On the Influence of Lifetime Learning on Selection Pressure

Evidence for both, acceleration and deceleration of evolution through learning can be found in the literature. We suggest a selection gradient model that allows to predict whether acceleration or deceleration is predominant. The main idea is that learning alters the genotype-to-fitness landscape, which determines selection pressure. Assuming that fitness can be split up into an innate and a learned component, conditions for the occurrence of the Baldwin and the Hiding effect are derived. We introduce learning curves to analyse what we term lifetime fitness. The influence of the shape of the learning curve on the interaction between evolution and learning is analysed. Introduction The interaction between evolution and learning has been studied from different perspectives. Straightforwardly, evolution influences learning by resulting in individuals that have the ability to learn. In artificial life, this mechanism is employed e.g. in evolutionary robotics (Harvey et al., 2005) and in virtual environments (Todd and Miller, 1991; Niv et al., 2002). Furthermore, learning also influences evolution. Learning may guide a population toward evolutionary paths, which would not have been taken in the absence of learning (Hinton and Nowlan, 1987). Besides this exploratory influence, learning also affects the rate of evolution (Keesing and Stork, 1991) if learning alters the selection pressure. These phenomena are usually subsumed under the terms Baldwin effect (Baldwin, 1896; Simpson, 1953) and Hiding effect (Mayley, 1997). In the next section, we briefly review the Baldwin and the Hiding effect and highlight some examples from the literature that have demonstrated either or both of these effects. Thereafter, we utilize a common method from quantitative biology, known as the selection gradient analysis (Lande and Arnold, 1983), to mathematically derive conditions under which selection pressure is increased (Baldwin effect) or decreased (Hiding effect). Based on this, we investigate how lifetime learning curves influence the Baldwin and the Hiding effect. We demonstrate that learning, even in a form where the learning curves are identical for all individuals in a population, can increase or decrease the selection pressure. Furthermore, we discuss the importance of the convexity and concavity of learning curves emphasizing their influence on the selection pressure. The influence of learning curves is important for both biological and artificial systems, whenever the system’s behavior is relevant throughout its lifetime, i.e., (already) while it is learning. Baldwin vs. Hiding Effect In the literature, the Baldwin effect has been widely discussed, but apparently there exists no clear definition of “the” Baldwin effect. As Depew states: The Baldwin effect does not reliably refer either to a theory-neutral empirical phenomenon, or to a single hypothesis, or to an identifiable mechanism (Depew, 2003, page 5). Seemingly, the term Baldwin effect subsumes a number of effects arising from the interdependency of phenotypic plasticity and genetic evolution. In this paper, we use the term Baldwin effect to describe the increase in selection pressure as an effect of learning (unfortunately, a comprehensive review of the Baldwin effect is beyond the scope of the paper). If learning alters the fitness landscape such that only the (genetically) very good individuals of the population reproduce, selection pressure is increased and evolution is accelerated. The other extreme, learning may allow innately weak individuals to catch up with innately strong individuals, and thus selection pressure is reduced. In the latter case, learning causes a reduction of fitness differences between weak and strong individuals and evolution is decelerated (Johnston, 1982). In (Mayley, 1997) this effect is named Hiding effect. Figure 1 illustrates both effects. In the artificial life and biological literature, examples can be found for both effects. In Hinton and Nowlan’s computer experiments learning leads to an acceleration of evolution (Hinton and Nowlan, 1987) and an analytical treatment supports the simulation results (Fontanari and Meir, 1990). Dopazo et al. find a halting effect of learning in an Baldwin effect Hiding effect in na te p he no ty pe ge no ty pe