reproduce.m

这个是关于neural fuzzy 算法的工具箱, 有例子程序, 请您用winzip解压缩

%% Reproduction -Main Evolutionary algorithm (Mutation, crossover, speciation) 

%% Neuro_Evolution_of_Augmenting_Topologies - NEAT 
%% developed by Kenneth Stanley ([e-mail]) & Risto Miikkulainen ([e-mail])
%% Coding by Christian Mayr ([e-mail])

function [new_population,updated_species_record,updated_innovation_record]=reproduce(population, species_record, innovation_record, initial, selection, crossover, mutation, speciation, generation, population_size);

% the following 'for' loop has these three objectives:

% 1.compute matrix of existing and propagating species from species_record (first row), assign their alloted number of offspring from the shared fitness (second row), and set their actual number of offspring to zero (third row) (will be incremented as new individuals are created from this species)

% 2.copy most fit individual in every species with more than initial.number_copy individuals unchanged into new generation (elitism) (But only if species is not dying out, i.e. has at least one individual allotted to itself in the new generation)
%   utilizes data stored in species_record.generation_record (index of individual in population having the highest fitness)

% 3.erase lowest percentage (initial.kill_percentage) in species with more than initial.number_for_kill individuals to keep them from taking part in reproduction
% Matlab actually doesn't offer a facility for redirecting the pointers which link one element in a structure with the next, so essentially this individual entry in the population structure cannot be erased. Rather, it will have its species ID set to zero, 
% which has the same effect of removing it from the rest of the reproduction cycle, since all reproduction functions access the population structure through the species ID and no species has an ID of zero

% Compute sum_average_fitnesses
sum_average_fitnesses=0;
for index_species=1:size(species_record,2)
   sum_average_fitnesses=sum_average_fitnesses+species_record(index_species).generation_record(2,size(species_record(index_species).generation_record,2))*(species_record(index_species).number_individuals>0);
end
% The following two lines only initialize the new_population structure. Since its species is set to 0, the rest of the algorithm will not bother with it. It gets overwritten as soon as the first really new individual is created
new_population(1)=population(1);
new_population(1).species=0;

overflow=0;
index_individual=0;
matrix_existing_and_propagating_species=[];
for index_species=1:size(species_record,2)
   if species_record(index_species).number_individuals>0 %test if species existed in old generation
      number_offspring=species_record(index_species).generation_record(2,size(species_record(index_species).generation_record,2))/sum_average_fitnesses*population_size; %compute number of offspring in new generation
      overflow=overflow+number_offspring-floor(number_offspring);
      if overflow>=1 %Since new species sizes are fractions, overflow sums up the difference between size and floor(size), and everytime this overflow is above 1, the species gets one additional individual alotted to it  
         number_offspring=ceil(number_offspring);
         overflow=overflow-1;
      else
         number_offspring=floor(number_offspring);
      end
      if number_offspring>0 %Check to see if species is dying out, only add those species to matrix_existing_and_propagating_species which have offspring in the new generation
         matrix_existing_and_propagating_species=[matrix_existing_and_propagating_species,[species_record(index_species).ID;number_offspring;0]]; %matrix (objective 1)
         if species_record(index_species).number_individuals>=initial.number_copy %check for condition for objective 2
            index_individual=index_individual+1;
            new_population(index_individual)=population(species_record(index_species).generation_record(4,size(species_record(index_species).generation_record,2))); % Objective 2
            matrix_existing_and_propagating_species(3,size(matrix_existing_and_propagating_species,2))=1; %Update matrix_existing_and_propagating_species
         end
      end
      
      if (species_record(index_species).number_individuals>initial.kill_percentage) & (ceil(species_record(index_species).number_individuals*(1-initial.kill_percentage))>2) %check condition for objective 3
         matrix_individuals_species=[find([population(:).species]==index_species);[population(find([population(:).species]==index_species)).fitness]];
         [sorted_fitnesses_in_species,sorting_vector]=sort(matrix_individuals_species(2,:));
         matrix_individuals_species=matrix_individuals_species(:,sorting_vector);
         sorting_vector=matrix_individuals_species(1,:);
 
         for index_kill=1:floor(species_record(index_species).number_individuals*initial.kill_percentage)
            population(sorting_vector(index_kill)).species=0; %objective 3
         end         
      end
      
   end
end

% generate reference of random individuals from every species from old population 
% cycle through species ID's, and add reference individuals from old population, new species of new population will be added during reproduction
index_ref=0;
for index_species_ref=1:size(species_record,2)
   if sum([population(:).species]==index_species_ref)>0 %Check if species exists in old population 
      index_ref=index_ref+1;
      [discard,index_ref_old]=max(([population(:).species]==index_species_ref).*rand(1,size(population,2)));
      population_ref(index_ref)=population(index_ref_old);
   end
end 
matrix_existing_and_propagating_species

