SIGGRAPH '94 Proceedings of the 21st annual conference on Computer graphics and interactive techniques
Efficient Reinforcement Learning Through Evolving Neural Network Topologies
GECCO '02 Proceedings of the Genetic and Evolutionary Computation Conference
| Hi-index | 0.00 |
We propose a novel algorithm for the evolution of body and control of three-dimensional, physically simulated virtual creatures controlled by artificial neural networks. The proposed algorithm is inspired by NeuroEvolution of Augmenting Topologies (NEAT) which efficiently evolves artificial neural networks. All three 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 the virtual creatures. Large-scale experiments have shown that the proposed algorithm evolves creatures using significantly less fitness evaluations than a standard genetic algorithm on all four tested fitness functions. Positive contribution of each component of the proposed algorithm has been confirmed with a series of supplementary ablation experiments.