Evolving Minimalistic Control for Complex Behavior
Keywords
Loading...
Authors
Issue Date
2010-07-20
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Traditionally, intelligence was thought to be associated with thinking, reasoning,
planning etcetera. More recently there has been an increasing interest in ‘lowlevel’
behavioral responses. It is likely that people make use of a combination of
a fast, ‘lower’, automatic system, and an adaptive, ‘higher’, deliberative system.
How these systems are combined to produce behavior is uncertain however. One
theory suggests the existence of a minimalistic control system that enables complex
behavior by biasing the automatic system. Here this theory about how automatic
and deliberative structures can be integrated is investigated. This is done by simulating
the evolution of simple neural networks in an environment with two different
states, during which the appropriateness of an action may also be different. It is
expected that purely reactive agents are not able to perform optimally in such an
environment. The theory predicts that a control structure/deliberative system will
evolve that is inhibitory/modulatory in relation to the automatic system. The control
system does not need to be active all the time, only in situations for which the
automatic system alone is not adequate. In contrast to expectations, many evolved
networks did not evolve hidden units, indicating that the task may not be difficult
enough. However, the contextual input units evolved to regulate behavior in a manner
similar to the hypothesized control structure, supporting the theory that natural
control systems are minimalistic in nature. Further research to clarify the results is
suggested.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen