Devrim-Celik / simple_echo_state_network

Echo state network framework, NARMA10 dataset generator as an demonstration supplied.

Echo State Network

In this repository you will find a very simply echo state network, which can be used as a scaffold for building a more sophisticated architecture. As a dataset, the model tries to approximate the NARMA10 series, which is generated by NARMA10.py and takes random noise as input.
The model can be found in the ESN.pyfile. The ESN_test.py file initializes an echo state network and fits it onto the NARMA10 series.
When execute, the script will save the plot of the reservoir activity (of neurons) and a plot of the prediction vs the real target values in the folder images. Furthermore, the complete reservoir history will be saved in a csv file in the folder csv_files.

---

Required Packages:

• Matplotlib
• Numpy
• Scipy

---

Getting Started:

After cloning this repository and installing the required packages, simply execute ESN_test.py. You can use the following flags:

Flags:

• -train_c or --training_cycles: specify number of training cycles (default 4000), e.g.:

python3 ESN_test.py -train_c 2000

• -test_c or --testing_cycles: specify number of testing cycles (default 1000), e.g.:

python3 ESN_test.py -test_c 500

• -rs or --resSize: specify reservoir size (default 1000), e.g.:

python3 ESN_test.py -rs 200

• -a or --alpha: specify leaking rate (default 0.8), e.g.:

python3 ESN_test.py -a 0.2

• -s or --show: show plots e.g.:

python3 ESN_test.py -s

---

NARMA10 Prediction

---

Reservoir Activity Sample