Indago 0.2.1

Numerical optimization framework

Indago

Indago is a Python 3 module for numerical optimization.

Indago containts several modern algorithms for real fitness function optimization over a real parameter domain. It was developed at the Department for Fluid Mechanics and Computational Engineering of the University of Rijeka, Faculty of Engineering, by Stefan Ivić, Siniša Družeta, and others.

Indago is developed for in-house research and teaching purposes and is not officially supported in any way, comes with no guarantees whatsoever and is not properly documented. However, we use it regulary and it seems to be working fine. Hopefully you will find it useful as well.

Important: After every Indago update please check this document since Indago methods and APIs can undergo significant changes at any time.

Installation

For easiest install use

pip3 install indago

or if you wish to update your existing Indago installation

pip3 install indago --upgrade

Dependencies

The following packages should be installed using aptitude:

• python3
• python3-pip
• python3-tk

sudo apt install python3 python3-pip python3-tk

After packages installation using above command, additional python packages should be installed using pip from requirements.txt:

pip install -r requirements.txt

Optimization problem setup

Using Indago is easy. The setup of the optimization problem in Indago is the same regardless of the used optimization algorithm (a.k.a. optimizer):

# Evaluation function
def evaluation(x):
    obj = np.sum(x ** 2)  # minimization objective
    constr1 = x[0] - x[1]  # constraint x_0 - x_1 <= 0
    constr2 = - np.sum(x)  # constraint sum x_i >= 0
    return obj, constr1, constr2

# Initialize the chosen algorithm
from indago import PSO # ...or any other Indago algorithm
optimizer = PSO()

# Optimization variables settings
optimizer.dimensions = 10 # number of variables (i.e. size of design vector x)
optimizer.lb = -10 # lower bound, given as scalar (equal for all variables)
optimizer.ub = 10 + np.range(pso.dimensions) # upper bounds, given as np.array (one bound value per variable)

# Set evaluation function
optimizer.evaluation_function = evaluation  

# Objectives and constraints settings
optimizer.objectives = 1  # number of objectives (optional parameter, default objectives=1), this is obj in evaluation function
optimizer.objective_labels = ['Squared sum minimization']  # labels for objectives (optional parameter, used in reporting)
optimizer.constraints = 2  # number of constraints (optional parameter, default constraints=0), these are constr1 and constr2 in evaluation function
optimizer.constraint_labels = ['Constraint 1', 'Constraint 2']  # labels for constraints (optional parameter, used in reporting)

# Print optimizer parameters
print(optimizer) # not necessary, but useful for checking the setup of the optimizer

# Run optimization
result = optimizer.optimize() # (using default algorithm parameters)

# Extract results
print(result.f) # minimum of obj with constr1 and constr2 satisfied
print(result.X) # design vector at minimum

Algorithms

As of now, Indago consists of several stochastic optimizers:

• Particle Swarm Optimization (PSO)
• Fireworks Algorithm (FWA)
• Squirrel Search Algorithm (SSA)
• Differential Evolution (DE)
• Bat Algorithm (BA)
• Electromagnetic Field Optimization (EFO)
• Manta Ray Foraging Optimization (MRFO)

These algorithms are available through a unified API, which was designed to be as accessible as possible. Indago relies heavily on NumPy, so the inputs and outputs of the optimizers are mostly NumPy arrays. Besides NumPy and a couple of other stuff here and there (a few SciPy functions and rich module for fancy monitoring), Indago is pure Python. Indago optimizers also include some of our original research improvements, so feel free to try those as well. And don't forget to cite. :)

Particle Swarm Optimization

Let us use PSO as a primary step-by-step example. First, we need to import NumPy and Indago PSO, and then initialize an optimizer object:

import numpy as np
from indago import PSO
pso = PSO() 

Then, we must provide a goal function which needs to be minimized, say:

def goalfun(x):	# must take 1d np.array
    return np.sum(x**2) # must return scalar number
pso.evaluation_function = goalfun

Now we can define optimizer inputs:

pso.method = 'Vanilla' # we will use Standard PSO, the other available option is 'TVAC' [1]; default method='Vanilla'
pso.dimensions = 20 # number of variables in the design vector (x)
pso.lb = np.ones(pso.dimensions) * -1 # 1d np.array of lower bound values (if scalar value is given, it will automatically be transformed to 1d np.array of size dimensions, filled with the value)
pso.ub = np.ones(pso.dimensions) * 1 # 1d np.array of upper bound values (if scalar value is given, it will automatically be transformed to 1d np.array of size dimensions, filled with the value)
pso.iterations = 1000 # default iterations=100*dimensions
pso.maximum_evaluations = 5000 # optional maximum allowed number of function evaluations; when surpassed, optimization is stopped (if reached before pso.iterations are exhausted)
pso.target_fitness = 10**-3 # optional fitness threshold; when reached, optimization is stopped (if it didn't already stop due to exhausted pso.iterations or pso.maximum_evaluations)

