@mastersthesis {pub4504,
	title = {Development of a Simulation Framework for Deep Reinforcement Learning in the Context of ConceptFormation},
	author = {Florian Schulze},
	year = {2020},
	abstract = {As modern robotics finds its way more and more into the human society, the need
for a natural interaction with artificial agents increases. However, the human world
is very complex. The sheer complexity of human interaction cannot be reached with
standard robotic systems. These are often based on a finite set of rules. Anyway, finite
rules make communication with robots uninteresting and unnatural to humans, as
humans tend to find the limitation of such system fast.
Humans do not behave, reason, and think based on a finite set of rules. Instead they use
symbols e.g. language, to communicate and think. The field of semiotics tries to define
such symbols and provide explanations for the human thought process. It becomes
clear, that these semiotic symbols have very few in common with what is known as
symbol systems in the robotics and artificial intelligence (AI) community. It becomes
clear that the physical symbol system, which is primarily used in robotics, is not suitable
to represent the symbols humans, use in their society. Symbol emergence, the ability
to dynamically emerge and evolve a human-like symbol system, is a multi-faceted
problem. The field involves cognitive sciences, psychology, and behavioral sciences. As
reinforcement learning (RL) learning deals with behavior-learning and decision making,
which are important parts of cognitive systems. Therefore, RL problems can be used
to represent parts of a symbol emergence system and are well suited to investigate
abstraction processes. However, hardware is expensive and slow. Simulations on the
other hand are cheap and can be run at any speed. They are unbound to the real world and thus provide huge potential to research important aspects of decision
and behavior processes. This can also be used to shift the attention towards concept
formation. A list of tools and requirements to enable efficient concept formation is
proposed. This list serves as a base to evaluate whether state-of-the-art simulation
platforms qualify for a flexible concept formation research. Based on these findings new
RL simulation framework was introduced. The ConFormSim framework aims to provide
a fast simulation, which allows different levels of abstraction in the observation and
action space and fulfills the requirements for an efficient concept formation simulator.
Therefore, it introduces new feature to the existing Unity ML-Agents platform. The
experiments have proven, that these features provide a benefit on RL research and
facilitate the training of agents and the design of new environments. However, it was
also shown, that the new features might introduce performance issues to a RL system,
as the implementation is computationally demanding.},
	publisher = {Ilmenau University of technology},
	booktitle = {Ilmenau University of technology}
}