In response to Siraj's Week #4 challege : Predict the magnitude of an earthquake, given the USGS dataset.
We are given a NN, modify it to predict the magnitudes of earthquakes. The base dataset
Steps:
-
Clean Data
-
Normalize Data -- Using minmax scaling z= (x-min(x))/(max(x)-min(x)) (Done in R)
-
Add extra features and repeat 1 & 2 (incomplete)
Neural Network source: https://github.com/sibyjackgrove/Earthquake_predict
-
Train
-
Predict -- Use random_search & grid_search for hyperparameter tuning (http://scikit-learn.org/stable/auto_examples/model_selection/randomized_search.html)
A challenge to myself is to increase the predictive capabalities of the net from the base USGS dataset, by adding the lunar-distance, and solar flare activity. But I'm having troubles...
https://www.fourmilab.ch/earthview/pacalc.html - lunar apogee/perigrees
According to USGS, a link between solar activity and earthquakes has not been established : https://www2.usgs.gov/faq/categories/9830/3282
However a look at the weights in the NN might be able to tell us if a machine-learning algorithm finds it to be relevant data for the prediction of earthquakes magnitues....