Evolving Controllers for Simulated Agricultural Scenarios Using Genetic Programming

Abstract

A lot of work in the agriculture world consists of guiding a vehicle on a field. The goal is to cover the entire field with e.g. a plow or a sowing machine. Today there are several systems available which assist the driver. However, no system is able to completely automate the tasks. This thesis investigates GPs potential for automating these tasks in a simulated environment. The investigation was performed by extending the well known Lawnmower problem, proposed by Koza, into a little more realistic version. A simulator was implemented and a set of tests conducted with two different types of controllers. Some of tests showed promise, but in most cases much of the area was left uncovered. The overall strategy used by the controllers seemed too simplistic for general use. Even though some of the results were promising, it is difficult to generalize and say something about its real world usefulness. A more in-depth investigation is needed before any definite conclusion can be made. One thing is certain, efficient and accurate automatic controllers will be greatly appreciated by a farmer.