dynamic_main.m

免疫算法实现多峰、多极值函数平面曲面拟合

% the main program of the dynamic training
% -------------------------------------
clear('all');
clc;
load('matlabhia31.mat');
% record how many times have the parameters been changed
dynamic_counter=1;
max_change_time=100;
per_evo_time=20;
result3=result2;
% -------------------------------------
proo=zeros(per_evo_time+2,max_change_time);
% main loop
% -------------------------------------
while dynamic_counter<=max_change_time
    % caculate the importance of each neuron    
    importance=judge(result2,t_x,t_y);
    % find out the global-fix neurons
    ss=20;
    global_neuron_idx=find(importance>ss);
    global_neuron_num=length(global_neuron_idx);
    while(global_neuron_num<10||global_neuron_num>20)
        if(global_neuron_num<10)
            ss=ss-1;
        end
        if(global_neuron_num>20)
            ss=ss+1;
        end
        global_neuron_idx=find(importance>ss);
        global_neuron_num=length(global_neuron_idx);
    end
    % the number of subtle-fix neurons 
    subtle_neuron_num=size(result2,1)-global_neuron_num;
    subtle_neuron_idx=1:size(result2,1);
    % remove the global neurons from the individual
    subtle_neuron_idx(global_neuron_idx)=[];
    
    % generate the dynamic training sets
    [t_x,t_y,te_x,te_y]=drosen;
    winner_fitness=fitness(result2,t_x,t_y);
    pro(per_evo_time+1,dynamic_counter)=winner_fitness;
    % calculate the fitness value of the static individual
    winner_fitness=fitness(result3,t_x,t_y);
    proo(per_evo_time+1,dynamic_counter)=winner_fitness;

    
    fprintf('dynamic counter: %d\ncurrent fitness: %f\n',dynamic_counter,winner_fitness);
    
    % re-modify if the error exceed the tolerance
    % ---------------------------------
    if winner_fitness<10000
        tic;
        % find the samples that need further modification 
        ss=0.0001;
        err_idx=find_error(result2,t_x,t_y,ss);
        while(length(err_idx)>40)
            ss=ss+0.0001;
            err_idx=find_error(result2,t_x,t_y,ss);
        end
        err_x=t_x(:,err_idx);
        if (subtle_neuron_num~=1 & subtle_neuron_num~=0 & ~isempty(err_x))
            fprintf('subtle-neuron-num: %d\n',subtle_neuron_num);
            % create new neurons
            [index,center]=kmeans(err_x',subtle_neuron_num,'distance','sqEuclidean');
        else
            center=result2(subtle_neuron_idx,2:size(result2,2)-1);
        end
        
        % generate new population
        % ------------------------------
        % the winner is preserved
        ini_pop=zeros(size(result2,1),size(result2,2),40);
        ini_pop(:,1,:)=1;
        for i=1:30
            ini_pop(:,:,i)=result2;
        end
        
        % add the subtle-fix neurons to the winner network
        ini_pop(subtle_neuron_idx,2:size(result2,2)-1,3)=center;
        ini_pop(subtle_neuron_idx,size(result2,2),3)=result2(subtle_neuron_idx,size(result2,2));
        for i=2:2:18
            ini_pop(subtle_neuron_idx,2:size(result2,2),i)=ini_pop(subtle_neuron_idx,2:4,1)+rand(subtle_neuron_num,3,1);
        end
        for i=20:2:38
            ini_pop(subtle_neuron_idx,2:size(result2,2),i)=ini_pop(subtle_neuron_idx,2:4,3)+rand(subtle_neuron_num,3,1);
        end
        ini_pop(:,2:size(result2,2)-1,29:2:37)=-2+4*rand(30,2,5);
        ini_pop(:,4,29:2:37)=1+7*rand(30,1,5);
        ini_pop(:,2:size(result2,2)-1,39:40)=-2+4*rand(30,2,2);
        ini_pop(:,4,39:40)=1+7*rand(30,1,2);
        % -----------------------------
        
        % evolve again
        [new_vault,pro2,result2]=hia1(ini_pop,t_x,t_y,vault,per_evo_time,inf);
        
        % refresh the vault
        vault=new_vault;
    % ---------------------------------
        timer=toc;
        fprintf('evolution time: %f\n',timer);
        pro(1:size(pro2,2),dynamic_counter)=pro2';
        pro(per_evo_time+2,dynamic_counter)=fitness(result2,te_x,te_y);
        
        result2=result3;
        init=zeros(size(result3,1),size(result3,2),40);
        init(:,1,:)=1;
        for i=1:30
            init(:,:,i)=result2;
        end
        for i=2:2:40
            init(:,2:size(result2,2),i)=init(:,2:4,1)+0*rand(size(init,1),size(result2,2)-1,1);
        end
        [pro2,result3]=sea(init,t_x,t_y,20);
        proo(1:size(pro2,2),dynamic_counter)=pro2';
        proo(per_evo_time+2,dynamic_counter)=fitness(result3,te_x,te_y);

        dynamic_counter=dynamic_counter+1;
    end
end
save('dynamic_result.mat');