Prediction as Internal Simulation: Taking Chances in What to Do Next

how the connections can be modulated and are flexible in different contexts. We approach these connections of internal representation through models of minimal cognitive systems in a bottom-up approach (Schilling and Cruse, submitted; Schilling and Cruse, 2008). An early neural network model of the own body is grounded in a biological-inspired framework that controls walking in a hexapod robot. While this internal model is very simple, it is quite flexible and can be recruited in service for multiple functions (Schilling, 2011). Importantly, the predictive capabilities of the body model allow applying the model in internal simulation. In internal simulation alternative possible behaviors can be tried out and the predicted consequences can be evaluated without actually carrying out probably dangerous behaviors (Hesslow, 2002). The system can plan ahead and becomes cognitive in the sense of McFarland and Bösser (1993). Interestingly, the behavioral system, which allows variation of existing or creation of new behaviors, implements a form of unified neuronal workspace (Dehaene and Naccache, 2001; on a more abstract level this might be termed a global workspace following; Baars and Franklin, 2007). In this global workspace, behavioral elements can be accessed (and possibly varied) in new contexts in order to find a (new) solution to an actual problem. The decisions of the system are not directly triggered by sensory input and are not predictable from outside. Instead action selection occurs at various levels of complexity. We argue that such a system shows properties of what has been termed access consciousness by Cleeremans (2005; which is distinguished from metacognition and phenomenological consciousness). Access consciousness refers to the ability of A commentary on