Structure and constraint in discrete adaptive networks

Whether arrived at by chance or by design, any network of adaptive logic elements starts with a fixed structure. Attendant with a fixed structure are constraints on performance capabilities. Any task such a network is to learn will be performed in some input environment, and in most practical situations the environment itself will be constrained. These latter constraints imply the existence of certain relationships among the inputs. It is held that for the net to perform any given task well, the structure of the net must be matched to the relationships in the environment which are pertinent to the task. This paper discusses these ideas and presents an approach, which uses the theory of decomposition of switching functions, for matching the structure of a network with the constraints in the environment pertinent to the task.