A bottom-up approach for cognitive control

Abstract

Cognitive skills including the ability to plan ahead require internal representations of the subject itself and its environment. The perspective on internal representations and assumptions on the importance of internal representations have changed over the last years: While traditional Artificial Intelligence tried to built intelligent systems relying solely on internal representations and working on a Knowledge level, behaviourist approaches tended to reject internal representations. Both approaches produced considerable results in different but complementing fields (examples are higher level tasks as mathematical proofs, or simple tasks for robots respectively). We believe that the most promising approach for more general tasks should connect both domains: For higher level and cognitive tasks internal representations are necessary or helpful. But these representations are not disconnected from the body and the lower-levels. The higher level is grounded in the lower levels and the robot is situated in an environment. While many projects today deal with the requirements of embodiment and situatedness, we focus on an approach like that of Verschure [55, 57, 58, 56]: The control system is built from the bottom up, growing towards higher levels and more complex tasks. The primitives of the different levels are constituted through the lower level and, as a main aspect, rely on neural networks. While Verschure only addressed the most relevant lower-level models in his work, we are constructing an architecture for cognitive control. The main idea for the cognitive control is the idea of mental simulation: Mental simulation sees planning as "probehandeln": This notion of trying a movement by mentally enacting it without performing the action physically relies strongly on the notion of an internal model. As a first step an internal model of the own body is constructed and used which later-on may be expanded to models of the environment. This body model is fully functional: it is constrained in the same way as the body itself. It can move and can be used in the same way as the body. Therefore, hypothetical movements can be tested on their consequences. For this purpose it must be possible to decouple the body itself from the action controlling modules to use the original controllers for control of the internal body model. Our approach shall address in particular (1) the structure and the construction of the mental models, (2) the process of using learnt behaviours to control the body or to modulate these behaviours in controlling the body or the internal model, (3) the decoupling of the body from the control structures to invoke the simulation, (4) the invention of new situation models and the decision when to construct new ones. An essential aspect of the approach proposed here is, apart from the idea that memory is composed of many individual situation models consisting of RNNs, that random selection of the connections is used, in this way introducing some kind of Darwinian aspect (see Edelmans Neural Darwinism). © Springer-Verlag Berlin Heidelberg 2009.