Even though Walknet is set up as a fixed structure consisting of hard-wired connections of the RNN, it can flexibly adapt to disturbances in the environment as needed during, for instance, crossing large gaps (Bläsing 2006). Nonetheless, the system might of course run into novel situations that require an even higher degree of adaption, and as such will require novel behaviors. As an example, think of a situation in which all the legs except the right hind leg are in the anterior part of the working range. When the right hind leg is forced to lift from the ground as it approaches a position very far to the rear, the whole system will become unstable, as the center of gravity is positioned very far towards the rear of the animal. In this case, the center of gravity would not be supported by the other legs, nor by the right hind leg that tries to start a swing movement. As a consequence, the agent would fall over, backwards. This problem could be detected by “problem detectors”, e.g., specific sensory input that reacts to the specific load distribution (a different solution is explained in section 8). In order to overcome this problem, the system would have to break out of its usual pattern of behavioral selection and try to select a different behavioral module that is usually not applicable in the given context. For instance, making a step backward with the right middle leg would be a possible solution, as this would provide support for the body and would afterwards allow going back to the normal walking behavior and the subsequent swing movement of the right hind leg. Usually, backward steps can only be selected in the context of backward walking.
Figure 6 shows an expansion that allows the system to search for solutions that are not connected to the current context. This expansion is termed the “attention controller”. We introduce a third layer of units (figure 6, in green), that is essentially a recurrent winner-take-all network (WTA-net). For each motivation unit there is a corresponding partner unit in this WTA-network. Currently-active motivation units suppress their winner-take-all (WTA) partner units (T-shaped connections in figure 6). Therefore, a random activation of this WTA-net will lead to the activation of one single unit not belonging to the currently- activated context. The random activation will be induced by another parallel layer, the “Spreading Activation Layer” (not depicted in figure 6, further details are described in Schilling & Cruse submitted). The winning unit of the WTA layer than activates its corresponding motivation unit. This triggers the connected behavior that can be tested as a solution to the problem at hand. The network follows a trial-and-error strategy as observed in, e.g., insects.
As has been proposed (Schilling & Cruse 2008), a further expansion of the system that is, most probably, not given in insects is not the testing of a behavior in reality, but instead the application of a newly-selected behavior on the body-model and the use of the model instead of the real body. The motor output is routed to the body-model instead of to the real body, and the real body is decoupled from the control system while testing new behaviors. Due to the predictive nature of the body-model, it can be used to predict possible consequences and to afterwards decide if a behavior solves the current problem and should be tried out on the real body. This procedure is called internal simulation and requires the introduction of switches that reroute motor output signals from the real body to the body model (figure 6, switch SW). Only after a successful internal simulation will the behavior be applied to the real body. McFarland & Bösser (1993) defined a cognitive system as a system that has the ability of planning ahead, i.e., that is able to perform internal simulations in order to predict possible outcomes of behaviors. Therefore, this latter expansion would make the control system cognitive (for details see Cruse & Schilling 2010b).