Based on Brooks' subsumption architecture, we introduce a relational definition of a layered module architecture. We present two straightforward implementations following the declarative and the procedural meaning of such architectures. Based on these representations and the notion of state snapshots we then explain how to learn relational, discrete descriptions of asynchronous and possible non-deterministic behaviour. We conclude with an outlook of how relational learning can be used to optimise module behaviour, how to prune redundant connections and how to induce functionalities of higher level competence.

Affective human-computer interaction requires a system to identify a user's emotional state. Such systems mostly use facial expressions or speech signals to recognize emotion. Another approach is to use physiological data that is known to be correlated with psychological evidence. Using a few signals, signal processing allows one to derive a variety of features from which one needs to learn a classifier. A standard approach is to learn a classifier from a set of feature data together with labels that indicate a target class. This article shows that even simple a model of target labels may create a hard learning problem and why predictive accuracy of many different learning algorithms cannot be improved beyond a certain point. Subsequently, we present the rather underestimated approach of rough set data analysis to explain this result. Simultaneously, we are able to derive a classifier that reaches a top predictive accuracy with literally no additional assumptions on the data and only very weak biases.