Genetic Programming and Neural Networks as Interpreters for a Distributive Tactile Sensing System

This paper describes performance of a neural network and genetic programming (GP) in identifying the state of contact in a distributive tactile sensing system. The chosen architecture for the neural network is a multilayer perceptron while that for the genetic programming is a structured representation on genetic algorithms for non-linear function fitting (STROGANOFF). The tactile system comprises a small matrix of sensors for detecting deformation of a tactile surface. The determination of contact state is completed using both simulated and experimental inputs. Because the system relies on few sensing positions hence a robust interpreting algorithm plays a vital role. The study involves the identification of the position of a pointed load for a range between 200-600 g which can be applied across the surface. The performance in determining the position is described in the form of absolute deviation from the actual applied position. The simulation result indicates that the multilayer perceptron is the best inference technique while the GP-based mapping model produces a better result in an experiment with a high load. The difference between the simulation and the experiment is the result of an inability of the simulation model at capturing true plate deflection characteristics.