Evolving neural networks to follow trajectories of arbitrary complexity.

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.

Machine learning of symbolic compositional rules with genetic programming: dissonance treatment in Palestrina.

We describe a method for automatically extracting symbolic compositional rules from music corpora. Resulting rules are expressed by a combination of logic and numeric relations, and they can therefore be studied by humans. These rules can also be used for algorithmic composition, where they can be combined with each other and with manually programmed rules. We chose genetic programming (GP) as our machine learning technique, because it is capable of learning formulas consisting of both logic and numeric relations. GP was never used for this purpose to our knowledge. We therefore investigate a well understood case in this study: dissonance treatment in Palestrina's music. We label dissonances with a custom algorithm, automatically cluster melodic fragments with labelled dissonances into different dissonance categories (passing tone, suspension etc.) with the DBSCAN algorithm, and then learn rules describing the dissonance treatment of each category with GP. Learning is based on the requirement that rules must be broad enough to cover positive examples, but narrow enough to exclude negative examples. Dissonances from a given category are used as positive examples, while dissonances from other categories, melodic fragments without dissonances, purely random melodic fragments, and slight random transformations of positive examples, are used as negative examples.

Evolving neural fields for problems with large input and output spaces.

We have developed an extension of the NEAT neuroevolution method, called NEATfields, to solve problems with large input and output spaces. The NEATfields method is a multilevel neuroevolution method using externally specified design patterns. Its networks have three levels of architecture. The highest level is a NEAT-like network of neural fields. The intermediate level is a field of identical subnetworks, called field elements, with a two-dimensional topology. The lowest level is a NEAT-like subnetwork of neurons. The topology and connection weights of these networks are evolved with methods derived from the NEAT method. Evolution is provided with further design patterns to enable information flow between field elements, to dehomogenize neural fields, and to enable detection of local features. We show that the NEATfields method can solve a number of high dimensional pattern recognition and control problems, provide conceptual and empirical comparison with the state of the art HyperNEAT method, and evaluate the benefits of different design patterns.

Neuroevolution and complexifying genetic architectures for memory and control tasks.

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.