A study on the effects of recency factors on prediction in real-world domains

Temporal difference methods have been proposed to solve the problem of prediction-that is, using past experience with an incompletely understood system to predict its future behavior. These methods utilize a recency factor that gives a weightage to successive predictions. Conventionally, this term has been modelled by an exponential function primarily because of its functional simplicity and its ability to simulate the `forgetting law´ of synaptic dynamics. However, in real-world problems like rainfall prediction, where modelling real neurons is not the goal, it is not appropriate because it has a large negative slope and does not lead to optimal predictions. We examine these issues and also suggest an alternative recency which leads to better predictions and still retains some functional advantages of the original function