Learning by statistical cooperation of self-interested neuron-like computing elements.

Since the usual approaches to cooperative computation in networks of neuron-like computating elements do not assume that network components have any "preferences", they do not make substantive contact with game theoretic concepts, despite their use of some of the same terminology. In the approach presented here, however, each network component, or adaptive element, is a self-interested agent that prefers some inputs over others and "works" toward obtaining the most highly preferred inputs. Here we describe an adaptive element that is robust enough to learn to cooperate with other elements like itself in order to further its self-interests. It is argued that some of the longstanding problems concerning adaptation and learning by networks might be solvable by this form of cooperativity, and computer simulation experiments are described that show how networks of self-interested components that are sufficiently robust can solve rather difficult learning problems. We then place the approach in its proper historical and theoretical perspective through comparison with a number of related algorithms. A secondary aim of this article is to suggest that beyond what is explicitly illustrated here, there is a wealth of ideas from game theory and allied disciplines such as mathematical economics that can be of use in thinking about cooperative computation in both nervous systems and man-made systems.