
        p.set(i, ev);
        return true;
}

bool GenoOperators::mutateProperty(ParamInterface &p, int i)
{
        double newval;
        ExtValue ev;
        p.get(i, ev);
        bool ok = getMutatedProperty(p, i, ev.getDouble(), newval);
        if (ok) { ev.setDouble(newval); p.set(i, ev); }
        return ok;
}

bool GenoOperators::getMutatedProperty(ParamInterface &p, int i, double oldval, double &newval)
{
        newval = 0;
        if (p.type(i)[0] != 'f' && p.type(i)[0] != 'd') return false; //don't know how to mutate
        const char *n = p.id(i), *na = p.name(i);
        if (strcmp(n, "si") == 0 && strcmp(na, "Sigmoid") == 0) newval = CustomRnd(distrib_sigmo); else
                if (strcmp(n, "in") == 0 && strcmp(na, "Inertia") == 0) newval = CustomRnd(distrib_inertia); else
                        if (strcmp(n, "fo") == 0 && strcmp(na, "Force") == 0) newval = CustomRnd(distrib_force); else
                        {
                double mn, mx, df;
                getMinMaxDef(&p, i, mn, mx, df);
                newval = mutateCreep(p.type(i)[0], oldval, mn, mx);
                        }
        return true;
}

double GenoOperators::mutateCreepNoLimit(char type, double current, double mn, double mx)
{
        double result = RndGen.Gauss(current, (mx - mn) / 2 / 5); // /halfinterval, 5 times narrower
        if (type == 'd') { result = int(result + 0.5); if (result == current) result += randomN(2) * 2 - 1; }
        else result = floor(result * 1000 + 0.5) / 1000.0; //round
        return result;
}

double GenoOperators::mutateCreep(char type, double current, double mn, double mx)
{
        double result = mutateCreepNoLimit(type, current, mn, mx); //TODO consider that when boundary is touched (reflect/absorb below), the default precision (3 digits) may change. Is it good or bad?
        //reflect:
        if (result > mx) result = mx - (result - mx); else
                if (result < mn) result = mn + (mn - result);
        //absorb (just in case 'result' exceeded the allowed range so much):
        if (result > mx) result = mx; else
                if (result < mn) result = mn;
        return result;
}

void GenoOperators::setIntFromDoubleWithProbabilisticDithering(ParamInterface &p, int index, double value) //TODO
{
        p.setInt(index, (paInt)value); //TODO value=2.5 will result in 2 but we want it to be 2 or 3 with equal probability. value=2.1 would be mostly 2, rarely 3. Careful with negative values (test it!)
}

void GenoOperators::linearMix(ParamInterface &p1, int i1, ParamInterface &p2, int i2, double proportion)
{
        char type1 = p1.type(i1)[0];
        char type2 = p2.type(i2)[0];
        if (type1 == 'f' && type2 == 'f')
        {
                double v1 = p1.getDouble(i1);
                double v2 = p2.getDouble(i2);
                p1.setDouble(i1, v1*proportion + v2*(1 - proportion));
                p2.setDouble(i2, v2*proportion + v1*(1 - proportion));
        }
        else
                if (type1 == 'd' && type2 == 'd')
                {
                int v1 = p1.getInt(i1);
                int v2 = p2.getInt(i2);
                setIntFromDoubleWithProbabilisticDithering(p1, i1, v1*proportion + v2*(1 - proportion));
                setIntFromDoubleWithProbabilisticDithering(p2, i2, v2*proportion + v1*(1 - proportion));
                }
                else
                        logPrintf("GenoOperators", "linearMix", LOG_WARN, "Cannot mix values of types '%c' and '%c'", type1, type2);
}

NeuroClass* GenoOperators::getRandomNeuroClass()
{
        vector<NeuroClass*> active;
        for (int i = 0; i < Neuro::getClassCount(); i++)
                if (Neuro::getClass(i)->genactive)
                        active.push_back(Neuro::getClass(i));
        if (active.size() == 0) return NULL; else return active[randomN(active.size())];
}

int GenoOperators::getRandomNeuroClassWithOutput(const vector<NeuroClass*>& NClist)
{
        vector<int> allowed;
        for (size_t i = 0; i < NClist.size(); i++)
                if (NClist[i]->getPreferredOutput() != 0) //this NeuroClass provides output
                        allowed.push_back(i);
        if (allowed.size() == 0) return -1; else return allowed[randomN(allowed.size())];
}

