CONSENSUS: A Statistical Learning Procedure in a Connectionist Network.

Abstract

We present a new scheme for the activity of neuron-like elements in a connectionist network. The CONSENSUS scheme is based on statistical inference. The guiding principle of CONSENSUS is that decisions should be deferred until sufficient evidence accumulates to make an informed choice. Consequently, large changes in network structure can be made with confidence. Nodes have an awareness of their role and utility in the network which allows them to increase their effectiveness. The reinforcement scheme utilizes the notion of confidence so that only nodes proven to contribute successfully issue reinforcements. Nodes are grouped into communities to exploit their collective knowledge which exceeds any individual member. The network was tested against several problems and was able to find suitable encodings to solve them.