%% Standard-reproduction (Main)
%% Cycle through all existing species
for index_species=1:size(matrix_existing_and_propagating_species,2)
   count_individuals_species=0;
   species_ID=matrix_existing_and_propagating_species(1,index_species); %this is ID of species which will be reproduced this cycle
   %IMPORTANT: index_species only has relevance to matrix_existing_and_propagating_species, all other mechanisms using species in some way must use species_ID
   
   % Linear Ranking and stochastic universal sampling
   % Ranking with selection.pressure
   fitnesses_species=[population(find([population(:).species]==species_ID)).fitness];
   index_fitnesses_species=find([population(:).species]==species_ID);
   [discard,sorted_fitnesses]=sort(fitnesses_species);
   ranking=zeros(1,size(fitnesses_species,2));
   ranking(sorted_fitnesses)=1:size(fitnesses_species,2);
   if size(fitnesses_species,2)>1
       FitnV=(2-selection.pressure+2*(selection.pressure-1)/(size(fitnesses_species,2)-1)*(ranking-1))';  
   else
       FitnV=2;
   end
   % stochastic universal sampling
   % First compute number of individuals to be selected (two parents required for every offspring through crossover, one for mutation) 
   number_overall=matrix_existing_and_propagating_species(2,index_species)-matrix_existing_and_propagating_species(3,index_species);
   number_crossover=round(crossover.percentage*number_overall);
   number_mutate=number_overall-number_crossover;
   Nind=size(fitnesses_species,2);
   Nsel=2*number_crossover+number_mutate;
   
   if Nsel==0 %rare case, in which a species with at least initial.number_copy individuals gets individual copied, but compares poorly to new generation, which results in this copied individual being 
      %the only individual of this species in new generation, so no crossover or mutation takes place. 
      %setting Nsel to 1 will prevent division by zero error, but will have no other effect since the propagation loop is governed by matrix_existing_and_propagating_species, not by Nsel
      Nsel=1;        
   end   
   
   % Perform stochastic universal sampling (Code Snippet from Genetic Algorithm toolbox 1.2 by Chipperfield et al)
   cumfit = cumsum(FitnV);
   trials = cumfit(Nind) / Nsel * (rand + (0:Nsel-1)');
   Mf = cumfit(:, ones(1, Nsel));
   Mt = trials(:, ones(1, Nind))';   
   [NewChrIx, ans] = find(Mt < Mf & [ zeros(1, Nsel); Mf(1:Nind-1, :) ] <= Mt);
   % Shuffle selected Individuals
   [ans, shuf] = sort(rand(Nsel, 1));
   NewChrIx = NewChrIx(shuf);
   % relate to indexes of individuals in population
   NewChrIx = index_fitnesses_species(NewChrIx);

   
   while matrix_existing_and_propagating_species(3,index_species)<matrix_existing_and_propagating_species(2,index_species) %cycle until actual number of offspring has reached allotted number of offspring                                                          
      index_individual=index_individual+1;
      count_individuals_species=count_individuals_species+1;
      % Crossover 
      if count_individuals_species<=number_crossover %O.k: we are doing crossover
         %Select Parents
         %select parent1          
         parent1=population(NewChrIx(2*count_individuals_species-1));
         %select parent2
         found_parent2=0;
         %sum([species_record(:).number_individuals]>0)
         if (rand<crossover.probability_interspecies) & (size(matrix_existing_and_propagating_species,2)>1) %select parent2 from other species (can only be done if there is more than one species in old population)          
            while found_parent2==0;                
               [discard,index_parent2]=max(rand(1,size(population,2)));
               parent2=population(index_parent2);
               found_parent2=((parent2.species~=0) & (parent2.species~=parent1.species)); %check if parent2.species is not species 0 (deleted individual) or species of parent1
            end
            parent2.fitness=parent1.fitness; %set fitnesses to same to ensure that disjoint and excess genes are inherited fully from both parents (tip from ken)
         else % O.K. we take parent2 from same species as parent1
             parent2=population(NewChrIx(2*count_individuals_species));
         end
             
