Robert T. Pennock

The evolutionary origin of complex features

A long-standing challenge to evolutionary theory has been whether it can explain the origin of complex organismal features. We examined this issue using digital organisms—computer programs that self-replicate, mutate, compete and evolve. Populations of digital organisms often evolved the ability to perform complex logic functions requiring the coordinated execution of many genomic instructions. Complex functions evolved by building on simpler functions that had evolved earlier, provided that these were also selectively favoured. However, no particular intermediate stage was essential for evolving complex functions. The first genotypes able to perform complex functions differed from their non-performing parents by only one or two mutations, but differed from the ancestor by many mutations that were also crucial to the new functions. In some cases, mutations that were deleterious when they appeared served as stepping-stones in the evolution of complex features. These findings show how complex functions can originate by random mutation and natural selection.

Evolving coordinated quadruped gaits with the HyperNEAT generative encoding

Legged robots show promise for complex mobility tasks, such as navigating rough terrain, but the design of their control software is both challenging and laborious. Traditional evolutionary algorithms can produce these controllers, but require manual decomposition or other problem simplification because conventionally-used direct encodings have trouble taking advantage of a problems regularities and symmetries. Such active intervention is time consuming, limits the range of potential solutions, and requires the user to possess a deep understanding of the problems structure. This paper demonstrates that HyperNEAT, a new and promising generative encoding for evolving neural networks, can evolve quadruped gaits without an engineer manually decomposing the problem. Analyses suggest that HyperNEAT is successful because it employs a generative encoding that can more easily reuse phenotypic modules. It is also one of the first neuroevolutionary algorithms that exploits a problems geometric symmetries, which may aid its performance. We compare HyperNEAT to FT-NEAT, a direct encoding control, and find that HyperNEAT is able to evolve impressive quadruped gaits and vastly outperforms FT-NEAT. Comparative analyses reveal that HyperNEAT individuals are more holistically affected by genetic operators, resulting in better leg coordination. Overall, the results suggest that HyperNEAT is a powerful algorithm for evolving control systems for complex, yet regular, devices, such as robots.

On the Performance of Indirect Encoding Across the Continuum of Regularity

This paper investigates how an evolutionary algorithm with an indirect encoding exploits the property of phenotypic regularity, an important design principle found in natural organisms and engineered designs. We present the first comprehensive study showing that such phenotypic regularity enables an indirect encoding to outperform direct encoding controls as problem regularity increases. Such an ability to produce regular solutions that can exploit the regularity of problems is an important prerequisite if evolutionary algorithms are to scale to high-dimensional real-world problems, which typically contain many regularities, both known and unrecognized. The indirect encoding in this case study is HyperNEAT, which evolves artificial neural networks (ANNs) in a manner inspired by concepts from biological development. We demonstrate that, in contrast to two direct encoding controls, HyperNEAT produces both regular behaviors and regular ANNs, which enables HyperNEAT to significantly outperform the direct encodings as regularity increases in three problem domains. We also show that the types of regularities HyperNEAT produces can be biased, allowing domain knowledge and preferences to be injected into the search. Finally, we examine the downside of a bias toward regularity. Even when a solution is mainly regular, some irregularity may be needed to perfect its functionality. This insight is illustrated by a new algorithm called HybrID that hybridizes indirect and direct encodings, which matched HyperNEATs performance on regular problems yet outperformed it on problems with some irregularity. HybrIDs ability to improve upon the performance of HyperNEAT raises the question of whether indirect encodings may ultimately excel not as stand-alone algorithms, but by being hybridized with a further process of refinement, wherein the indirect encoding produces patterns that exploit problem regularity and the refining process modifies that pattern to capture irregularities. This paper thus paints a more complete picture of indirect encodings than prior studies because it analyzes the impact of the continuum between irregularity and regularity on the performance of such encodings, and ultimately suggests a path forward that combines indirect encodings with a separate process of refinement.

The sensitivity of HyperNEAT to different geometric representations of a problem

HyperNEAT, a generative encoding for evolving artificial neural networks (ANNs), has the unique and powerful ability to exploit the geometry of a problem (e.g., symmetries) by encoding ANNs as a function of a problems geometry. This paper provides the first extensive analysis of the sensitivity of HyperNEAT to different geometric representations of a problem. Understanding how geometric representations affect the quality of evolved solutions should improve future designs of such representations. HyperNEAT has been shown to produce coordinated gaits for a simulated quadruped robot with a specific two-dimensional geometric representation. Here, the same problem domain is tested, but with different geometric representations of the problem. Overall, experiments show that the quality and kind of solutions produced by HyperNEAT can be substantially affected by the geometric representation. HyperNEAT outperforms a direct encoding control even with randomized geometric representations, but performs even better when a human engineer designs a representation that reflects the actual geometry of the robot. Unfortunately, even choices in geometric layout that seem to be inconsequential a priori can significantly affect fitness. Additionally, a geometric representation can bias the type of solutions generated (e.g., make left-right symmetry more common than front-back symmetry). The results suggest that HyperNEAT practitioners can obtain good results even if they do not know how to geometrically represent a problem, and that further improvements are possible with a well-chosen geometric representation.

