Search

Many experiments have been performed that use evolutionary algorithms for learning the topology and connection weights of a neural network that controls a robot or virtual agent. These experiments are not only performed to better understand basic biological principles, but also with the hope that with further progress of the methods, they will become competitive for automatically creating robot behaviors of interest. However, current methods are limited with respect to the (Kolmogorov) complexity of evolved behavior. Using the evolution of robot trajectories as an example, we show that by adding four features, namely (1) freezing of previously evolved structure, (2) temporal scaffolding, (3) a homogeneous transfer function for output nodes, and (4) mutations that create new pathways to outputs, to standard methods for the evolution of neural networks, we can achieve an approximately linear growth of the complexity of behavior over thousands of generations. Overall, evolved complexity is up to two orders of magnitude over that achieved by standard methods in the experiments reported here, with the major limiting factor for further growth being the available run time. Thus, the set of methods proposed here promises to be a useful addition to various current neuroevolution methods.

It has been suggested recently that it is a reasonable abstraction of evolutionary processes to use evolutionary algorithms that select individuals based on the novelty of their behavior instead of their fitness. Here we study the performance of fitness- and novelty-based search on several neuroevolution tasks. We also propose several new algorithms that select both for fit and for novel individuals, but without weighting these two criteria directly against each other. We find that behavioral speciation, behavioral near neutral speciation, and behavioral novelty speciation perform best on most tasks. Pure novelty search, as well as a number of hybrid methods without speciation mechanism, do not perform well on most tasks. Using behavioral criteria for speciation often yields better results than using genetic criteria.

We demonstrate that neural agents can evolve behavioral sequences of arbitrary length. In our framework, agents in a two-dimensional arena have to find the secure one among two possible patches, and which of them is secure changes over time. Evolution of arbitrarily long behavioral sequences is achieved by extending the neuroevolution method NEAT with two techniques: Only newly evolved network structure is subject to mutations, and inputs to the neural network are provided in an incremental fashion during evolution. It is suggested that these techniques are transferable to other neuroevolution methods and domains, and constitute a step towards achieving open-ended evolution. Furthermore, it is argued that the proposed techniques are strongly simplified models of processes that to some degree occur naturally in systems with more flexible genetic architectures.

The control of multilegged robots is challenging because of the large number of sensors and actuators involved. However, the regularity inherent to gait control can be taken into account to design controllers for multilegged robots. In this paper, we show that NEATfields, a method designed for the evolution of large neural networks, can exploit this regularity to evolve significantly better gaits than those evolved by the standard NEAT method. We also show how evolved networks can control a robot with a ball-like morphology to move on a rough terrain. The success in evolving large neural networks suggests that the NEATfields method is a promising tool for studying complex behaviors in robotics and artificial life. © 2011 IEEE.

The way genes are interpreted biases an artificial evolutionary system towards some phenotypes. When evolving artificial neural networks, methods using direct encoding have genes representing neurons and synapses, while methods employing artificial ontogeny interpret genomes as recipes for the construction of phenotypes. Here, a neuroevolution system (neuroevolution with ontogeny or NEON) is presented that can emulate a well-known neuroevolution method using direct encoding (neuroevolution of augmenting topologies or NEAT), and therefore, can solve the same kinds of tasks. Performance on challenging control and memory benchmark tasks is reported. However, the encoding used by NEON is indirect, and it is shown how characteristics of artificial ontogeny can be introduced incrementally in different phases of evolutionary search.