as learning (Shanks, 1995) and stimulus control (see e.g. Enquist & Ghirlanda,
1998). In essence, they provide a potential common framework for understanding
and modelling behavioural mechanisms relevant for both simple as well as com-
plex organisms. Here we will focus on stimulus control. In his classic text-book
The Study of Instinct, Tinbergen (1951) devotes a significant part to how external
stimuli influence behaviour. Similarly, Hinde (1970) in his attempt to synthe-
size ethology and animal experimental psychology, dedicates several chapters to
stimulus control. Stimulus control has also been and still is a central subject in
experimental psychology (see e.g. Terrace, 1966; Mackintosh, 1974; Pearce et al.,
1997). One of the fundamental observations on how stimuli control response is
referred to as stimulus generalization in psychology (Guttman & Kalish, 1956;
Mackintosh, 1974). This refers to the fact that an organism responds in similar
ways to many variants of stimulation. For instance, if an animal has been trained
to react to a particular stimulus, it will also react to stimuli that are somewhat
different. The strength of response is often described by a generalization gradient
(over some stimulus dimension) with a maximum of responding usually at or near
the training stimulus.
The findings within ethology are very similar, although different concepts have
been used and the focus is not on learning but on stimuli that naturally occur in
the wild, for example social signals. Ethologists have studied the importance of
stimulus components and concluded that certain aspects often are more impor-
tant than others, and that naturally occurring stimuli can be stripped of many of
their components and still can be potent in eliciting the response (Hinde, 1970;
Baerends, 1982). In practise this is the same thing as generalization: the animal
does not only react to stimuli that occur in nature but also to variants which lack
certain aspects or have extra components added.
Both ethologists and experimental psychologists have also shown that certain
stimulus variants may even be more efficient than the naturally occurring ones or
the training stimuli. Within ethology this is called a supernormal stimulus (Tinber-
gen, 1951; Hinde, 1970), and psychologists talk about peak-shifts in responding
(Hanson, 1959; Mackintosh, 1974).
Empirical studies of stimulus control have resulted in a set of principles (e.g.
animals do generalize) which directly reflect observations, but there is very little
theory in the strictest sense. One exception is gradient-interaction theory (based
on Spence 1936, 1937; Hull 1943), which allows one to calculate response in
a complex situation given that one knows how the animal generalizes in all the
simple situations that build up the complex case. Note, however, that this model
cannot predict the degree of generalization in the individual cases: this has to
obtained in experiments.
Artificial neural networks have several advantages compared with traditional
theory of stimulus control. First, these models do not rest on empirical findings on