Page - 98 - in Limina - Grazer theologische Perspektiven, Volume 3:2

98 | www.limina-graz.eu Sara Lumbreras and Lluis Oviedo | Belief networks as complex systems Reinforcement Learning as a metaphor for decision making Another AI technique that can enrich the brain-computer metaphor is Re- inforcement Learning (RL). RL simulates an agent that can take actions in her environment. These actions can lead to a reward. The agent learns, by trial and error, the consequences of her actions. Then, she can take the de- cision that is optimal for her situation (we will refer to this as her state).1 We define “optimal” according to a given objective, described by means of an objective mathematical formula. The power of RL and Dynamic Control, its underlying technique, is that by defining only the value of the differ- ent final outcomes, it can derive the value of the intermediate outcomes, so that the method arrives at the optimal strategy at each stage even if it is not the final one, as the final value cascades into the nearer ones. For in- stance, if we apply RL to playing chess, the final outcome can be winning or not. The next-to-last states can be described in terms of their probability of resulting in a winning situation. Then, the ones before can be evaluated in terms of the next ones, and so on. Reinforcement Learning therefore provides a basic mechanism to understand how, from meta-goals, we can derive intermediate goals and optimal strategies. Belief systems have several functions as has been recognized in the litera- ture (Frank 1977). One of these functions is to evaluate a given state in or- der to select the next most desirable course of action, so that they have a key role in these processes. 3 The limitations of the pattern-recognition simile The brain is far more complex than an ANN. A single neuron is an entity of a remarkable complexity, which cannot be reduced to a mere simple math- ematical operation. We should remember that even unicellular organisms exhibit complex behaviour: amoebas hunt, protozoa build complex shells from minuscule specks of dust. A living cell is a wonderfully complex being, and we should keep in mind that ANNs are only vaguely based on the work- ings of the brain; they are not modelled as its equivalent. Reinforcement Learning provides a mechanism to understand how, from meta-goals, we can derive intermediate goals and optimal strategies. 1 For more context on these tech- niques, we refer the reader to Sut- ton/Barto 2018.