def _update_particles(self):
"""
Update the velocities and positions for all particles.
This does not update the fitness for each particle.
"""
# Random values between zero and one. One random value per particle.
rand_g = tools.rand_uniform(size=self.num_particles)
# Update velocity for all particles using numpy operations.
# For an explanation of this formula, see the research papers referenced above.
# Note that self.best is the swarm's best-known position aka. global-best.
self.velocity = (self.omega * self.velocity.T \
+ self.phi_g * rand_g * (self.best - self.particle).T).T
# Fix de-normalized floating point values which can make the execution very slow.
self.velocity = tools.denormalize_trunc(self.velocity)
# Bound velocity.
self.velocity = tools.bound(self.velocity, self.velocity_lower_bound, self.velocity_upper_bound)
# Update particle positions in the search-space by adding the velocity.
self.particle = self.particle + self.velocity
# Bound particle position to search-space.
self.particle = tools.bound(self.particle, self.problem.lower_bound, self.problem.upper_bound)
def _update_particles(self):
"""
Update the velocities and positions for all particles.
This does not update the fitness for each particle.
"""
# Random values between zero and one. One random value per particle.
rand_g = tools.rand_uniform(size=self.num_particles)
# Update velocity for all particles using numpy operations.
# For an explanation of this formula, see the research papers referenced above.
# Note that self.best is the swarm's best-known position aka. global-best.
self.velocity = (self.omega * self.velocity.T \
+ self.phi_g * rand_g * (self.best - self.particle).T).T
# Fix de-normalized floating point values which can make the execution very slow.
self.velocity = tools.denormalize_trunc(self.velocity)
# Bound velocity.
self.velocity = tools.bound(self.velocity, self.velocity_lower_bound,
self.velocity_upper_bound)
# Update particle positions in the search-space by adding the velocity.
self.particle = self.particle + self.velocity
# Bound particle position to search-space.
self.particle = tools.bound(self.particle, self.problem.lower_bound,
self.problem.upper_bound)
def _new_agents(self):
"""
Calculate a new population of agents using the DE method.
The fitness is not calculated in this function.
"""
# Convenience variables.
population = self.population # 2-d array with the population of agents.
num_agents = self.num_agents # Number of agents in population.
dim = self.problem.dim # Search-space dimensionality.
differential_weight = self.differential_weight # Control parameter for DE.
crossover_probability = self.crossover_probability # Control parameter for DE.
# Create a new population of agents.
for i in range(num_agents):
# This loop only processes a single agent in each iteration.
# It might be possible to make a faster implementation using Numpy
# to process all agents simultaneously. However, this
# implementation is simpler and easier to understand, and the
# greatest runtime usage is typically in the fitness function.
# Pick random and distinct indices for agents in the population.
a, b, c, k = tools.rand_choice(num_agents, size=4, replace=False)
# Indices a, b, c, k are all different from each other.
# Now ensure that indices a, b, c are also different from i.
# Simply check if an index equals i and then replace it with k,
# which is different from i, a, b, c.
if i == a:
a = k
elif i == b:
b = k
elif i == c:
c = k
assert i != a != b != c
# Original position in the search-space for the agent to be updated.
original = population[i, :]
# Calculate crossover of randomly selected agents from the population.
crossover = population[a, :] + differential_weight * (population[b, :] - population[c, :])
# Create a new agent (i.e. position) in the search-space from the crossover.
# The crossover probability decides whether to use the crossover or
# the agent's original position in the search-space.
new_agent = np.where(tools.rand_uniform(dim) < crossover_probability, crossover, original)
# Ensure at least one element of the agent's new position is from the crossover.
rand_index = tools.rand_int(lower=0, upper=dim)
new_agent[rand_index] = crossover[rand_index]
# Bound the agent's new position to the search-space boundaries.
new_agent = tools.bound(x=new_agent,
lower=self.problem.lower_bound,
upper=self.problem.upper_bound)
# Assign the agent's new position to the temporary population so the
# fitness can be calculated in another function.
self.new_population[i, :] = new_agent
def _new_agents(self):
"""
Calculate a new population of agents using the DE method.
The fitness is not calculated in this function.
"""
# Convenience variables.
population = self.population # 2-d array with the population of agents.
num_agents = self.num_agents # Number of agents in population.
dim = self.problem.dim # Search-space dimensionality.
differential_weight = self.differential_weight # Control parameter for DE.
crossover_probability = self.crossover_probability # Control parameter for DE.
# Create a new population of agents.
for i in range(num_agents):
# This loop only processes a single agent in each iteration.
# It might be possible to make a faster implementation using Numpy
# to process all agents simultaneously. However, this
# implementation is simpler and easier to understand, and the
# greatest runtime usage is typically in the fitness function.
# Pick random and distinct indices for agents in the population.
a, b, c, k = tools.rand_choice(num_agents, size=4, replace=False)
# Indices a, b, c, k are all different from each other.
# Now ensure that indices a, b, c are also different from i.
# Simply check if an index equals i and then replace it with k,
# which is different from i, a, b, c.
if i == a:
a = k
elif i == b:
b = k
elif i == c:
c = k
assert i != a != b != c
# Original position in the search-space for the agent to be updated.
original = population[i, :]
# Calculate crossover of randomly selected agents from the population.
crossover = population[a, :] + differential_weight * (population[b, :] - population[c, :])
# Create a new agent (i.e. position) in the search-space from the crossover.
# The crossover probability decides whether to use the crossover or
# the agent's original position in the search-space.
new_agent = np.where(tools.rand_uniform(dim) < crossover_probability, crossover, original)
# Ensure at least one element of the agent's new position is from the crossover.
rand_index = tools.rand_int(lower=0, upper=dim)
new_agent[rand_index] = crossover[rand_index]
# Bound the agent's new position to the search-space boundaries.
new_agent = tools.bound(x=new_agent,
lower=self.problem.lower_bound,
upper=self.problem.upper_bound)
# Assign the agent's new position to the temporary population so the
# fitness can be calculated in another function.
self.new_population[i, :] = new_agent
