Date of Award

Spring 2019

Project Type


Program or Major

Computer Science

Degree Name

Master of Science

First Advisor

Marek Petrik

Second Advisor

Wheeler Ruml

Third Advisor

Laura Dietz


Solving reinforcement learning problems using value function approximation requires having good state features, but constructing them manually is often difficult or impossible. We propose Fast Feature Selection (FFS), a new method for automatically constructing good features in problems with high-dimensional state spaces but low-rank dynamics. Such problems are common when, for example, controlling simple dynamic systems using direct visual observations with states represented by raw images. FFS relies on domain samples and singular value decomposition to construct features that can be used to approximate the optimal value function well. Compared with earlier methods, such as LFD, FFS is simpler and enjoys better theoretical performance guarantees. Our experimental results show that our approach is also more stable, computes better solutions, and can be faster when compared with prior work.