Crate evo_rl 

Evo RL is a machine learning library built on the concept of Neuroevolution – evolving an architecture for neural networks as opposed to pre-specifying it. This library is best suited for Reinforcement Learning tasks in which a reward (or fitness) score can be assigned to agents.

Evo RL is a WIP and is in the pre-alpha state.

Modules

agent_wrapper

A wrapper class for neural network that creates an interface to other environments such as Python

api

doctest

This module contains constants and examples used in unit tests.

enecode

Module which specifies the genetic encoding for neural network and neural properties. The genetic encoding is used to reconstruct neural network graphs and for recombination of offspring.

graph

This module implements neural networks in the form of directed graphs (from petgraph). Evolutionary changes are projected onto the graph first before being encoded genetically.

neuron

Module for the neuron struct which defines the individual unit of computation for the neural network.

population

This module focuses on implementing an evolutionary algorithm for neural network optimization. It uses Stochastic Universal Sampling (SUS) and Truncation for selection within a population of neural network agents.

Functions

ez_input

Convenience function to easily create a vector of owned Strings.

hash_em

Convenience function to easily create maps when specifying gene strings.

increment_innovation_number

Increments the ID of a neuron when creating a daughter

progenitor_code

Returns progenitor code base for an innnovation number

relu

rng_box

Returns an rng based on Option

setup_logger

Utility function for logging unit tests

sigmoid

sort_genes_by_neuron_type

Sorts topology gene in the order of Input, Hidden, Output and then by innovation number