Using Neuroevolution with gradient descent to construct and train logic circuits

Overview

This was an extension of a previous project I did where I used neuroevolution to build neural network (NN) agents to play OpenAI games. However, here, I made it so the weights aren't searched for randomly, they're trained via gradient descent, which speeds things up a ton.

In addition, I made it so a successful NN can be packaged into an "atom", to be inserted in other, higher level NNs. This part was less successful, but still has promise.

A long summary of the results can be found on my blog here.

Main run scripts

These are the main .py scripts that I use, in the main directory for simplicity. They are:

• hyper_evo.py - For running scripts.

Classes

These are the classes that the run scripts use, which live in the classes directory. Here, they're listed in order from low to high level.

Software:

• Atom.py - Represents a single discrete entity, either a simple classic NN node, or a more complex, functional module atom.
• HyperEPANN.py - Assembles Atom objects, handles propagation through the NN, and other stuff.
• EvoAgent.py - Handles the NN, and interfaces with whatever environment you want to use it with (OpenAI gym, or your own).
• Population.py - For building a populatinon of EvoAgent objects. Also has multi_run_hist() for running the same evolution scenario multiple times, to generate statistics.