Also, we can provide optimization method parameters:

pso.params['swarm_size'] = 15 # number of PSO particles; default swarm_size=dimensions
pso.params['inertia'] = 0.8 # PSO parameter known as inertia weight w (should range from 0.5 to 1.0), the other available options are 'LDIW' (w linearly decreasing from 1.0 to 0.4) and 'anakatabatic'; default inertia=0.72
pso.params['cognitive_rate'] = 1.0 # PSO parameter also known as c1 (should range from 0.0 to 2.0); default cognitive_rate=1.0
pso.params['social_rate'] = 1.0 # PSO parameter also known as c2 (should range from 0.0 to 2.0); default social_rate=1.0

If we want to use our novel adaptive inertia weight technique [2], we invoke it by:

pso.params['inertia'] = 'anakatabatic'

and then we need to also specify the anakatabatic model:

pso.params['akb_model'] = 'Languid' # [3,4], other options are 'FlyingStork', 'MessyTie', 'RightwardPeaks', 'OrigamiSnake' [2]

We can enable reporting during the optimization process by providing the monitoring argument:

pso.monitoring = 'basic' # other options are 'none' and 'dashboard'; default monitoring='none'

Finally, we can start the optimization and retrieve the results:

result = pso.optimize()
min_f = result.f # fitness at minimum, scalar number
x_min = result.X # design vector at minimum, 1d np.array

And that's it!

Fireworks Algorithm

If we want to use FWA [5], we just have to import it instead of PSO:

from indago import FWA
fwa = FWA()

Now we can proceed in the same manner as with PSO.

For FWA, the only method available is basic FWA, which is implemented in two versions: 'Vanilla' (ignores contraints) and 'Rank' (supports using constraints):

fwa.method = 'Rank' # the other option is 'Vanilla' which does not support constraints; default method='Rank'

In FWA we can set the following method parameters:

fwa.params['n'] = 20 # default n=dimensions
fwa.params['m1'] = 10 # default m1=dimensions/2
fwa.params['m2'] = 10 # default m2=dimensions/2

Squirrel Search Algorithm

We can try our luck also with SSA [6]. We initialize it like this:

from indago import SSA
ssa = SSA()

In SSA, the only available method is 'Vanilla' (which is set as default), and there is only one important method parameter:

ssa.params['acorn_tree_attraction'] = 0.6 # ranges from 0.0 to 1.0; default acorn_tree_attraction=0.5

If we want to fine-tune the algorithm, we can define a few other SSA parameters:

ssa.params['predator_presence_probability'] = 0.1 # default
ssa.params['gliding_constant'] = 1.9 # default
ssa.params['gliding_distance_limits'] = [0.5, 1.11] # default

Differential Evolution

If we want to use DE [7], we initialize it in the same way as with the other methods:

from indago import DE
de = DE()

There are two DE methods implemented, namely SHADE and LSHADE. Say we want to use LSHADE:

de.method = 'LSHADE' # default method='SHADE'

Both DE methods use the following parameters:

de.params['initial_population_size'] = 200 # default initial_population_size=dimensions*18
de.params['external_archive_size_factor'] = 2.6 # default
de.params['historical_memory_size'] = 4 # default historical_memory_size=6
de.params['p_mutation'] = 0.2 # default p_mutation=0.11

DE implementation does not (yet) support using constraints.

Bat Algorithm

For using BA [8], we initialize it in the same way as with the other methods:

from indago import BA
ba = BA()

The only BA version implemented is the original Bat Algorithm [8] with mutation modified to make it fitness-scalable. We specifiy it as:

ba.method = 'Vanilla'

The following parameters are used:

ba.params['bat_swarm_size'] = 15 # default bat_swarm_size=dimensions
ba.params['loudness'] = 1 # default
ba.params['pulse_rate'] = 0.001 # default
ba.params['alpha'] = 0.9 # default 
ba.params['gamma'] = 0.1 # default 
ba.params['freq_range'] = [0, 1] # default

Electromagnetic Field Optimization

To use EFO [9], we initialize it in the same way as with the other methods:

from indago import EFO
efo = EFO()

In EFO, the only available method is 'Vanilla' (which is set as default):

efo.method = 'Vanilla'

The following parameters are used:

efo.params['population_size'] =  100 # default population_size=10*dimensions
efo.params['R_rate'] = 0.25 # should range from 0.1 to 0.4; default R_rate=0.25
efo.params['Ps_rate'] = 0.25 # should range from 0.1 to 0.4; default Ps_rate=0.25
efo.params['P_field'] = 0.075 # should range from 0.05 to 0.1; default P_field=0.075
efo.params['N_field'] = 0.45 # should range from 0.4 to 0.5; default N_field=0.45

Currently, parallelization in EFO is not allowed due to it being entirely ineffective for this method.

Manta Ray Foraging Optimization

If we want to use MRFO [10], we initialize it in the same way as with the other methods:

from indago import MRFO
mrfo = MRFO()

In MRFO, the only available method is 'Vanilla' (set as default):

mrfo.method = 'Vanilla'

The following parameters are used:

mrfo.params['manta_population'] = 3 # default 
mrfo.params['somersault_factor'] = 2 # default (added for experimentation purposes, probably best be left at default)

Multiple objectives and constraints handling

The optimization algorithms implemented in Indago are able to consider nonlinear constraints defined as c(x) <= 0. The constraints handling is enabled by the multi-level comparison which is able to compare multi-constraint solution candidates.

Multi-objective optimization problems can also be treated in Indago by automatically constructed weighted sum fitness, hence reducing the problem to single-objective.

The following example prepares a PSO optimizer for an evaluation which returns two objectives and two constraints:

pso.objectives = 2
pso.objective_labels = ['Route length', 'Passing time']
pso.objective_weights = [0.4, 0.6]
pso.constraints = 2
pso.constraint_labels = ['Obstacles intersection length', 'Curvature limit']

The evaluation function needs to be modified accordingly:

def evaluate(x):
    # Calculate minimization objectives o1 and o2
    # Calculate constraints c1 and c2
    # Constraints are defined as c1 <= 0 and c2 <= 0
    return o1, o2, c1, c2

Stopping criteria

Five distinct criteria can be enabled for stopping the Indago optimization:

• Stop when reached maximum number of iterations (optimizer.iterations),
• Stop when reached maximum number of evaluations (optimizer.maximum_evaluations),
• Stop when reached target fitness (optimizer.target_fitness), and
• Stop when reached maximum number of iterations with no progress (optimizer.maximum_stalled_iterations)
• Stop when reached maximum number of evaluations with no progress (optimizer.maximum_stalled_evaluations)

The optimization stops when any of the specified criteria is reached. The maximum number of iterations (optimizer.iterations) is a mandatory stopping condition. If not set by the user, it is automatically set to the default value calculated as iterations = 100 * dimensions. Maximum number of evaluations, target fitness criterion and stall criteria are enabled only if they are specified by the user. Stopping criteria can be used in any combination.

Optimization monitoring

Three different modes of optimization process monitoring can be used by specifiying the parameter optimizer.monitoring. The available options are:

• 'none' - no output is displayed (this is the default behavior),
• 'basic' - one line of output per iteration is provided, comprising basic convergence parameters, and
• 'dashboard' - a live dashboard is shown, featuring progress bars and continuously updated values of parameters most important for tracking optimization convergence.

Parallel evaluation

Indago is able to evaluate a group of solution candidates (e.g. a swarm in PSO) in parallel mode. This is especially useful for expensive (in terms of computational time) engineering problems whose evaluation relies on simulations such as CFD or FEM.

Indago utilizes the multiprocessing module for parallelization and it can be enabled by specifying the number_of_processes parameter:

optimizer.number_of_processes = 4 # use 'maximum' for employing all available processors/cores

Note that the implemented parallelization scales well only on relatively slow goal functions. Also keep in mind that Python multiprocessing sometimes does not work when initiated from imported code, so you need to have the optimization run call (optimizer.optimize()) wrapped in if __name__ == '__main__':.

When dealing with numerical simulations, one mostly needs to specify input files and a directory in which the simulation runs. If execution is performed in parallel, these file/directory names need to be unique to avoid possible conflicts in simulation files. In order to facilitate this, Indago offers the option of passing a unique string over to the evaluation function, thus enabling execution of simulations without no conflicts.

To enable the passing of a unique string to evaluation function, set forward_unique_str to True:

optimizer.forward_unique_str = True

Note that the evaluation function needs an additional argument through which the unique string is received:

def evaluation(X, unique_str=None):
    # Prepare a simulation case in a new file and/or a new directory with names based on unique_str
    # Run external simulation and extract results
    return objective

Failing evaluation function

Sometimes, for whatever reason, goal function (i.e. optimizer.evaluation_function) may fail to compute. Indago features a built-in (semi-experimental) scheme for handling such cases, which are identified by evaluation function returning np.nan. You can control this scheme by setting the optimizer.eval_fail_behavior to one of the following:

