Applies the Differential evolution algorithm to minimize a function.
tfp.optimizer.differential_evolution_minimize( objective_function, initial_population=None, initial_position=None, population_size=50, population_stddev=1.0, max_iterations=100, func_tolerance=0, position_tolerance=1e-08, differential_weight=0.5, crossover_prob=0.9, seed=None, name=None )
Differential Evolution is an evolutionary optimization algorithm which works
on a set of candidate solutions called the population. It iteratively
improves the population by applying genetic operators of mutation and
recombination. The objective function
f supplies the fitness of each
candidate. A candidate
s_1 is considered better than
f(s_1) < f(s_2).
This method allows the user to either specify an initial population or a single candidate solution. If a single solution is specified, a population of the specified size is initialized by adding independent normal noise to the candidate solution.
The implementation also supports a multi-part specification of the state. For example, consider the objective function:
# x is a tensor of shape [n, m] while y is of shape [n]. def objective(x, y): return tf.math.reduce_sum(x ** 2, axis=-1) + y ** 2
The state in this case is specified by two input tensors
apply the algorithm to this objective function, one would need to specify
either an initial population as a list of two tensors of shapes
[population_size, k] and
[population_size]. The following code shows the
population_size = 40 # With an initial population and a multi-part state. initial_population = (tf.random.normal([population_size]), tf.random.normal([population_size])) def easom_fn(x, y): return -(tf.math.cos(x) * tf.math.cos(y) * tf.math.exp(-(x-np.pi)**2 - (y-np.pi)**2)) optim_results = tfp.optimizers.differential_evolution_minimize( easom_fn, initial_population=initial_population, seed=43210) print (optim_results.converged) print (optim_results.position) # Should be (close to) [pi, pi]. print (optim_results.objective_value) # Should be -1. # With a single starting point initial_position = (tf.constant(1.0), tf.constant(1.0)) optim_results = tfp.optimizers.differential_evolution_minimize( easom_fn, initial_position=initial_position, population_size=40, population_stddev=2.0, seed=43210)
objective_function: A Python callable that accepts a batch of possible solutions and returns the values of the objective function at those arguments as a rank 1 real
Tensor. This specifies the function to be minimized. The input to this callable may be either a single
Tensoror a Python
Tensors. The signature must match the format of the argument
population. (i.e. objective_function(*population) must return the value of the function to be minimized).
initial_population: A real
Tensoror Python list of
Tensors. If a list, each
Tensormust be of rank at least 1 and with a common first dimension. The first dimension indexes into the candidate solutions while the rest of the dimensions (if any) index into an individual solution. The size of the population must be at least 4. This is a requirement of the DE algorithm.
initial_position: A real
Tensorof any shape. The seed solution used to initialize the population of solutions. If this parameter is specified then
initial_populationmust not be specified.
population_size: A positive scalar int32
Tensorgreater than 4. The size of the population to evolve. This parameter is ignored if
initial_populationis specified. Default value: 50.
population_stddev: A positive scalar real
Tensorof the same dtype as
initial_position. This parameter is ignored if
initial_populationis specified. Used to generate the population from the
initial_positionby adding random normal noise with zero mean and the specified standard deviation. Default value: 1.0
max_iterations: Positive scalar int32
Tensor. The maximum number of generations to evolve the population for. Default value: 100
Tensorof the same dtype as the output of the
objective_function. The algorithm stops if the absolute difference between the largest and the smallest objective function value in the population is below this number. Default value: 0
Tensorof the same real dtype as
initial_population. The algorithm terminates if the largest absolute difference between the coordinates of the population members is below this threshold. Default value: 1e-8
differential_weight: Real scalar
Tensor. Must be positive and less than 2.0. The parameter controlling the strength of mutation in the algorithm. Default value: 0.5
crossover_prob: Real scalar
Tensor. Must be between 0 and 1. The probability of recombination per site. Default value: 0.9
intor None. The random seed for this
None, no seed is applied. Default value: None.
name: (Optional) Python str. The name prefixed to the ops created by this function. If not supplied, the default name 'differential_evolution_minimize' is used. Default value: None
optimizer_results: An object containing the following attributes: converged: Scalar boolean
Tensorindicating whether the minimum was found within the specified tolerances. num_objective_evaluations: The total number of objective evaluations performed. position: A
Tensorcontaining the best point found during the search. If the search converged, then this value is the argmin of the objective function within the specified tolerances. objective_value: A
Tensorcontaining the value of the objective function at the
position. If the search converged, then this is the (local) minimum of the objective function. final_population: The final state of the population. final_objective_values: The objective function evaluated at the final population. initial_population: The starting population. initial_objective_values: The objective function evaluated at the initial population. num_iterations: The number of iterations of the main algorithm body.
ValueError: If neither the initial population, nor the initial position are specified or if both are specified.