A Neural Network Abductive Model

Abduction is defined as a reasoning to a best explanation for a set of given data. Generally, finding an explanation for the given data is computationally expensive. Neural network computing is known to be strong in solving computationally difficult tasks. In this paper, a neural network abductive model is presented. One feature of this model is that, a causal network is directly used as the neural network without any further transformation. The causal network can capture both conjunctive and disjunctive relations which enable the model to represent monotonic abductive problems and independent abductive problems as well. The second feature of this model is that the manifestation distance is used to guide the network computing. This may be different from all the existing abductive models regarding the explanation judgement. An important issue for abduction is to deal with abductive problems with dependent hypotheses. For most existing models, the hypothesis interaction has been ignored. The present model takes the hypothesis interactions into account. The key idea of solving the abductive problems with dependent hypotheses is to reduce the impact caused by ignoring the interactions rather than to calculate the interactions. The experimental results show that the rate of correct solutions derived by this model is high.