Evolutionary exploration of dynamic swarm behaviour
Hobin Kwong, Christian J. Jacob · 2003
In general, it is difficult to make a highly dynamic swarm system follows explicit behaviour patterns. Multiple, simultaneous interactions among a large number of agents make the non-linear relationship between a parameter change and the corresponding effect on global behaviour non-intuitive and, consequently, hard to control. This paper presents breeding experiments of dynamic swarm behaviour patterns using an interactive evolutionary algorithm. Specifically, eight scalar parameters that influence swarm behaviour dynamics in a 3D swarm simulation are bred to produce agents that collectively fly in line, ring, and figure-eight formations. Our initial examples demonstrate that a 'swarm breeding' system can partly eliminate the manual tuning of control parameters and provides a viable approach to design swarm systems through interactive genetic programming.