The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9750

Title:Evolution and Co-Evolution of Pursuit-Evasion Tactics
Date:Sep 1997
Abstract:Genetic programming is a technique which hasthe potential to evolve and co-evolve pursuit-evasion controllers for simulated realistic robots in complex environments. Its effectiveness depends upon choosing the correct controller representation, function set, terminal set and fitness function. By using a modified Khepera simulator and a modified genetic programming package, experiments are described that attempt to evolve an evader against a fixed behaviour pursuer in three different environments. Each environment's layout varies in the position and number of obstacles. These experiments are analysed in detail to see if genetic programming is effective in evolving good controllers. A scond set of experiments then allows both pursuer and evader to co-evolve in the same three environments. By studying co-evolution we are able to see how the dynamic fitness landscapes and different opponent selection strategies affects the evolution of the pursuer's and evader's behaviours.

[Search These Pages] [DAI Home Page] [Comment]