头脑风暴算法(Brain Storming Optimization Algorithm, BSO)是一种新型群智能优化算法,该算法灵感来源于头脑风暴法,当一群人围绕一个特定的兴趣领域产生新观点的时候,这种情境就叫做头脑风暴。头脑风暴法是一种充分开发人类创造性思维解决问题的方法,具有进化能力强、搜索速度快、寻优能力强的特点。该成果于2011年发表在知名SCI期刊Advances in Swarm Intelligence上,目前在谷歌学术上被引773次。
在一个头脑风暴的过程中,通常会涉及到三种人,即主持人、问题负责人和一组头脑风暴人员。主持人主要是为了引导头脑风暴过程顺利进行而不带偏见,这意味着一个有主持经验但不了解问题的主持人将是首选;头脑风暴小组是指在主持人的引导下,通过头脑风暴过程实际产生想法的一群人;问题所有者负责在每一轮创意生成中挑选他们认为更好的创意。一般来说,每一个头脑风暴的过程都要经过三轮的想法产生过程,因为经过三轮的想法产生过程,人就会筋疲力尽,效率也就不高了。
1、算法原理
(1)头脑风暴阶段
在BSO算法中,一个人代表一个想法,而这个想法又代表了待解决问题的潜在解决方案。类似于其他群体智能算法,均匀随机初始化是首选,特别是当没有先验知识是已知的问题要解决的。在原始的BSO中,通过根据以下公式将高斯随机值添加到选定的个体来生成新的个体:
其中,xi new表示新生成的个体的第i个维度,可以看作是产生的新想法;xi selected表示基于当前个体群体的所选个体的第i个维度,可以看作是原来的想法;random()是一个随机函数,它是原BSO中的高斯随机函数。
在原始BSO中,系数ξ(t)是加权高斯随机值对新生成值的贡献的步长,ξ(t)值由下式求得
其中,logsig()为对数sigmoid传递函数,T为最大迭代次数,t为当前迭代次数,k是控制logsig( )斜率的系数,rand()返回(0,1)范围内的随机值。
其中,原来的想法xi selected可以根据以下公式获得
其中,xi selected为xold1和xold2的加权和,rand为随机产生的0到1之间的系数,xold1和xold2是分别从两个集群中选取的想法,通过这个随机值实现两个个体想法的交互和融合。
(2)算法伪代码:
1.初始化种群规模I和维度dim、最大迭代次数1
2.初始化参数和种群 Xi(i = 1,2,..., )
3. While t < T do
4. 利用 k-means 聚类算法将 n 个个体聚类为 m个聚类
5. 计算每个个体的适应度值并排名,将最佳个体记录为每个聚类中的聚类中心
6. If rand < 0.2 then
7. 随机生成一个个体来替换随机选取的一个聚类的中心
8. End If
9. For i = 1 to N do
10. If rand < 0.8 then
11. 选取一个聚类的中心或随机个体
12. Else
13. 选取两个聚类的中心或随机个体
14. End lf
15. 通过式(1)产生新的个体,计算适应度值
16. 将新生成的个体与现有个体进行比较,将保留较好的个体
17. End For
18. End While
19. 输出最终解
MATLAB核心代码
function best_fitness = bso(fun,n_p,n_d,n_c,rang_l,rang_r,max_iteration)
% fun = fitness_function
% n_p; population size
% n_d; number of dimension
% n_c: number of clusters
% rang_l; left boundary of the dynamic range
% rang_r; right boundary of the dynamic range
prob_one_cluster = 0.8; % probability for select one cluster to form new individual;
stepSize = ones(1,n_d); % effecting the step size of generating new individuals by adding random values
popu = rang_l + (rang_r - rang_l) * rand(n_p,n_d); % initialize the population of individuals
popu_sorted = rang_l + (rang_r - rang_l) * rand(n_p,n_d); % initialize the population of individuals sorted according to clusters
n_iteration = 0; % current iteration number
% initialize cluster probability to be zeros
prob = zeros(n_c,1);
best = zeros(n_c,1); % index of best individual in each cluster
centers = rang_l + (rang_r - rang_l) * rand(n_c,n_d); % initialize best individual in each cluster
centers_copy = rang_l + (rang_r - rang_l) * rand(n_c,n_d); % initialize best individual-COPY in each cluster FOR the purpose of introduce random best
best_fitness = 1000000*ones(max_iteration,1);
fitness_popu = 1000000*ones(n_p,1); % store fitness value for each individual
fitness_popu_sorted = 1000000*ones(n_p,1); % store fitness value for each sorted individual
indi_temp = zeros(1,n_d); % store temperary individual
% calculate fitness for each individual in the initialized population
for idx = 1:n_p
fitness_popu(idx,1) = fun(popu(idx,:));
end
while n_iteration < max_iteration
cluster = kmeans(popu, n_c,'Distance','cityblock','Start',centers,'EmptyAction','singleton'); % k-mean cluster
% clustering
fit_values = inf*ones(n_c,1); % assign a initial big fitness value as best fitness for each cluster in minimization problems
number_in_cluster = zeros(n_c,1); % initialize 0 individual in each cluster
for idx = 1:n_p
number_in_cluster(cluster(idx,1),1)= number_in_cluster(cluster(idx,1),1) + 1;
% find the best individual in each cluster
if fit_values(cluster(idx,1),1) > fitness_popu(idx,1) % minimization
fit_values(cluster(idx,1),1) = fitness_popu(idx,1);
best(cluster(idx,1),1) = idx;
end
end
% form population sorted according to clusters
counter_cluster = zeros(n_c,1); % initialize cluster counter to be 0
acculate_num_cluster = zeros(n_c,1); % initialize accumulated number of individuals in previous clusters
for idx =2:n_c
acculate_num_cluster(idx,1) = acculate_num_cluster((idx-1),1) + number_in_cluster((idx-1),1);
end
%start form sorted population
for idx = 1:n_p
counter_cluster(cluster(idx,1),1) = counter_cluster(cluster(idx,1),1) + 1 ;
temIdx = acculate_num_cluster(cluster(idx,1),1) + counter_cluster(cluster(idx,1),1);
popu_sorted(temIdx,:) = popu(idx,:);
fitness_popu_sorted(temIdx,1) = fitness_popu(idx,1);
end
% record the best individual in each cluster
for idx = 1:n_c
centers(idx,:) = popu(best(idx,1),:);
end
centers_copy = centers; % make a copy
if (rand() < 0.2) % select one cluster center to be replaced by a randomly generated center
cenIdx = ceil(rand()*n_c);
centers(cenIdx,:) = rang_l + (rang_r - rang_l) * rand(1,n_d);
end
% calculate cluster probabilities based on number of individuals in
% each cluster
for idx = 1:n_c
prob(idx,1) = number_in_cluster(idx,1)/n_p;
if idx > 1
prob(idx,1) = prob(idx,1) + prob(idx-1,1);
end
end
% generate n_p new individuals by adding Gaussian random values
for idx = 1:n_p
r_1 = rand(); % probability for select one cluster to form new individual
if r_1 < prob_one_cluster % select one cluster
r = rand();
for idj = 1:n_c
if r < prob(idj,1)
if rand() < 0.4 % use the center
indi_temp(1,:) = centers(idj,:);
else % use one randomly selected cluster
indi_1 = acculate_num_cluster(idj,1) + ceil(rand() * number_in_cluster(idj,1));
indi_temp(1,:) = popu_sorted(indi_1,:);
end
break
end
end
else % select two clusters
% pick two clusters
cluster_1 = ceil(rand() * n_c);
indi_1 = acculate_num_cluster(cluster_1,1) + ceil(rand() * number_in_cluster(cluster_1,1));
cluster_2 = ceil(rand() * n_c);
indi_2 = acculate_num_cluster(cluster_2,1) + ceil(rand() * number_in_cluster(cluster_2,1));
tem = rand();
if rand() < 0.5 %use center
indi_temp(1,:) = tem * centers(cluster_1,:) + (1-tem) * centers(cluster_2,:);
else % use randomly selected individuals from each cluster
indi_temp(1,:) = tem * popu_sorted(indi_1,:) + (1-tem) * popu_sorted(indi_2,:);
end
end
stepSize = logsig(((0.5*max_iteration - n_iteration)/20)) * rand(1,n_d);
indi_temp(1,:) = indi_temp(1,:) + stepSize .* normrnd(0,1,1,n_d);
% if better than the previous one, replace it
fv = fun(indi_temp);
if fv < fitness_popu(idx,1) % better than the previous one, replace
fitness_popu(idx,1) = fv;
popu(idx,:) = indi_temp(1,:);
end
end
% keep the best for each cluster
for idx = 1:n_c
popu(best(idx,1),:) = centers_copy(idx,:);
fitness_popu(best(idx,1),1) = fit_values(idx,1);
end
n_iteration = n_iteration +1;
% record the best fitness in each iteration
best_fitness(n_iteration, 1) = min(fit_values);
end
参考文献
[1] Shi Y. Brain storm optimization algorithm[C]//Advances in Swarm Intelligence: Second International Conference, ICSI 2011, Chongqing, China, June 12-15, 2011, Proceedings, Part I 2. Springer Berlin Heidelberg, 2011: 303-309.
完整代码获取方式:后台回复关键字:
TGDM100