layg 0.0.1.dev0

Learn-as-you-go emulator with error estimation

Python implementation of the Learn As You Go algorithm published in http://arxiv.org/abs/arXiv:1506.01079. TODO: reference ICLR paper

The package defines a decorator that can be applied to functions to convert them to functions which learn outputs as they go and emulate the true function when expected errors are low. Two emulators are included: the k-nearest neighbors Monte Carlo accelerator described there, and a simple neural network.

The basic usage of the emulator code is something like this:

@emulate(CholeskyNnEmulator)
def loglike(x):
    """
    Your complex and expensive function here
    """
    return -np.dot(x,x)

This decorates the function loglike so that it is an instance of the Learner class. It can be used similarly to the original function: just call it as loglike(x). The __call__(x) method hides some extra complexity: it uses the Learn As You Go emulation scheme.

It learns both the output of loglike(x) and the difference between the emulator and the true value of loglike(x) so that we can make a prediction for the error residuals. We then put a cutoff on the amount of error that one will allow for any local evaluation of the target function. Any call to the emulator that has a too-large error will be discarded and the actual function loglike(x) defined above will be evaluated instead.

The logic for generating training sets and returning a value from either the true function or the emulated function are contained in the Learner class. The Learner class relies on an emulator class to do the emulation.

You can define you own emulator. Define a class that inherits from BaseEmulator and define two methods on it: set_emul_func(self, x_train: np.ndarray, y_train: np.ndarray) and set_emul_error_func(self, x_train: np.ndarray, y_train: np.ndarray) that set functions for, respectively, self.emul_func and self.emul_error_func. An example of this definition is provided.

Installation

There are a small number of python dependencies. If you use anaconda you can create an appropriate environment and install to your python path by running

conda env create --file environment.yml
pip install -e .

from this directory.

The pytorch dependency is only needed if you are using the neural network emulator or running the associated tests.