Can’t philosophers tell the difference between science and religion?: Demarcation revisited

In the 2005 Kitzmiller v Dover Area School Board case, a federal district court ruled that Intelligent Design creationism was not science, but a disguised religious view and that teaching it in public schools is unconstitutional. But creationists contend that it is illegitimate to distinguish science and religion, citing philosophers Quinn and especially Laudan, who had criticized a similar ruling in the 1981 McLean v. Arkansas creation-science case on the grounds that no necessary and sufficient demarcation criterion was possible and that demarcation was a dead pseudo-problem. This article discusses problems with those conclusions and their application to the quite different reasoning between these two cases. Laudan focused too narrowly on the problem of demarcation as Popper defined it. Distinguishing science from religion was and remains an important conceptual issue with significant practical import, and philosophers who say there is no difference have lost touch with reality in a profound and perverse way. The Kitzmiller case did not rely on a strict demarcation criterion, but appealed only to a “ballpark” demarcation that identifies methodological naturalism (MN) as a “ground rule” of science. MN is shown to be a distinguishing feature of science both in explicit statements from scientific organizations and in actual practice. There is good reason to think that MN is shared as a tacit assumption among philosophers who emphasize other demarcation criteria and even by Laudan himself.

Models, simulations, instantiations, and evidence: the case of digital evolution

What is the difference between a simulation of X and simply another instance of X? Is there a point at which the “virtual reality” of a model becomes the real thing? This paper examines these questions using cases taken from recent developments in evolutionary engineering and artificial life research. By implementing the Darwinian mechanism and setting it to work on a design problem, scientists and engineers find that evolution not only can improve prior designs, but also produce novel technological solutions. Artificial life systems Tierra and Avida which operate at a higher level of abstraction than evolutionary engineering applications. I analyze simulation as a rational concept “S simulates R” and argue that it always includes some relevant property P, of R, that is captured but that there is always also some other that it omits, and that pragmatic factors fix what counts as relevant. The border between a simulation and an instance or R can change depending upon the context. I show that in one sense, evo-technology and artificial life simulate organic evolution, but in another relevant sense they are instances of evolution itself. Biologists can use such systems to experimentally test evolutionary hypotheses such as those involving the evolution of complex features and altruism. This analysis suggests lines for future research on broader questions about models classification and confirmation.

Should Creationism Be Taught in the Public Schools

This article discusses philosophicalarguments relevant to the question of teachingcreationism, especially with regard to developments inthe debate since the early 1990s.Section 1 reviews the newfactions within the creationist movement, and theoverlapping views from ‘young earth’ to ‘intelligentdesign’ creationism, as well as non-Christianvarieties. It also considers what are the relevantdifferences for the policy question for private,public schools, and for home schoolers, as well aspossible differences in what it means to ‘teach’creationism. Sections 2 & 3 discuss the main legal argumentsthat have ruled in the public school case, as well asarguments from academic freedom, fairness, censorship,parental rights and majority rule. Section 4 evaluates theepistemological issues regarding competing claims oftruth, and the contention that excluding ‘whatChristians know’ (Alvin Plantinga) amounts to‘viewpoint discrimination’ (Phillip Johnson). Section 5argues that religious protection arguments actuallyfavor excluding creationism more than including it. Section 6 considers the goals of education, especiallyDeweys views on science education, and what theseimply regarding the teaching of a ‘theistic science’. In Section 7, I review a new argument of Alvin Plantingabased upon a purported Rawlsian basic right of aparent not to have her children taught anything thatviolates her comprehensive beliefs, and show whyRawlsian agents would reject it.

Naturalism, evidence and creationism: The case of Phillip Johnson

Phillip Johnson claims that Creationism is a better explanation of the existence and characteristics of biological species than is evolutionary theory. He argues that the only reason biologists do not recognize that Creationists negative arguments against Darwinism have proven this is that they are wedded to a biased ideological philosophy —Naturalism — which dogmatically denies the possibility of an intervening creative god. However, Johnson fails to distinguish Ontological Naturalism from Methodological Naturalism. Science makes use of the latter and I show how it is not dogmatic but follows from sound requirements for empirical evidential testing. Furthermore, Johnson has no serious alternative type of positive evidence to offer for Creationism, and purely negative argumentation, despite his attempt to legitimate it, will not suffice.

HybrID: a hybridization of indirect and direct encodings for evolutionary computation

Evolutionary algorithms typically use direct encodings, where each element of the phenotype is specified independently in the genotype. Because direct encodings have difficulty evolving modular and symmetric phenotypes, some researchers use indirect encodings, wherein one genomic element can influence multiple parts of a phenotype. We have previously shown that Hyper-NEAT, an indirect encoding, outperforms FT-NEAT, a direct-encoding control, on many problems, especially as the regularity of the problem increases. However, HyperNEAT is no panacea; it had difficulty accounting for irregularities in problems. In this paper, we propose a new algorithm, a Hybridized Indirect and Direct encoding (HybrID), which discovers the regularity of a problem with an indirect encoding and accounts for irregularities via a direct encoding. In three different problem domains, HybrID outperforms HyperNEAT in most situations, with performance improvements as large as 40%. Our work suggests that hybridizing indirect and direct encodings can be an effective way to improve the performance of evolutionary algorithms.

How a Generative Encoding Fares as Problem-Regularity Decreases

It has been shown that generative representations, which allow the reuse of code, perform well on problems with high regularity (i.e. where a phenotypic motif must be repeated many times). To date, however, generative representations have not been tested on irregular problems. It is unknown how they will fare on problems with intermediate and low amounts of regularity. This paper compares a generative representation to a direct representation on problems that range from having multiple types of regularity to one that is completely irregular. As the regularity of the problem decreases, the performance of the generative representation degrades to, and then underperforms, the direct encoding. The degradation is not linear, however, yet tends to be consistent for different types of problem regularity. Furthermore, if the regularity of each type is sufficiently high, the generative encoding can simultaneously exploit different types of regularities.