Evolution and analysis of self-synthesized minimalist neural controllers for collective robotics using Pareto multi-objective optimization

In this paper, we investigate the utilization of a multi-objective approach for evolving artificial neural networks (ANNs) that act as controllers for a radio frequency (RF) based collective box-pushing task of a group of virtual E-puck robots simulated in a 3D, physics-based environment. The modified Pareto-frontier Differential Evolution (PDE) algorithm is used to generate the Pareto optimal sets of ANN that optimize the conflicting objectives of maximizing the virtual E-puck robots' behaviors for pushing a box towards a wall based on RF-localization as well as minimizing the number of hidden neurons used in its feed-forward ANN controller. A new fitness function used during the collective robotics' optimization process is proposed. The experimentation results showed the virtual E-puck robots were capable of moving towards to the RF signal area. Thereafter, the robots were capable in self-assembling in the signal source area as well as completing the box-pushing towards the target wall with very small neural network architectures. Then, the genetic structures of the generated controllers have been further analyzed using the Hinton analysis. Hence, the results demonstrated that the utilization of the PDE approach in evolutionary robotics can be practically used to generate neural-based controllers for collective robotics' behaviors.