From Tutorial on Neural Systems Modeling, Chapter 7:
Simultaneous perturbative reinforcement learning is effective not only in two-layered but also in three-layered neural networks. Input-hidden, hidden-output, and bias weights can all be perturbed simultaneously. We will use perturbative reinforcement learning (the directed drift algorithm) in this section to show how two different input signals are distributed over the hidden units in a three-layered, feedforward neural network. Specifically, we will use perturbative reinforcement learning to reproduce the results on the formation of a non-uniform distributed representation that we obtained in Chapter 6 using back-propagation. This simulation was a simplified version of a neural network model of distributed parallel processing in the vestibulo-oculomotor system (see Chapter 6).
Researchers should cite this work as follows:
Tutorial on Neural Systems Modeling, Copyright 2010 Sinauer Associates Inc.
Author: Thomas J. Anastasio
AbderRahman N Sobh, Jessica S Johnson, NanoBio Node (2014), "[Illinois]: Perturbative Reinforcement Learning to Develop Distributed Representations," https://nanohub.org/resources/pertdistrep. (DOI: 10.4231/D3FF3M12N).