Computational mechanisms for action selection
Imagine a zebra in the African savannah. At each moment in time this zebra has to weigh up alternative courses of action before deciding which will be most beneficial to it. For instance, it may want to graze because it is short of food, or it may want to head towards a water hole because it is short of water, or it may want to remain motionless in order to avoid detection by the predator it can see lurking nearby. This is an example of the problem of action selection: how to choose, at each moment in time, the most appropriate out of a repertoire of possible actions. This thesis investigates action selection in a novel way and makes three main contribu¬ tions. Firstly, a description is given of a simulated environment which is an extensive and detailed simulation of the problem of action selection for animals. Secondly, this simulated environment is used to investigate the adequacy of several theories of ac¬ tion selection such as the drive model, Lorenz's hydraulic model and Maes' spreading activation network. Thirdly, a new approach to action selection is developed which determines the most appropriate action in a principled way, and which does not suffer from the inherent shortcomings found in other methods.