A python machine learning toolkit for economists.
econlearn contains a number of supervised learning (regression), unsupervised learning and reinforcement learning algorithms, suitable for problems with large sample sizes but relatively few dimensions (large N, small D).
econlearn was developed to solve problems that arise in economics, particularly stochastic dynamic optimisation problems (Markov Decision Processes MDPs). A key feature is the batch reinforcement learning method 'Fitted QV iteration with tile coding', for solving MDPs by simulation (described in more detail in my thesis).
econlearn is designed to be fast. All algorithms are implemented in Cython and several make use of multi-threading. Many of the methods make use of a tile coding (i.e., overlapping grid) data structure. For pure speed of both fitting and prediction, tile coding is hard to beat. In higher dimensions tile coding can run into memory problems. econlearn includes a simple implementation of 'hashing' to help limit memory usage.
- Tile coding
- Radial Basis Functions
- Local quadratic
- Distance weighted OLS
- Sample grid (distance)
- Sample grid (tile coding)
- Density estimation (tile coding)
- Approximate nearest neighbors (tile coding)
- Fitted QV Iteration with Tile coding
This code requires installation of python with packages: cython, numpy, and scikit-learn. A good option is to install Anaconda.
Next download or clone this repository. Then you need to compile the Cython modules. To do this navigate to your local econlearn directory and run the setup.py file
python setup.py build_ext --inplace
This code has been developed and tested on linux machines only. In windows a compiler with openMP support will be required. If I get a chance I will provide more instructions on installation under Windows later.
econlearn is a work in progress. So far only the tile coding methods are available. In time I will hopefully add more algorithms including: distance based sample grid, radial basis function regression, distance weighted regression and local quadratic regression.
While not documented, the underlying Cython modules may be of use to those wanting to access these learning methods quickly within larger applications / models. Including those wanting to implement the reinforcement learning methods in a multi-agent simulation model.
The documentation is also a work in progress. The next job is to add some examples.