Thoughts

This document contains some thoughts regarding what can be refined in the algorithms used in the pipe_breaks app.

Aritificial Neural Network

The model built by Sudharshan only uses a lot of Dense layers and does Binary Classification. The model architecture can be changed a little.

Tip: Could be interesting to find out, increasing depth of the network vs increasing the width?

Tree based models

• Grid Search and Random Search for hyperparameter tuning
• Need for better tracking of the experiments done with various algorithms
• Mostly update all the results in the PostgreSQL database

LSTM

• Single layer or 2 layer networks will not scale well
• Needs a better algorithm for forecasting
• Two/three possible solutions
  • Use a different library altogether, something like darts, sktime or prophet
  • Create a bigger model for forecasting, trying different loss functions for better peak prediction
  • Trying a tree based model like XGBoost for time series forecasting
• Problems to be addressed is that,
  • Need to find the solution for multivariate multistep time series forecasting (I think darts has a module for that)

For temporal data

• Need to revise the train_test_split as you can't use future data to predict past results

Comparison of results

• Need a nice layout for seeing test results and better comparison
• Take a look at the layouts from tensorboard and weights and biases
• Interesting pip package called stickyland for creating dashboards etc