Abstract
In this paper we compare systematically the most promising neuroevolutionary methods and two new original methods on the double-pole balancing problem with respect to: the ability to discover solutions that are robust to variations of the environment, the speed with which such solutions are found, and the ability to scale-up to more complex versions of the problem. The results indicate that the two original methods introduced in this paper and the Exponential Natural Evolutionary Strategy method largely outperform the other methods with respect to all considered criteria. The results collected in different experimental conditions also reveal the importance of regulating the selective pressure and the importance of exposing evolving agents to variable environmental conditions. The data collected and the results of the comparisons are used to identify the most effective methods and the most promising research directions.
Highlights
Neuroevolution, namely the evolution of neural networks selected for the ability to perform a given function [1], constitutes a general and effective method that can be applied to a wide range of problems
The data reported in the table are those obtained with the best combination of parameters: xNES, Parallel Stochastic Hill Climber (PSHC) (MutRate 50%, Stochasticity 10%, Interbreeding 10%, NumHiddens 1), Coevolved Synapses (CoSyNE) (MutRate 90%, SubPopulations 5, NumHiddens 1), Stochastic Steady State (SSS), Cartesian Genetic Programming of Artificial Neural Network (CGPANN) (MutRate 3%; NumIncomingConnections 8, NumHiddens 2), and NeuroEvolution of Augmenting Topology (NEAT)
The data reported in the table are those obtained with the best combination of parameters: NEAT, CoSyNE, CGPANN; varying positions: MutRate 1%, Incoming Connections 4), xNES, SSS, PSHC
Summary
Neuroevolution, namely the evolution of neural networks selected for the ability to perform a given function [1], constitutes a general and effective method that can be applied to a wide range of problems. It presents several advantages with respect to alternative training methods for neural networks. Since it does not depend on gradient information, it can be applied to problems in which this information is unavailable, too noisy, or too costly to be obtained. By operating on the basis of a scalar fitness value that rates the extent to which the network solves the given problem, it can be applied to any problem. It is effective in partially observable domains since, unlike most Reinforcement Learning methods, does not rely on the Markov assumption. It can be applied to any type of neural network and can be used to adapt all the characteristics of the network including the architecture of the network, the transfer function of the neurons, and the characteristics of the system (if any) in which the network is embedded
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
