Nemesis

Feed Forward Neural Networks for Reinforcement Learning (Soon to be extended)

Estimator

Estimator is a general interface for estimators. An estimator consist of:

• an input type and output type that define the expected type of input vector values and output vector values.
• an expected number of input and output attributes
• predict; an estimator should produce an output from a given input vector using predict
• fit; an estimator should be trained using fit

MLP

The MLP class is a Multi-Layer Perceptron. The MLP class is a subclass of Estimator.

A simple example of how to construct this class is this,

MLP<float, float, float> (1, 1)

where each float type defines the input type, output type, and weight type respectively for the network.

The (1, 1) constructor constructs a neural network expecting 1 input and 1 output.

learning_rate: a variable that can be accessed directly; it represents the learning rate of the MLP

decay_rate: a variable that can be accessed directly; it represents the learning rate of the MLP

Currently, an MLP can be saved and loaded to a json file using the MLP's load and save functions.

TDLearner

TDLearner is an interface for reinforcement learning agents using TD-learning methods.

It requires an Estimator to estimate and update the value of state action pairs.

set_value_estimator: replaces the current estimator with a new estimator

update_values function: updates the TD function for the state reward pair(s) (TODO: make update_values action_index a set)

get_value: gets the values for a state reward pair

Todo:

• remove nhlohmann's json file and add as a requirement