| Automatic Feature Selection in Neuroevolution
Feature selection is the process of finding the best set of inputs for a machine learning algorithm. Feature selection can be critical to performance, since excluding important features makes it impossible to find optimal solutions, while including superfluous features impedes learning by increasing the dimensionality of the problem. In supervised learning, many methods for automating feature selection already exist. However, few if any are applicable in reinforcement learning, since they typically assume the existence of labeled training data or a complete model by which candidate feature sets can be evaluated. The goal of this project is to develop a new method for automating feature selection in reinforcement learning problems using neuroevolution. Neuroevolution, in which evolutionary methods optimize populations of neural networks, can tackle difficult reinforcement learning problems. The most sophisticated methods can evolve, not only the weights of the networks, but their internal topologies too. This project aims to extend such methods to evolve the features that serve as inputs to the networks, yielding a feature selection method avoids the limitations of filters and wrappers and is consequently practical for reinforcement learning tasks.
This project involves formulating a new feature selection
methodology, identifying appropriate evaluation benchmarks, and
rigorously comparing to existing feature selection techniques. It
may also include the formulation of and comparison to feature
selection methods based on traditional temporal difference methods
for reinforcement learning.
|
|
| Keywords: |
Feature Selection, Reinforcement Learning, Neuroevolution Learning |
| Study: |
Artificial Intelligence |
| Contact: |
Shimon Whiteson |
| Location: |
Universiteit van Amsterdam |
| References: |
|