Towards a ´neural´ architecture for abductive reasoning

The authors formulate the general task of abduction as a nonlinear nonmonotonic constrained optimization problem. They then consider a linear monotonic version of the general abductive problem, and propose a neural network for solving it. The neurons in this network represent the elementary explanatory hypotheses and the connections between them are symmetric. It is found that representing the abductive problem as minimization of an energy function requires a network of order greater than two. The authors outline a second ´neural´ architecture that reflects the structure of the abductive problem. In this model, the constraints of the problem are represented explicitly, the network is composed of functional modules, and the connections between the ´neurons´ are asymmetric. Suggestions are made as to how this second-order network can accommodate certain interactions between the elementary hypotheses.<>