int GenoOperators::getRandomNeuroClassWithInput(const vector<NeuroClass*>& NClist)
{
        vector<int> allowed;
        for (size_t i = 0; i < NClist.size(); i++)
                if (NClist[i]->getPreferredInputs() != 0) //this NeuroClass wants one input connection or more                  
                        allowed.push_back(i);
        if (allowed.size() == 0) return -1; else return allowed[randomN(allowed.size())];
}

int GenoOperators::getRandomChar(const char *choices, const char *excluded)
{
        int allowed_count = 0;
        for (size_t i = 0; i < strlen(choices); i++) if (!strchrn0(excluded, choices[i])) allowed_count++;
        if (allowed_count == 0) return -1; //no char is allowed
        int rnd_index = randomN(allowed_count) + 1;
        allowed_count = 0;
        for (size_t i = 0; i < strlen(choices); i++)
        {
                if (!strchrn0(excluded, choices[i])) allowed_count++;
                if (allowed_count == rnd_index) return i;
        }
        return -1; //never happens
}

NeuroClass* GenoOperators::parseNeuroClass(char*& s)
{
        int maxlen = (int)strlen(s);
        int NClen = 0;
        NeuroClass *NC = NULL;
        for (int i = 0; i<Neuro::getClassCount(); i++)
        {
                const char *ncname = Neuro::getClass(i)->name.c_str();
                int ncnamelen = (int)strlen(ncname);
                if (maxlen >= ncnamelen && ncnamelen>NClen && (strncmp(s, ncname, ncnamelen) == 0))
                {
                        NC = Neuro::getClass(i);
                        NClen = ncnamelen;
                }
        }
        s += NClen;
        return NC;
}

Neuro* GenoOperators::findNeuro(const Model *m, const NeuroClass *nc)
{
        if (!m) return NULL;
        for (int i = 0; i < m->getNeuroCount(); i++)
                if (m->getNeuro(i)->getClass() == nc) return m->getNeuro(i);
        return NULL; //neuron of class 'nc' was not found
}

int GenoOperators::neuroClassProp(char*& s, NeuroClass *nc, bool also_v1_N_props)
{
        int len = (int)strlen(s);
        int Len = 0, I = -1;
        if (nc)
        {
                Param p = nc->getProperties();
                for (int i = 0; i<p.getPropCount(); i++)
                {
                        const char *n = p.id(i);
                        int l = (int)strlen(n);
                        if (len >= l && l>Len && (strncmp(s, n, l) == 0)) { I = 100 + i; Len = l; }
                        if (also_v1_N_props) //recognize old properties symbols /=!
                        {
                                if (strcmp(n, "si") == 0) n = "/"; else
                                        if (strcmp(n, "in") == 0) n = "="; else
                                                if (strcmp(n, "fo") == 0) n = "!";
                                l = (int)strlen(n);
                                if (len >= l && l > Len && (strncmp(s, n, l) == 0)) { I = 100 + i; Len = l; }
                        }
                }
        }
        Neuro n;
        Param p = n.extraProperties();
        for (int i = 0; i<p.getPropCount(); i++)
        {
                const char *n = p.id(i);
                int l = (int)strlen(n);
                if (len >= l && l>Len && (strncmp(s, n, l) == 0)) { I = i; Len = l; }
        }
        s += Len;
        return I;
}

bool GenoOperators::isWS(const char c)
{
        return c == ' ' || c == '\n' || c == '\t' || c == '\r';
}

void GenoOperators::skipWS(char *&s)
{
        if (s == NULL)
                logMessage("GenoOperators", "skipWS", LOG_WARN, "NULL reference!");
        else
                while (isWS(*s)) s++;
}

bool GenoOperators::areAlike(char *g1, char *g2)
{
        while (*g1 || *g2)
        {
                skipWS(g1);
                skipWS(g2);
                if (*g1 != *g2) return false; //when difference
                if (!*g1 && !*g2) break; //both end
                g1++;
                g2++;
        }
        return true; //equal
}

char* GenoOperators::strchrn0(const char *str, char ch)
{
        return ch == 0 ? NULL : strchr((char*)str, ch);
}

bool GenoOperators::isNeuroClassName(const char firstchar)
{
        return isupper(firstchar) || firstchar == '|' || firstchar == '@' || firstchar == '*';
}