Active training of backpropagation neural networks using the learning by experimentation methodology

This paper proposes the Learning by Experimentation Methodology to facilitate the active training of neural networks. In an active learning paradigm, a learning mechanism can actively interact with its environment to acquire new knowledge and revise itself. The learning by experimentation is an active learning strategy. Experiments are conducted to form hypotheses, and the evaluation of those hypotheses feeds back to the learning mechanism to revise knowledge. We use a backpropagation neural network as the learning mechanism. We also adopt a weight space analysis method and a heuristic to select salient attributes to perform new experiments in order to revise the network. Finally, we illustrate performance by solving the sonar signal classification problem.