Flocking with GA/PSO SI Course Project Yvan Bidiville & Thomas Thurnherr
The Goal • Evolve a controller for robots to move as flock. • Explore the effectiveness of GA and PSO, with both homogeneous and heterogeneous learning.
The Idea Behind the Given Code Neural Network
The First Try Neural Network Average Direction
Problems • Assuming too much information on the robots side: They cannot figure out the direction of their neighbours. • Bad approach, which does not really make use of the neural network.
Neural Network The Second Try
The Fitness Function • Former fitness function: fit[i] = 1.0 - #neighbours/#robots • New fitness function fit[i] = 1.0 - (#neighbours/#robots) · min(1.0, d[i]), where d[i] = (x last [i] - x first [i] ) 2 + (y last [i] - y first [i] ) 2 , with x and y coordinates of the centre of mass. • Fitness value within the interval [0, 1]
Webots-Movie of a Simulation QuickTime™ and a decompressor are needed to see this picture.
Any Questions?
Recommend
More recommend
