A neural network framework for combining different task types and motivations in motivated reinforcement learning

Download files
Access & Terms of Use
open access
Copyright: Ismail, Hafsa
Altmetric
Abstract
Combining different motivation models for different task types within artificial agents has the potential to produce agents capable of a greater range of behaviours in different situations. This thesis will explore the ability to produce agents that can identify different types of tasks online then learn solutions for these tasks while driven by one or more motivations in isolation or combination to produce different agent behaviour that will focus on different task types. The main contributions of this research are: Definitions for four task types for use in motivated reinforcement learning agents; a multi-layer neural network model for combining different motivations for different task types; a study of agent ability to identify different types of tasks in a simulated game scenario. A review of existing computational models of motivation and the layered approaches to the design of goal generation and intrinsically motivated learning agents inspire the multi-layer neural network framework presented and examined experimentally in this thesis. The first set of experiments examines topological and non-topological adaptive resonance theory networks that are used to classify sensed states and identify different types of tasks. Different distance metrics and weight update rules are used to determine the impact of the network structure on identified task types. The most promising network structure, distance metric and update rule is then chosen to implement the task identification process in conjunction with motivated reinforcement learning. In the second set of experiments, three motivation models, novelty, interest and competence, are examined separately in conjunction with our task identification algorithm. Established metrics and statistical methods are used to quantify learning behaviour and the complexity and variety of repetitive behavioural structures. Finally, we examine agent behaviour when the agent can identify more than one type of task with a single motive model that drives the learning process. Parallel and series task types combination modes are explored. Results show that a series approach to changing the type of task the agent can identify is most effective for producing motivated reinforcement learning agents that can both learn effectively and exhibit quantifiable changes in their characteristic learning behaviour over time.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Ismail, Hafsa
Supervisor(s)
Merrick, Kathryn
Barlow, Michael
Creator(s)
Editor(s)
Translator(s)
Curator(s)
Designer(s)
Arranger(s)
Composer(s)
Recordist(s)
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
2014
Resource Type
Thesis
Degree Type
Masters Thesis
UNSW Faculty
Files
download public version.pdf 3.68 MB Adobe Portable Document Format
Related dataset(s)