Hybrid tuning of activation functions in feedforward neural networks

摘要: Tuning procedures for activation functions significantly increases the flexibility and nonlinear approximation capability of feedforward neural networks in supervised learning tasks. As a consequence, process presents better performance, with final state network being kept away from undesired saturation regions. Based on hybrid architecture combining gradient strategy fuzzy decision model, an auto-tuning algorithm is derived to adjust additional parameters associated functions. The other conventional parameters, connection weights between layers, are adjusted using powerful second-order approach based conjugate algorithm. To demonstrate performance proposed method we compare this technique standard solely descent method. three applied several artificial real world benchmarks.