Intro

• Gradient-free optimization of neural network parameters

---

• Example: f(x) = x^2 (loss = log10(MSE(y_predict, x**2)))

• Genetic algorithm:

• AdamW (for comparison):

Usage

pip install grad-free-optim

import torch as tc
from pymoo.algorithms.soo.nonconvex.ga import GA as Algo
from trading_models.simple_models import MLP
from trading_models.utils import mod_keys, train_model

from grad_free_optim.pymoo import GradFreeOptim


def make_data():
    x = tc.rand((100, 1))
    yt = x**2
    return dict(x=x, yt=yt)


def loss_fn(data, mode):
    x, yt = data["x"], data["yt"]
    y = net(x)
    loss = tc.log10(((y - yt) ** 2).mean())
    info = {"SCORE": -loss}
    plots = {}
    return loss, mod_keys(info, mode), mod_keys(plots, mode)


tc.manual_seed(0)
net = MLP([1, 32, 32, 1])
train_data, test_data = make_data(), make_data()

if 1:
    bound = [-1.0, 1.0]
    opt = GradFreeOptim(bound, net, train_data, test_data, loss_fn)
    opt.run(Algo())
else:  # AdamW
    train_model(net, train_data, test_data, loss_fn)