Seite - 101 - in Limina - Grazer theologische Perspektiven, Band 3:2

101 | www.limina-graz.eu Sara Lumbreras and Lluis Oviedo | Belief networks as complex systems they can also read any other word. They have unconsciously learnt to iden- tify the letters and can generalize in their ability to read but have learnt much more quickly than if the letters and combinations are shown to them directly. This method has been proven to work even with toddlers who can- not yet speak, given that understanding visual information is much easier than articulating words. After all, human beings speak well before under- standing grammar, and walk without any conscious idea of the inner work- ings of their anatomy. The second type of change in belief systems is belief change, where an alter- native definition of a concept is found to have more explanatory power or to be less subject to ambiguity (Bendixen 2002). If the new concept definition is evaluated as considerably superior to the one previously held, then the belief changes. We discuss some further details of these processes in con- nection to complexity theory in the next sections. An additional missing issue that has not yet been adequately introduced in machine learning is the consideration of logical constraints. However, there have been some very interesting developments in understanding how beliefs can be related to each other through logical constraints and how in- dividuals influence each other (Friedkin et al. 2016). This is especially im- portant: cognitive dissonance emerges when these logical constraints are not respected. 4 The limitations of the Reinforcement Learning perspective The main issue in the RL perspective is the definition of the objective func- tion, that is, the evaluation of the different outcomes for the agent. When a programmer is developing the system, she creates a mathematical function that expresses her preferences (i. e. what states are preferred compared to others). However, it is not clear how this function would be built in the case of a living organism. It seems intuitive to choose survival, or reproduction, as the final goal for a living being. Then, some intermediate goals could be chosen by natural selection, given that they will favour survival and reproduction. Evolution may have intuitively selected homeostasis as a process that supports sur- vival. This would include, for example, physical integrity or feeding when hungry. Then, individuals can derive intermediate goals that get them to a point where keeping homeostasis is easier or more difficult. RL provides an in-