This work is an extension of the work done by Randall Beer and Brain Yamauchi in the paper Sequential Behavior and Learning in Evolved Dynamical Neural Networks (sorry, it's hidden behind a paywall!). We implemented our networks in a very similar way with some small modifications, and attempted to scale the result up in order to solve a useful task. The first task we have designed is to solve a non-trivial maze in the least amount of steps.
The full mathematical details of the algorithm are in the file docs/README.pdf.
As an example, the network was able to learn how to solve this maze in 14 decisions ‒ the optimal solution.
Open a terminal and navigate to where you'd like this program. Enter
git clone https://github.com/Garrett-R/evolved-dynamical-neural-network.git ednn
cd ednn
make
The binary will be called "ednn". You can see how to execute it by typing "./ednn --help". Since there are so many parameters required, we also supplied a helper script that sets the parameters and executes the program. To use just run
./run_ednn.sh
GNU GPL version 2 or any later version