“Towards Efficient Evolutionary Design of Autonomous Robots”, 2008 ():
Recent works explored the possibility of designing physical robots using evolutionary algorithms.
We propose a novel algorithm for the evolution of morphology and control of autonomous robots controlled by artificial neural networks. The proposed algorithm is inspired by NeuroEvolution of Augmenting Topologies (NEAT) which efficiently evolves artificial neural networks. All 3 main components of NEAT algorithm (protecting evolutionary innovation through speciation, effective crossover of neural networks with different topologies and incremental growth from minimal structure) are applied to the evolution of both morphology and control system of a robot.
Large-scale experiments with simulated robots have shown that the proposed algorithm uses substantially less fitness evaluations than a standard genetic algorithm on all 4 tested fitness functions. [but also struggled with reward hacking]
Positive contribution of each component of the proposed algorithm has been confirmed with a series of supplementary ablation experiments.
[Keywords: autonomous robot, evolutionary design, neural connection, ablation experiment, standard genetic algorithm]