<p>This thesis aims to evaluate four artificial neural network architectures, each of which implements the sensory-motor mapping in an embodied, situated, and autonomous agent set up to reach a goal area in one out of six systematically varied T-maze environments. In order to reach the goal the agent has to turn either to the left or to the right in each junction in the environment, depending on the placement of previously encountered light sources. The evaluation is broken down into (i) measuring the reliability of the agents' capacity to repeatedly reach the goal area, (ii) analyzing how the agents work, and (iii) comparing the results to related work on the problem.</p><p>Each T-maze constitutes an instance of a broad class of problems known as delayed response tasks, which are characterized by a significant (and typically varying) delay between a stimulus and the corresponding appropriate response. This thesis expands this notion to include, besides simple tasks, repeated and multiple delayed response tasks. In repeated tasks, the agent faces several stimulus-delay-response sequences after each other. In multiple tasks, the agent faces several stimuli before the delay and the corresponding appropriate responses. Even if simple at an abstract level, these tasks raise some of the fundamental issues within cognitive science and artificial intelligence such as whether or not an internal objective world model is necessary and/or suitable to achieve the appropriate behavior. For such reasons, these problems also constitute an interesting base for evaluating alternative ideas within these fields.</p><p>The work leads to several interesting insights. Firstly, purely reactive controllers (as represented by a feed-forward network) may be sufficient, in interaction with the environment, to solve both simple and repeated delayed response tasks. Secondly, an extended sequential cascaded network that selectively replaces its own sensory-motor mapping achieves significantly better performance than the other networks. This indicates that selective replacement of the sensory-motor mapping may be more powerful than both modulation (as represented by a simple recurrent network) and replacement in each step (as represented by a standard sequential cascaded network). Thirdly, this thesis demonstrates that even reactive controllers may contribute to behavior, which, from an observer's point of view, may seem to require an internal rational capacity, i.e. the ability to represent and explore alternatives internally.</p>

This thesis aims to evaluate four artificial neural network architectures, each of which implements the sensory-motor mapping in an embodied, situated, and autonomous agent set up to reach a goal area in one out of six systematically varied T-maze environments. In order to reach the goal the agent has to turn either to the left or to the right in each junction in the environment, depending on the placement of previously encountered light sources. The evaluation is broken down into (i) measuring the reliability of the agents' capacity to repeatedly reach the goal area, (ii) analyzing how the agents work, and (iii) comparing the results to related work on the problem. Each T-maze constitutes an instance of a broad class of problems known as delayed response tasks, which are characterized by a significant (and typically varying) delay between a stimulus and the corresponding appropriate response. This thesis expands this notion to include, besides simple tasks, repeated and multiple delayed response tasks. In repeated tasks, the agent faces several stimulus-delay-response sequences after each other. In multiple tasks, the agent faces several stimuli before the delay and the corresponding appropriate responses. Even if simple at an abstract level, these tasks raise some of the fundamental issues within cognitive science and artificial intelligence such as whether or not an internal objective world model is necessary and/or suitable to achieve the appropriate behavior. For such reasons, these problems also constitute an interesting base for evaluating alternative ideas within these fields. The work leads to several interesting insights. Firstly, purely reactive controllers (as represented by a feed-forward network) may be sufficient, in interaction with the environment, to solve both simple and repeated delayed response tasks. Secondly, an extended sequential cascaded network that selectively replaces its own sensory-motor mapping achieves significantly better performance than the other networks. This indicates that selective replacement of the sensory-motor mapping may be more powerful than both modulation (as represented by a simple recurrent network) and replacement in each step (as represented by a standard sequential cascaded network). Thirdly, this thesis demonstrates that even reactive controllers may contribute to behavior, which, from an observer's point of view, may seem to require an internal rational capacity, i.e. the ability to represent and explore alternatives internally.