• 'abort' - optimization is stopped at the first event of evaluation function returning np.nan (default)
• 'ignore' - optimizer will ignore any np.nan values returned by the evaluation function and (note that Vanilla FWA does not support this)
• 'retry' - optimizer will try to resolve the issue by repeatedly receding a failed design vector a small step towards the best solution thus far

When using optimizer.eval_fail_behavior = 'retry' the retrying mechanism can be fine-tuned by setting additional parameters:

optimizer.eval_retry_attempts = 5 # retry at most 5 times to evaluate the unevaluated design vector; default eval_retry_attempts=10
optimizer.eval_retry_recede = 0.05 # at each retry move the unevaluated design vector 5% towards the hitherto best solution; any value in range (0,1) is allowed; default eval_retry_recede=0.01

Note that setting optimizer.eval_retry_recede = 0 yields pure evaluation retries without design vector modification, which might be useful for randomly failing fitness functions.

Failed evaluations are detected by optimizer.evaluation_function returning np.nan. Thus the function should be prepared in such a way so that it returns np.nan if it fails to compute. However, if you want Indago to handle this for you, you can enable

optimizer.safe_evaluation = True # default safe_evaluation=False

Although this treatment of failing evaluation functions will probably solve the problem, it may also hamper the efficiency of the optimization algorithm. Therefore keep in mind that it is always better to make sure that your goal function never fails to compute.

Results and convergence plot

Some intermediate optimization results are stored in optimizer.results object, which can be explored/analyzed after the optimization is finished.

Also, a utility function is available for visualizing optimization convergence, which produces convergence plots for all defined objectives and constraints:

optimizer.results.plot_convergence()

CEC 2014

Indago also includes the CEC 2014 test suite [11], comprising 30 test functions for real optimization methods. You can use it by importing it like this:

from indago.benchmarks import CEC2014

Then, you have to initialize it for a specific dimensionality of the test functions:

test = CEC2014(20) # initialization od 20-dimension functions, you can also use 10, 50 and 100

Now you can use specific test functions (test.F1, test.F2, ... up to test.F30), they all take 1d np.array of size 10/20/50/100 and return a scalar number. Alternatively, you can iterate through the built-in list of them all:

test_results = []
for f in test.functions:
    optimizer.evaluation_function = f
    test_results.append(optimizer.optimize().f)

Have fun!

References:

1. Ratnaweera, A., Halgamuge, S. K., & Watson, H. C. (2004). Self-organizing hierarchical particle swarm optimizer with time-varying acceleration coefficients. IEEE Transactions on evolutionary computation, 8(3), 240-255.

2. Družeta, S., & Ivić, S. (2020). Anakatabatic Inertia: Particle-wise Adaptive Inertia for PSO, arXiv:2008.00979 [cs.NE].

3. Družeta, S., & Ivić, S. (2017). Examination of benefits of personal fitness improvement dependent inertia for Particle Swarm Optimization. Soft Computing, 21(12), 3387-3400.

4. Družeta, S., Ivić, S., Grbčić, L., & Lučin, I. (2019). Introducing languid particle dynamics to a selection of PSO variants. Egyptian Informatics Journal, 21(2), 119-129.

5. Tan, Y., & Zhu, Y. (2010, June). Fireworks algorithm for optimization. In International conference in swarm intelligence (pp. 355-364). Springer, Berlin, Heidelberg.

6. Jain, M., Singh, V., & Rani, A. (2019). A novel nature-inspired algorithm for optimization: Squirrel search algorithm. Swarm and evolutionary computation, 44, 148-175.

7. Tanabe, R., & Fukunaga, A. S. (2014). Improving the search performance of SHADE using linear population size reduction. In Proceedings of the 2014 IEEE Congress on Evolutionary Computation (CEC), pp. 1658–1665, Beijing, China.

8. Yang, X. S., & Gandomi, A. H. (2012). Bat algorithm: a novel approach for global engineering optimization. Engineering computations.

9. Abedinpourshotorban, H., Shamsuddin, S. M., Beheshti, Z., & Jawawi, D. N. (2016). Electromagnetic field optimization: a physics-inspired metaheuristic optimization algorithm. Swarm and Evolutionary Computation, 26, 8-22.

10. Zhao, W., Zhang, Z., & Wang, L. (2020). Manta ray foraging optimization: An effective bio-inspired optimizer for engineering applications. Engineering Applications of Artificial Intelligence, 87, 103300.

11. Liang, J. J., Qu, B. Y., & Suganthan, P. N. (2013). Problem definitions and evaluation criteria for the CEC 2014 special session and competition on single objective real-parameter numerical optimization. Computational Intelligence Laboratory, Zhengzhou University, Zhengzhou China and Technical Report, Nanyang Technological University, Singapore, 635.