Catálogo de publicaciones - libros

Standard Genetic Programming operators are highly disruptive, with the concomitant risk that it may be difficult to converge to an optimal structure. The Tree Adjoining Grammar (TAG) formalism provides a more flexible Genetic Programming tree representation which supports a wide range of operators while retaining the advantages of tree-based representation. In particular, minimal-change point insertion and deletion operators may be defined. Previous work has shown that point insertion and deletion, used as local search operators, can dramatically reduce search effort in a range of standard problems. Here, we evaluate the effect of local search with these operators on a real-World ecological time series modelling problem. For the same search effort, TAG-based GP with the local search operators generates solutions with significantly lower training set error. The results are equivocal on test set error, local search generating larger individuals which generalise only a little better than the less accurate solutions given by the original algorithm.

A technique used to visualize DNA sequences is adapted to visualize large numbers of individuals in a genetic programming population. This is used to examine how the content diversity of a population changes during evolution and how this correlates with changes in fitness.

We present an agent-based, 3D model of the lactose () operon, a gene regulatory system the bacterium The operon is a prime example of a ‘real genetic programming’ system, which has been studied extensively and lends itself to rigorous mathematical analysis and computational simulations. We suggest natural gene regulatory systems, as observed within , to serve as testbeds for future genetic programming systems.

We examine the behavior of an evolutionary search on neutral networks in a simple linear genetic programming system of a Boolean function space problem. To this end we draw parallels between notions in RNA-folding problems and in Genetic Programming, observe parameters of neutral networks and discuss the population dynamics via the occupation probability of network nodes in runs on their way to the optimal solution.

Bloat is a common and well studied problem in genetic programming. Size and depth limits are often used to combat bloat, but to date there has been little detailed exploration of the effects and biases of such limits. In this paper we present empirical analysis of the effects of size and depth limits on binary tree genetic programs. We find that size limits control population average size in much the same way as depth limits do. Our data suggests, however that size limits provide finer and more reliable control than depth limits, which has less of an impact upon tree shapes.

This chapter gives a systematic view, based on the experience from The Dow Chemical Company, of the key issues for applying symbolic regression with Genetic Programming (GP) in industrial problems. The competitive advantages of GP are defined and several industrial problems appropriate for GP are recommended and referenced with specific applications in the chemical industry. A systematic method for selecting the key GP parameters, based on statistical design of experiments, is proposed. The most significant technical and non-technical issues for delivering a successful GP industrial application are discussed briefly.

This chapter describes how genetic programming might be integrated as a tool into the human context of discovery. To accomplish this, a comparison is made between GP and a well-regarded strategy in open-ended problem solving. The comparison indicates which tasks and skills are likely to be complemented by GP. Furthermore, the comparison also indicates directions in research that may need to be taken for GP to be further leveraged as a tool that assists discovery.

Genetic programming has proved its potential for automated synthesis of a variety of engineering systems such as electrical, control, and mechanical systems. Given any of these application domains, a set of generic GP functions can be developed for its synthesis. In this chapter, however, we illustrate that while a generic GP system can often be used to prove a concept, realistic or industrial automated synthesis often requires domain-specific GP configuration, especially of the GP function sets. As a case study, it is shown how the open-ended topology search capability of GP readily exploits “loopholes” in a generic bond-graph-based GP function set and evolves high-performance but unrealistic mechanical vibration absorbers, even though the bond graphs would be readily implementable in, for example, the electrical domain. The preliminary attempt to constrain evolved topologies to only those that would be readily implementable in the mechanical domain was not sufficiently restrictive.

This paper investigates the application of genetic programming to problems in industrial analog computer-aided design (CAD). One CAD subdomain, analog structural synthesis, is an often-cited success within the genetic programming (GP) literature, yet industrial use remains elusive. We examine why this is, by drawing upon our own experiences in bringing analog CAD tools into industrial use. In sum, GP-synthesized designs need to be more robust in very specific ways. When robustness is considered, a GP methodology of today on a reasonable circuit problem would take 150 years on a 1,000-node 1-GHz cluster. Moore’s Law cannot help either, because the problem itself is ‘Anti-Mooreware’ — it becomes more difficult as Moore’s Law progresses. However, we believe the problem is still approachable with GP; it will just take a significant amount of ‘algorithm engineering.’ We go on to describe the recent application of GP to two other analog CAD subdomains: symbolic modeling and behavioral modeling. In contrast to structural synthesis, they are easier from a GP perspective, but are already at a level such that they can be exploited in industry. Not only is GP the only approach that gives interpretable SPICE-accurate nonlinear models, it turns out to outperform nine other popular blackbox approaches in a set of six circuit modeling problems.