nevergrad meta issues a warning by default

[lesshaste]

I am putting this as a discussion as it is quite likely to be a user error on my part. My code:

import optimagic as om
import numpy as np

def three_minima_function(params):
    x, y = params
    res =  -2 * np.exp(-((x - 1) ** 2 + (y - 1) ** 2) / 0.2) \
        - 1.5 * np.exp(-((x + 1) ** 2 + (y + 1) ** 2) / 0.5) \
        - 1 * np.exp(-((x - 1) ** 2 + (y + 1) ** 2) / 1)
    res = np.ceil(res * 100) / 100
    return res



bounds = om.Bounds(lower=np.array([-5, -5]), upper=np.array([5, 5]))

algo = om.algos.nevergrad_meta
res=om.minimize(
    fun=three_minima_function,
    params=np.arange(2),
    algorithm=algo,
    bounds=bounds,
    algo_options={"stopping_maxfun": 50},
)
print(algo, three_minima_function(res.x), res.nfev)

This outputs:

UserWarning: COBYLA: Invalid RHOEND; it should be a positive number and RHOEND <= RHOBEG; it is set to 1e-06
  warn(f'{solver}: Invalid RHOEND; it should be a positive number and RHOEND <= RHOBEG; it is set to {rhoend}')

Should I have set a parameter to avoid this warning?

[timmens]

Hey, thanks for reaching out!

I tried to reproduce your warning but couldn't. Can you let me know which operating system you are using and what versions of Python, optimagic, and nevergrad you have installed?

[lesshaste]

om.__version__
Out[2]: '0.5.2'

nevergrad.__version__
Out[4]: '1.0.12'

Python 3.12.8

Ubuntu 24.04.2 LTS

[timmens]

Thanks! I've managed to reproduce it on Python 3.12. It does not reproduce on Python 3.10!

1. As far as I can tell, you cannot govern these parameters via optimagic. It would be very interesting to see whether you can reproduce the warnings (and no warnings in Python 3.10) by using nevergrad directly. By default, optimagics nevergrad_meta uses the Shiwa optimizer.
2. Does this influence the optimization result?

[lesshaste]

To answer question 1, I tried:

import numpy as np
import nevergrad as ng
from nevergrad.optimization.optimizerlib import Shiwa

def three_minima_function(params):
    x, y = params
    res = -2 * np.exp(-((x - 1) ** 2 + (y - 1) ** 2) / 0.2) - 1.5 * np.exp(-((x + 1) ** 2 + (y + 1) ** 2) / 0.5) - 1 * np.exp(-((x - 1) ** 2 + (y + 1) ** 2) / 1)
    res = np.ceil(res * 100) / 100
    return res

# Define the parameter space with bounds [-5, 5] for both x and y
parametrization = ng.p.Array(shape=(2,)).set_bounds(lower=[-5, -5], upper=[5, 5])

# Create the Shiwa optimizer
optimizer = Shiwa(parametrization=parametrization, budget=10000)

# Run the optimization
recommendation = optimizer.minimize(three_minima_function)

# Get the optimal parameters and function value
optimal_params = recommendation.value
optimal_value = three_minima_function(optimal_params)

print(f"Optimal parameters: x = {optimal_params[0]:.6f}, y = {optimal_params[1]:.6f}")
print(f"Optimal function value: {optimal_value:.6f}")

# Optional: Run multiple times to check consistency
print("\nRunning multiple optimizations for comparison:")
for i in range(5):
    optimizer = Shiwa(parametrization=parametrization, budget=10000)
    recommendation = optimizer.minimize(three_minima_function)
    params = recommendation.value
    value = three_minima_function(params)
    print(f"Run {i+1}: x = {params[0]:.4f}, y = {params[1]:.4f}, f(x,y) = {value:.4f}")

It doesn't print any warnings. It also doesn't do a great job at optimising. The optimal value is -2.01.

[lesshaste]

Actually, if you reduce the budget to 59 you will see the warning:

UserWarning: COBYLA: Invalid RHOEND; it should be a positive number and RHOEND <= RHOBEG; it is set to 1e-06
warn(f'{solver}: Invalid RHOEND; it should be a positive number and RHOEND <= RHOBEG; it is set to {rhoend}')

[timmens]

Thanks! Very interesting. In this case I would suggest asking for help / guidance from the nevergrad developers. It seems there is little we can do from the optimagic side. If you agree, you can close this discussion.

Thanks again!