7. Modeling Evolution: Evolutionary Computation

By Luis M. Rocha

Lecture notes for I485 - : Biologically Inspired Computing. School of Informatics, Indiana University. Also available in adobe acrobat pdf format

“How does evolution produce increasingly fit organisms in environments which are highly uncertain for individual organisms? How does an organism use its experience to modify its behavior in beneficial ways (i.e. how does it learn or ‘adapt under sensory guidance')? How can computers be programmed so that problem-solving capabilities are built up by specifying ‘what is to be done' rather than ‘how to do it'?" [Holland, 1975, page 1]

These were some of the questions concerning John Holland when he thought of Genetic Algorithms (GA's) in the 1960's. All these questions were shown to be reducible to a problem of optimizing multi-parameter functions. Nature's "problem" is to create organisms that reproduce more (are more fit) in a particular environment: the environment-organism coupling dictates the selective pressures, and the solutions to these pressures are organisms themselves. In the language of optimization, the solutions to a particular problem (say, an engineering problem), will be selected according to how well they solve that problem. GA's are inspired by natural selection as the solutions to our problem are not algebraically calculated, but rather found by a population of solution alternatives which is altered in each time step of the algorithm in order to increase the probability of having better solutions in the population. In other words, GA's, or other Evolutionary Strategies (ES) such as Evolutionary Programming (EP), explore the multi-parameter space of solution alternatives for a particular problem, by means of a population of encoded strings (standing for alternatives) which undergo variation (crossover and mutation) and are reproduced in a way as to lead the population to ever more promising regions of this search space (selection) [Goldeberg, 1989; Mitchell, 1999; De Jong, 2006].

7.1 Evolutionary strategies and self-organization

The underlying idea of computational ES is the separation of solutions for a particular problem (e.g. a machine) from descriptions of those solutions (memory). GA's work on these descriptions and not on the solutions themselves, that is, variation is applied to descriptions, while the respective solutions are evaluated, and the whole (description-solution) selected according to this evaluation. Such machine/description separation follows von Neumann's self-reproducing scheme (see chapter 6) which is able to increase the complexity of the (organization of) machines described. Therefore, the form of organization evolved by GA's is not self-organizing in the sense of a boolean network or cellular automata (see chapter 4). Even though the solutions are obtained from the interaction of a population of elements, and in this sense following the general rules usually observed by computationally emergent systems (e.g. Langton [1988], Mitchell [1992]), they do not self-organize since they rely on the selective pressures of some environment (in ES, defined by an explicit or implicit fitness function). The order so attained is not a result of the internal dynamics of a collection of interacting elements, but is instead dictated by the external selection criteria. In this sense, ES follow an organizing scheme that is driven by external selection of encoded symbolic descriptions (a “Turing tape” ). It is perhaps useful to think that ES are modeling the most fundamental design principle of biological systems: natural selection. While self-organizing systems model the dynamical characteristics of matter, ES model the existence of, external, selective pressures on populations of symbolic descriptions of some system. While self-organization models material dynamics, ES models the selection of information about dynamics.

7.2 Development and morphogenesis: self-organization and selection come together

Since the original introduction of GA's, many subsequent developments had to do with the inclusion of a developmental stage, or intermediate layers between genotype and phenotype; in other words, the creation of some artificial morphogenesis or regulation. The idea has been to encode rules that will themselves self-organize to produce a phenotype, rather than the direct encoding of the phenotype itself, or the introduction of gene regulation . As discussed in class, these rules often use L-System grammars which dictate production system programs [Wilson, 1988] leading to some phenotype. The most important advantage of this intermediate stage, as explored by Kitano [1990], Gruau [1993], Belew [1992] and others, is the ability to code for much larger structures than a direct encoding allows. In practical terms, they have solved some of the scalability problems of encoding (e.g.) neural networks in GA's, by reducing the search space dramatically.

TL-system grammars are higher-level descriptions of self-organizing developmental processes. However, these first approaches used solely context-free, state-determined, L-System grammars, compromising epistasis (or mutual, non-linear, influence of genetic descriptions amongst each other) in the simulation of self-organizing development. Dellaert and Beer [1994] and Kitano [1994], for instance, used Boolean networks to simulate genetic epistasis and self-organization. In other words, the GA encodes rules which construct Boolean networks whose nodes stand for aspects of the phenotypes we wish to evolve on some physical simulation. In Dellaert and Beer’s model, the nodes stand for cell mitosis and other characteristics. This way, the solutions of the GA are self-organizing systems whose attractor behavior dictates pre-defined phenotypic traits.

These approaches in effect offer an emergent morphology, that is, they encode rules which will themselves self-organize into some phenotype (instead of strict programming of morphology). The indirect encoding further allows the search to occur in a reduced space, amplified through development. An interesting side effect has to do with the appearance of modularity traits on the evolved phenotypes [Wagner, 1995]. Subsequent developments paid even more attention to the contextual regulation that indirect encodings afford to the search [Rocha 1995, 1997]. More recently, given our expanded view of genomics, other intermediate layers between genotype and phenotype have been explored, such as transcription regulation [Reil, 1999; Hallinan & Wiles, 2004] and RNA Editing [Rocha et al, 2006]. The inclusion of more sophisticated regulation of genetic information prior to translation, while not necessarily including a self-organizing component, allows us to model a much more realistic genotype/phenotype/environment interaction. Instead of genotypes used exclusively for Mendelian inheritance (see chapter 5) of (directly encoded) phenotypic traits, ES with genotype regulation allow us to model the contextual, plastic development of phenotypes we have come to understand via modern Genomics—thus also learning additional design principles for bio-inspired computation [Huang et al, 2007].

The most important aspect of GA’s with emergent morphologies is the utilization in the same model of an external selection engine (the GA) coupled to a particular self-organizing dynamics (e.g. Boolean networks) standing for some materiality. Such schemes bring together, computationally, the two most important aspects of evolutionary systems: self-organization and selection. These models belong to a category of self-organization referred to as Selected Self-Organization which is based on symbolic memory [Rocha, 1996, 1997, 1998]. Selected Self-Organization with distributed memory is also possible in autocatalytic structures, though its evolutionary potential is much smaller than the local memory kind [Rocha, 2001][Vasas, 2010]. The reason lies in Von Neumann’s notion of self-reproduction (see chapter 6). The introduction of symbolic descriptions allows a much more sophisticated form of communication: structures are constructed from static descriptions and do not have to reproduce through some complicated, and limited process of self-inspection. In other words, separate descriptions can be used to reliably construct any kind of structure in an open-ended manner, while self-inspection relies on only those structures that happen to be able to make copies of themselves. As an example, a non-genetic protein-based life form, would have to rely only on those proteins that could make direct copies of themselves [Rocha, 2001].

7. Modeling Evolution: Evolutionary Computation

By Luis M. Rocha

7.1 Evolutionary strategies and self-organization

7.2 Development and morphogenesis: self-organization and selection come together

Further Readings and References