/Meta-Learning

This is the repository from my undeargrad thesis, available at:

Primary LanguagePython

Meta-Learning

This is the repository from my undergrad thesis. We used an evolutionary algorithm (Differential Evolution) for hyper-parameter optimization of five time series forecasting models, as ilustrated in the image:

alt text

Each model was optimized 30 times, resulting in 30 different models:

From the 95% confidence interval for the mean we compute this plots:

More details available at:

https://1drv.ms/u/s!ApdlapclXdxHtUwjAVYbPFle9fGe?e=Ba7eY0

For usage:

1 - pip install -r requirements

2 - let a file named series.csv in the directory. Note that we used 3 time series, so small modifications will be >necessary for single time series.

3 - run main.py

WARNING, for a 160 sample time series, using DE for an LSTM archtecture optimizaiton tooks rough one week.