Reinforcement learning of a neural network using a fuzzy logic controller: application to seismic tomographic data

A neural network-fuzzy logic controller (NN-FLC) methodology is developed to train a neural network in the absence of a training set. In our method, the outputs from the neural network are transformed back to the form of the inputs of the network with an application-dependent transformation module. These values are then compared to the inputs given to the neural network to calculate the errors, thus a basis for the evaluation of the network is established. If the errors are not within a tolerable limit, these errors are then evaluated by the fuzzy logic controller. The fuzzy logic controller provides the updates to the outputs of the neural network. The differences between updated and computed outputs are then backpropagated to train the network. Thus the fuzzy logic controller acts as a teacher to the neural network. The training is continued until the sum of errors is within a tolerable limit. If the reduction in cumulative error over successive NN-FLC iterations is converged to a steady value which is still larger than the predefined acceptable level, or if the cumulative error begins to oscillate, then the analysis is continued with a fine-tuning module. The methodology is applied to the inversion of seismic tomographic data for obtaining the slowness pattern of an interwell region. A number of examples with synthetic and field datasets gave encouraging results