modules/evotorch/docs/quickstart.md

metadata

hide:
  - navigation

Quickstart

This section gives a brief overview of the EvoTorch API to solve a simple minimisation problem.

There are four main stages to using EvoTorch:

1. Creating a problem to solve.
2. Creating a searcher to optimise the problem.
3. Attaching loggers to the searcher
4. Running the algorithm

Let's start by importing the relevant packages.

from evotorch import Problem
from evotorch.algorithms import SNES
from evotorch.logging import StdOutLogger
import torch

Problem definition

For this simple example, we're going to consider the classic 'sphere' minimisation problem. The objective is to find a $d$-dimensional vector $x*$ that minimises

[ f(x) = \sum_{i=1}^d x_i^2. ]

Implementing this in PyTorch we have

def sphere(x: torch.Tensor) -> torch.Tensor:
    return torch.sum(x.pow(2.0))

To make this function visible to EvoTorch's algorithms, we simply wrap it up as a [Problem][evotorch.core.Problem] instance. To do this, we will need to specify that we want to minimise ("min") the function, and the solution_length is $d$, in this case $d=10$. We will also specify that the initial bounds for solutions is in the range $(-1, 1)$, so that our algorithm knows roughly where to start.

problem = Problem("min", sphere, solution_length=10, initial_bounds=(-1, 1))

Creating a searcher

Now we can search for solutions for the problem we've defined. In this example, we'll use the Separable Natural Evolution Strategies algorithm with default parameters, and we will specify that the initial standard deviation (scale) of the search distribution is 5 with stdev_init=5.

searcher = SNES(problem, stdev_init=5)

Attaching a logger

To keep an eye on what's happening as we run the algorithm, we'll also create a logger. In this case, we'll use the [StdOutLogger][evotorch.logging.StdOutLogger] which will print the status of the evolutionary algorithm to the terminal.

logger = StdOutLogger(searcher)

Running the searcher

Now we can run the algorithm for one iteration by calling the searcher.step() method.

searcher.step()

???+ abstract "Output" iter : 1 mean_eval : 268.27362060546875 median_eval : 269.3234558105469 pop_best_eval : 119.95197296142578 best_eval : 119.95197296142578 worst_eval : 473.68804931640625

Or if we want to, we can run it for as many iterations as we want using the searcher.run() method. Let's try running it for 3 iterations.

searcher.run(3)

???+ abstract "Output" ``` iter : 2 mean_eval : 244.1479034423828 median_eval : 223.21856689453125 pop_best_eval : 128.8501434326172 best_eval : 119.95197296142578 worst_eval : 473.68804931640625

        iter : 3
    mean_eval : 276.6123352050781
  median_eval : 207.94456481933594
pop_best_eval : 88.03515625
    best_eval : 88.03515625
  worst_eval : 688.3544921875

        iter : 4
    mean_eval : 284.18206787109375
  median_eval : 224.14187622070312
pop_best_eval : 83.8626937866211
    best_eval : 83.8626937866211
  worst_eval : 688.3544921875
```

A complete script

Putting everything together we have:

from evotorch import Problem
from evotorch.algorithms import SNES
from evotorch.logging import StdOutLogger
import torch


# Define a function to minimize
def sphere(x: torch.Tensor) -> torch.Tensor:
    return torch.sum(x.pow(2.0))


# Define a Problem instance wrapping the function
# Solutions have length 10
problem = Problem("min", sphere, solution_length=10, initial_bounds=(-1, 1))

# Instantiate a searcher
searcher = SNES(problem, stdev_init=5)

# Create a logger
logger = StdOutLogger(searcher)

# Evolve!
searcher.run(3)

Next steps

Now that you have completed your first evolutionary learning with EvoTorch, we recommend that you continue to our User Guide. Alternatively, you can take a look at our Examples if you want to dive into some advanced use-cases.