         %Crossover
         %inherit nodes from both parents
         new_individual.nodegenes=[];
         matrix_node_lineup=[[parent1.nodegenes(1,:);1:size(parent1.nodegenes,2);zeros(1,size(parent1.nodegenes,2))],[parent2.nodegenes(1,:);zeros(1,size(parent2.nodegenes,2));1:size(parent2.nodegenes,2)]];
         [discard,sort_node_vec]=sort(matrix_node_lineup(1,:));
         matrix_node_lineup=matrix_node_lineup(:,sort_node_vec);
         node_number=0;
         for index_node_sort=1:size(matrix_node_lineup,2)   
            if node_number~=matrix_node_lineup(1,index_node_sort)
               if matrix_node_lineup(2,index_node_sort)>0
                  new_individual.nodegenes=[new_individual.nodegenes,parent1.nodegenes(:,matrix_node_lineup(2,index_node_sort))];
               else
                  new_individual.nodegenes=[new_individual.nodegenes,parent2.nodegenes(:,matrix_node_lineup(3,index_node_sort))];
               end               
               node_number=matrix_node_lineup(1,index_node_sort);
            end
         end
         %Crossover connection genes
         %first do lineup of connection genes
         matrix_lineup=[[parent1.connectiongenes(1,:);1:size(parent1.connectiongenes,2);zeros(1,size(parent1.connectiongenes,2))],[parent2.connectiongenes(1,:);zeros(1,size(parent2.connectiongenes,2));1:size(parent2.connectiongenes,2)]];
         [discard,sort_vec]=sort(matrix_lineup(1,:));
         matrix_lineup=matrix_lineup(:,sort_vec);
         final_matrix_lineup=[];
         innovation_number=0;
         for index_sort=1:size(matrix_lineup,2)   
            if innovation_number~=matrix_lineup(1,index_sort)
               final_matrix_lineup=[final_matrix_lineup,matrix_lineup(:,index_sort)];
               innovation_number=matrix_lineup(1,index_sort);
            else
               final_matrix_lineup(2:3,size(final_matrix_lineup,2))=final_matrix_lineup(2:3,size(final_matrix_lineup,2))+matrix_lineup(2:3,index_sort);
            end
         end            
         % O.K. Connection Genes are lined up, start with crossover
         new_individual.connectiongenes=[];

         for index_lineup=1:size(final_matrix_lineup,2)
            if (final_matrix_lineup(2,index_lineup)>0) & (final_matrix_lineup(3,index_lineup)>0) %check for matching genes, do crossover
               if rand<0.5 %random crossover for matching genes
                  new_individual.connectiongenes=[new_individual.connectiongenes,parent1.connectiongenes(:,final_matrix_lineup(2,index_lineup))];
               else
                  new_individual.connectiongenes=[new_individual.connectiongenes,parent2.connectiongenes(:,final_matrix_lineup(3,index_lineup))];
               end
               if rand>crossover.probability_multipoint %weight averaging for offspring, otherwise the randomly inherited weights are left undisturbed
                  new_individual.connectiongenes(4,size(new_individual.connectiongenes,2))=(parent1.connectiongenes(4,final_matrix_lineup(2,index_lineup))+parent2.connectiongenes(4,final_matrix_lineup(3,index_lineup)))/2;
               end                              
            end                   
            parent1_flag=sum(final_matrix_lineup(2,index_lineup+1:size(final_matrix_lineup,2)));  % test if there exist further connection genes from index_lineup+1 to end of final_matrix_lineup for parent1 (to detect excess)
            parent2_flag=sum(final_matrix_lineup(3,index_lineup+1:size(final_matrix_lineup,2)));  % test if there exist further connection genes from index_lineup+1 to end of final_matrix_lineup for parent1 (to detect excess)
            % Two cases to check (excess is taken care of in the disjoint gene checks)
            if (final_matrix_lineup(2,index_lineup)>0) & (final_matrix_lineup(3,index_lineup)==0) %Disjoint parent1
               if parent1.fitness>=parent2.fitness
                  new_individual.connectiongenes=[new_individual.connectiongenes,parent1.connectiongenes(:,final_matrix_lineup(2,index_lineup))];
               end         
            end
            if (final_matrix_lineup(2,index_lineup)==0) & (final_matrix_lineup(3,index_lineup)>0) %Disjoint parent2
               if parent2.fitness>=parent1.fitness
                  new_individual.connectiongenes=[new_individual.connectiongenes,parent2.connectiongenes(:,final_matrix_lineup(3,index_lineup))];
               end              
            end             
         end
         new_individual.fitness=0; %has no impact on algorithm, only required for assignment to new population
         new_individual.species=parent1.species; %will be species hint for speciation
      else % no crossover, just copy a individual of the species and mutate in subsequent steps