Convergence Issues with Particle Swarm Optimization

[SimonBlanke]

Following up on issue #84:

PSO fails to fully converge on simple convex problems.

Expected Behavior
PSO should converge effectively on simple convex optimization problems like the sphere function.

Actual Behavior
PSO fails to converge properly. PSO comes close the the global optimum but fails to fully converge.

Test Case

from gradient_free_optimizers import ParticleSwarmOptimizer

opt = ParticleSwarmOptimizer(search_space, population=20)
opt.search(sphere_function, n_iter=10000)
print(f"PSO Best score: {opt.best_score}")

[Soham-47]

Hi @SimonBlanke , maybe the issue is due to not explicitly setting the inertia, cognitive_weight (c1), and social_weight (c2) parameters? I trained a PSO with the relevant parameters and got these results:-

Code:

import numpy as np
from gradient_free_optimizers import ParticleSwarmOptimizer

def sphere_function(params):
    return -(params["x1"]**2 + params["x2"]**2)

search_space = {
    "x1": np.arange(-10, 10, 0.1),
    "x2": np.arange(-10, 10, 0.1),
}


optimizer = ParticleSwarmOptimizer(
    search_space=search_space,
    population=20,
    inertia=0.7,           
    cognitive_weight=1.5,   
    social_weight=1.5,          
    rand_rest_p=0.02       
)

optimizer.search(
    objective_function=sphere_function,
    n_iter=500
)

print("Best score:", optimizer.best_score)
print("Best parameters:", optimizer.best_para)

Results: 'sphere_function'
   Best score: -2.5243548967072378e-27
   Best parameter:
      'x1' : -3.552713678800501e-14
      'x2' : -3.552713678800501e-14

   Random seed: 1136852140

   Evaluation time   : 0 sec    [0.0 %]
   Optimization time : 0.04750561714172363 sec    [100.0 %]
   Iteration time    : 0.04750561714172363 sec    [10525.07 iter/sec]

Best score: -2.5243548967072378e-27
Best parameters: {'x1': np.float64(-3.552713678800501e-14), 'x2': np.float64(-3.552713678800501e-14)}

[SimonBlanke]

Hello @Soham-47,

thank you for taking interest in this bug issue. The optimizer implementations have changed to support mixed search-space dimensions (categorical and continuous). The optimizer classes now also follow a template method pattern, which overrides private methods of the core-optimizer. The new PSO implementation is here:
https://github.com/SimonBlanke/Gradient-Free-Optimizers/blob/main/src/gradient_free_optimizers/optimizers/pop_opt/particle_swarm_optimization.py

If you like you can investigate if this issue persists for this new implementation.