An Incremental Approach to the Evolutionary Design of Autonomous Controllers for Micro-unmanned Aerial Vehicles

摘要

The work presented herein aims to provide a quantitative measure of the impact deriving from the adoption of an incremental approach to evolution within the context of Evolutionary Robotics. Notwithstanding the large amount of published researches relying on incremental evolution, little quantitative analysis have been performed so far to provide the answer to a basic question: is incremental evolution beneficial for evolutionary approaches to autonomous robotics? The application we use as testbed is a computer-based model of MAV (Micro-unmanned Aerial Vehicles) autonomous naviga- tion. Either single individuals embedding a neural network controller or teams made of several MAVs are subjected to different tasks across a simulated three-dimensional world. These tasks are: (1) navigation to a target area, (2) tracking of a moving target, and (3) execution of a behaviour requiring spatial and temporal coordination among members of the same team. The results obtained are compared with those generated via direct evolution. There are no evidences of systematic advantages deriving from the adoption of an incremental evolution approach.