SALMON: A TEMPERAMENTAL program that learns
Abstract
A new approach to machine learning is introduced that utilizes semantic information in a connectionist network. The approach is implemented in a program that learns to act appropriately in the dynamic environment of a children's game of tag. The model is interesting in several respects including the ability to begin with no connections and then make and break them according to its experience, the ability to adjust the weights on its connections and the ability to interact with its environment.