Skip to content

Supervised Neuroevolution

Supervised Neuroevolution is the application of Neuroevolution to supervised learning. Typically, a supervised learning setting has a set of example inputs \(X = x_1 \dots x_N\) and a set of example outputs \(Y = y_1 \dots y_n\) and the objective is to find a function \(f\) that minimizes a well-defined loss function \(L\) across the dataset \((X, Y)\):

\[L*(f, X, Y) = \sum_{i=1}^N L(f(x_i), y_i)\]

In Supervised Neuroevolution, \(f\) is a neural network which is optimized via neuroevolution.

Overview of SupervisedNE

EvoTorch provides direct support for Supervised Neuroevolution throough the SupervisedNE class. Consider a dataset, \(N=100\), generated from the function \(y = x_1 + x_2\) by sampling \(x_1, x_2\) from the standard normal distribution:

import torch
from torch import nn

N = 100
X = torch.randn((N, 2))
Y = X.sum(dim=-1, keepdim=True)

This can be wrapped up as a torch.utils.data.TensorDataset instance for convenience:

from torch.utils.data import TensorDataset, DataLoader

train_dataset = TensorDataset(X, Y)

Creating a SupervisedNE instance for this dataset is straightforward:

from evotorch.neuroevolution import SupervisedNE

sum_of_problem = SupervisedNE(
    dataset=train_dataset,  # Use the training dataset generated earlier
    network=nn.Sequential(nn.Linear(2, 32), nn.ReLU(), nn.Linear(32, 1)),  # Simple MLP
    minibatch_size=32,  # Solutions will be evaluated on minibatches of size 32
    loss_func=nn.MSELoss(),  # Solutions will be evaluated using MSELoss
)

Then each network evaluated by sum_of_problem will be assigned a fitness based on how well it minimizes MSELoss on a 32 samples drawn from the train_dataset. Training the simple MLP will show a clear progress::

from evotorch.algorithms import SNES
from evotorch.logging import PandasLogger

searcher = SNES(sum_of_problem, popsize=50, radius_init=2.25)
logger = PandasLogger(searcher)
searcher.run(500)
logger.to_dataframe().mean_eval.plot()
Output

Unless your data is incompatible with torch.utils.data.DataLoader and/or the notation that minibatches drawn from the dataloader consist of data and targets:

X, Y = minibatch

then SupervisedNE should work in most cases. If not, then you can create custom functionality wherever you see fit:

class CustomSupervisedNE(SupervisedNE):
    def _make_dataloader(self) -> DataLoader:
        # Override to generate a custom dataloader
        ...

    def _loss(self, y_hat: Any, y: Any) -> Union[float, torch.Tensor]:
        # Override to define a custom loss function on network output yhat vs. target output y
        ...

    def _evaluate_using_minibatch(
        self, network: nn.Module, batch: Any
    ) -> Union[float, torch.Tensor]:
        # Override to modify how a network is evaluated on a minibatch
        ...

    def _evaluate_network(self, network: nn.Module) -> torch.Tensor:
        # Override to completely change how a network is evaluated
        ...

Manipulating Minibatches

SupervisedNE includes some particular features to manipulate how networks are evaluated on minibatches. These are:

  • minibatch_size: int, which defines the size of the minibatch that each network is evaluated on. This argument is passed as the batch_size when the SupervsiedNE instance instantiates its dataloader.
  • num_minibatches: int, which defines the number of minibatches that each network is evaluated on. Each minibatch will have size minibatch_size. This argument is useful in conjunction with minibatch_size, for example, when the loss_func is a non-linear function of the minibatch_size, or the GPU memory does not permit a larger minibatch_size, or some other reason that minibatch_size must take a particular value. In any of these cases, setting num_minibatches > 1 allows you to repeatedly evaluate each network on different minibatches, with the overall loss (fitness) averaged across the minibatches.
  • common_minibatch: bool, which specifies whether the same minibatch(es) should be used to evaluate all solutions when a SolutionBatch instance is passed to the SupervisedNE instance's evaluate method. As noted in recent work, it is sometimes more effective to evaluate all solutions on the same sets of minibatches, as this may reduce noise, for example, when approximating gradients in distribution-based evolution strategies. This is particularly true when common_minibatch = True is used in conjunction with num_actors > 1, as each actor will evaluate its sub-population on its own fixed set of minibatches.

For example:

  • minibatch_size = 16, num_minibatches = 2, common_minibatch = False will mean that each network is evaluated on its own set of 2 minibatches of size 16, with the loss averaged across these 2 minibatches.
  • minibatch_size = 64, num_minibatches = 1, common_minibatch = True, will mean that every network described by a SolutionBatch passed to the evaluate method will be evaluated on the same minibatch of size 16.
  • minibatch_size = 4, num_minibatches = 8, common_minibatch = True, num_actors = 16, will mean that there will be 16 actors, each of which will evaluate a sub-population of the SolutionBatch passed to the evaluate method. Each of the 16 actors will generate 8 minibatches of size 4, and will use those 8 minibatches to evaluate all of the solutions in its assigned sub-population.