Using Mixture Density Networks to model continuous action priors
No Thumbnail Available
Date
2019
Authors
Zwane, Sicelukwanda
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Given enough data and access to an environment, model-free deep reinforcement
learning algorithms allow for direct unsupervised behaviour
learning in an environment without relying on a model of the environment's
dynamics. As such, these algorithms have typically been favoured
when learning tasks de ned in continuous environments such as robotics
where such models are hard to come by. Despite their success here, these
approaches still su er from high sample complexity and other drawbacks
imposed by the use of the function approximators they rely on. This
includes poor generalisation across tasks, poor initial performance and
generally unsafe learning procedures due to inherent randomness in the
learning process where actions sampled from random policies are executed
for the purpose of exploration in the environment.
In this work we address these issues by presenting a method for transferring
knowledge from multiple, previously solved tasks (source tasks) to
new target tasks in the same continuous environment. Equipping successive
agents with this body of prior knowledge allows for easier acquisition
of subsequent behaviours in the given continuous environment. We also
empirically show that our proposed method \mixture density network action
priors" demonstrates improved safety as well as improved learning
compared to a standard model-free deep reinforcement learning agent in
continuous environments.
Description
A dissertation submitted for the degree
Master of Science
March 2019