Genetic algorithms for swarm parameter tuning

• Main Author: Chee, Glenn Jun Yuan
• Other Authors: Zinovi Rabinovich
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Computer science and engineering::Computer applications::Social and behavioral sciences; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Simulation and modeling
• Online Access: http://hdl.handle.net/10356/77042
• Institution: Nanyang Technological University

Swarming behaviour is based on the aggregation of simple drones exhibiting basic instinctive reactions to stimuli. However, to achieve overall balanced/interesting behaviours, the relative importance of these instincts as well their internal parameters must be tuned. This project attempts to achieve a series of non-trivial swarm-level behaviours by applying Genetic Programming as means of such tuning. To measure the performance of a swarm, a swarm simulator was designed in Python. The simulator receives parameters for a swarm and a scenario, and outputs a score based on the swarm’s performance for that scenario. The score is used by the Multi-Objective Genetic Algorithm (MOGA) as fitness to determine the swarm’s chance of breeding; higher the fitness, higher the chance of breeding. The Genetic Algorithm (GA) terminates after 1000 generations and evaluations are done thereafter on the resulting population for analysis. It was observed that the GA did indeed improve individual specialisations substantially, but improved multi-specialisations only marginally better. Uniform crossover was evident to be the better crossover function, possibly owing to the genes being independent. Interesting behaviours, namely swarm cohesion and bait ball, were also observed on the resultant. In conclusion, MOGA shows promise of optimising multiple objectives together but is negligible thus far. However, MOGA succeeded in manufacturing certain behaviours by tuning swarm parameters. Further study needs to be done to determine the full potential of MOGA.