Created by W.Langdon from gp-bibliography.bib Revision:1.7175
The acquisition of models is based on data logged when a human pilot was controlling the vehicle, and is carried out by an evolutionary algorithm based on competitive coevolution. Models in the form of symbolic expressions are coevolved along with the portions of the training data that are used to compute their fitness. This results in an effective and computationally efficient way of constructing models.
The modelling method is applied to a small toy car, a full sized aeroplane and two different types of small quadrotor helicopters. For comparison, models of the same vehicles are also derived using standard modelling techniques that exploit platform knowledge. The models produced by our technique are shown to be as accurate or better than those produced manually. Importantly after a limited amount of rearrangement of terms, the models also prove to be interpretable.
A method is presented for reproducing in the models the noise and uncertainties that characterise real world platforms. The evolved deterministic models produced are augmented with a simple yet computationally efficient Gaussian noise model, and a principled method based on unscented Kalman filtering is used to estimate the noise parameters. The augmented models are demonstrated to reproduce most of the variability shown by real vehicles.
The automatic design of controllers considers both monolithic and modular structures based on recurrent neural networks. Conventional steady state evolution is used to evolve monolithic controllers, and cooperative coevolution is applied to modular controllers. Manually designed controllers are also developed for purposes of comparison. Controllers are mainly evolved for path-following tasks, but other tasks like imitating game players' abilities are also considered.
In general monolithic controllers are shown to be very effective in controlling the toy car, but have limitations when applied to the helicopters. Modular networks show a better ability to scale to more demanding platforms, and in simulation reach levels of performance comparable to or better than controllers designed manually.
Tests show that for both the toy car and quadrotor helicopters, the evolved controllers successfully transfer to the real vehicles, although a certain amount of mismatch exists between the performances predicted in simulation and those on the real platforms.",
Genetic Programming entries for Renzo De Nardi