An Empirical Comparison of Genetically Evolved Programs and Evolved Neural Networks for Multi-agent Systems Operating under Dynamic Environments
J.J. Davila · 2015
This paper expands on the research presented in [12] by comparing the performance of genetically evolved programs operating under dynamic game environments with that of neural networks with evolved weights. On the genetic programming side, the maximum allowed tree depth was varied in order to study its effect on the evolutionary process. For evolution of neural networks, encoding included direct encoding of weights and three different L-Systems. Empirical results show that genetic evolution of neural networks weights provided better performance under dynamic environments when evolved to choose which of several high-level actions to perform, such as "defend" or "attack". On the other hand, genetic programming evolved better solutions for low-level actions, such as "move left," "move right," or "accelerate." Solutions are analyzed in order to explain these differences.