We describe a way to construct an artificial agent whose architecture is characterized by a number of local, reactive procedures controlled by an RNN, termed motivation unit network. This network is able to adopt various attractor states, or internal states, which are able to protect the complete system from sensory input not belonging to the current internal state. No strict hierarchy can be observed in this network. Instead, internal states may be represented by partly overlapping state vectors.
Where required, further procedures have been introduced that can be interpreted as forming explicit representations of parts of the environment. Specifically, an internal model of the agent’s own body is introduced that can, as a “manipulable” body-model, be used for planning new behaviors via internal simulation. Internal manipulation is possible because the body-model, like a marionette puppet, able to adopt all configurations the real body can assume. This expansion allows the agent to switch between reactive control and cognitive control (in the sense of McFarland & Bösser 1993).
When aiming to study higher mental properties, at least in human beings, we have to deal with the phenomenal aspect of these properties. A number of experimental results suggest that, i) some, but not all neuronal activities are, under specific—and unknown in any detail—conditions equipped with a phenomenal aspect, i.e., show subjective experience, but that ii) there is no specific function of this phenomenal aspect apart from the functions that can be ascribed to the physical properties of the system. Note that this does not mean that the phenomenal aspect has no function. Rather, a network adopts the function only when, at the same time, the phenomenal aspect is given. This view allows us to focus the analysis on the functional aspect of the procedure (see section 7). However, due to our lack of knowledge, as an external observer we cannot decide whether a given internal state is a mental state or not (if mental states are understood as internal states that are equipped with a phenomenal aspect).
The complete network represents a collection of hypotheses that can be tested by comparing their properties with experimental data and by trying to match them with theoretical concepts. Examples studied in this article concern behaviors that, for an external observer, may be conceptualized as various gait patterns, or navigation using an internal map, on the “lower” level. On a higher level, we deal with inventing new behaviors and planning ahead, as well as phenomena attributed to mental states like emotions, attention, intention, and volition. Last but not least we compare the properties of our approach with different aspects of consciousness, such as access consciousness (including global accessibility) and metacognition. We claim that, at least in their basic form, these phenomena can be attributed to internal states emerging from the cooperation of decentralized elements of our network.