Global Search Algorithms:

Searching over the entire feasible set

1
Global Search Algorithms
• When we have local minima then the gradient based methods can fail as they can get stuck in those
local minima point.

• Instead of gradient information we can use a stochastic or random search where even if we reach a
minimum point we will get out to a worse solution and search again.

• This way we can make sure that we are in fact looking for a global minima and get unstuck from
any local minima point.

2
Simulated Annealing
•

3
Simulated Annealing
•

4
Simulated Annealing
•

5
Simulated Annealing
•

6
Simulated Annealing
•

7
Simulated Annealing
•

8
Simulated Annealing: Example
•

9
10
Simulated Annealing: Example
%%%%%% Intialize
x1 = linspace(-5,5,1000);
x2 = linspace(-5,5,1000);

gamma = 2;
iter = 5000;

x_val(1,:) = [-1 1];

i=0;
x_best= x_val(1,:);
f_best = obj(x_best(1,:));

11
Simulated Annealing: Example
for k=1:iter
%%%%% Generate neighborhood point z(k,:) at random for x_val(k,:) over the entire search space
z(k,1) = -5 + (5-(-5))*rand(1,1);
z(k,2) = -5 + (5-(-5))*rand(1,1);
fz(k) = obj(z(k,:)); %%%% Objective function at z
fx(k) = obj(x_val(k,:)); %%%% Objective function at x_val
T(k) = gamma/(log(k+2)); %%%% Temperature: Cooling as the iteration progresses

if fz(k)<fx(k) %%%%% If neighbourhood point z gives better performance replace x_val with z
x_val(k+1,:) = z(k,:);
fx(k+1) = fz(k);

12
Simulated Annealing: Example
else
p_val(k,:) = prob([fz(k),fx(k),T(k)]); %%%% If not replace with the probability (coin toss)
if p_val(k,:)>rand(1,1)
x_val(k+1,:) = z(k,:); fx(k+1) = fz(k);
else
x_val(k+1,:) = x_val(k,:); fx(k+1) = fx(k); %% If coin toss does not allow then keep as it is
end
end
if fx(k)<f_best
x_best = x_val(k,:); %%%% Keep track of the best result so far
f_best = fx(k);
end
end

13
Simulated Annealing: Example
%%%% Objective function
function f_val = obj(x)
% f_val =
3*(1-x(1)^2)*exp(-x(1)^2-(x(2)+1)^2)-10*(x(1)/5-x(1)^3-x(2)^5)*exp(-x(1)^2-x(2)^2)-(exp(-(x(1)+1)^2-x(2)^2))/3;
% f_val = (x(1)-3)^2+(x(2)+2)^2+x(1)*x(2);
f_val = 0.2*x(1)^2-0.1*cos(3*pi*x(1))+0.2*x(2)^2-0.1*cos(4*pi*x(2));
end

%%%% Probability function

end

14
Simulated Annealing: Example
f_best = -0.1615

x_best = [0.0114 0.0706]

At k = 2165

15
Particle Swarm Optimization (PSO)
• Presented by James Kennedy (a social psychologist) and Russell C. Eberhart (an engineer) in 1995.

• Instead of updating a single candidate we update a population or swarm of particles.

• We start with a randomly selected swarm where the particles interact with each other and move toward a
common solution (optimal value).

• Each particle keeps track of its best-so-far position—this is the best position it has visited so far (with
respect to the value of the objective function) and it is called a personal best (pbest).

• In contrast, the overall best-so-far position (best among all the positions encountered so far by the entire
population) is called a global best (gbest).

16
Particle Swarm Optimization (PSO)
•

17
Particle Swarm Optimization (PSO)
•

18
Particle Swarm Optimization (PSO)
•

19
Particle Swarm Optimization (PSO)
•

20
PSO: Example
•

21
PSO: Example
%%%%% Intialize
x1 = linspace(-5,5,1000);
x2 = linspace(-5,5,1000);
d = 10; % Particles
iter = 1000;
va = -0.1;
vb = 0.1;
x_1(1,:) = -5 + (5-(-5))*rand(10,1); %%% Intialize
x_2(1,:) = -5 + (5-(-5))*rand(10,1);
v1(1,:) = va + (vb-va)*rand(10,1); %%% Intialize
v2(1,:) = va + (vb-va)*rand(10,1);
w = 0.9;
c1 = 2; c2 = 2;

22
PSO: Example
%%%% Intial p_best and g_best selection
p1_best(1,:) = x_1(1,:);
p2_best(1,:) = x_2(1,:);
f(1) = obj([x_1(1,1),x_2(1,1)]);

for i=2:d
f(i) = obj([x_1(1,i),x_2(1,i)]);
if f(i-1)<f(i)
gbest(1,:) = [x_1(1,i-1),x_2(1,i-1)];
else
gbest(1,:) = [x_1(1,i),x_2(1,i)];
end
end

23
PSO: Example
%%%%% Update the velocities and positions
for k=1:iter
r = rand(1,10);
s = rand(1,10);

fx(k) = obj([x_1(k,i) x_2(k,i)]);

v1(k+1,:) = w*v1(k,:)+c1*r.*(p1_best(k,:)-x_1(k,:))+c2*s.*(gbest(k,1)*ones([1,10]) -x_1(k,:));

x_1(k+1,:) = x_1(k,:)+v1(k+1,:);
x_2(k+1,:) = x_2(k,:)+v2(k+1,:);

24
PSO: Example
%%%% For each generted particle limit it within the search space limits
for i=1:d

if abs(x_1(k+1,i))>5
x_1(k+1,i)=5*sign(x_1(k+1,i));
end
if abs(x_2(k+1,i))>5
x_2(k+1,i)=5*sign(x_2(k+1,i));
end

25
PSO: Example
%%%% Select p_best for iteration k
if obj([x_1(k+1,i) x_2(k+1,i)])<obj([x_1(k,i) x_2(k,i)])
p1_best(k+1,i) = x_1(k+1,i);
p2_best(k+1,i) = x_2(k+1,i);
else
p1_best(k+1,i) = p1_best(k,i);
p2_best(k+1,i) = p2_best(k,i);
end

26
PSO: Example
%%%% Select gbest for iteration k and end the loop
if obj([p1_best(k+1,i) p2_best(k+1,i)])<obj([gbest(k,1) gbest(k,2)])
gbest(k+1,:) = [p1_best(k+1,i) p2_best(k+1,i)];
fbest = obj([gbest(k+1,1) gbest(k+1,2)]);
k_best = k;
else
gbest(k+1,:) = gbest(k,:);
fbest = obj([gbest(k,1) gbest(k,2)]);
k_best = k;
end
end

end

27
PSO: Example

gbest = [-0.0091 0.0053]

fbest = -0.1994

28