Department of Linguistics
University of Edinburgh
hoplite@usa.net

Despite the considerable differences of surface structures of various languages, it can be agreed that all natural languages are equally complex; perhaps the most complex system in any cognitive faculties. The question arises, then, how any child, wherever they may be in the world, can begin to acquire such a complicated system. Everyone intuitively knows linguistic input plays a crucial role in language acquisition. Linguistic input, however, employed to construct knowledge of a language is often ill-formed, incoherent, and most importantly, insufficient (Chomsky 1965). Thus claims have been made that the process of language acquisition is neither completely a process of postnatal learning nor a product of innately fully specified linguistic knowledge. Chomsky’s original formulation of the nature of the language acquisition device (LAD) and its core theory –Principles & Parameters theory- (Chomsky 1981) were introduced to solve this embarrassing complication of language acquisition. This strategy tries to come to terms with this complication by putting forward both aspects of language acquisition as being equally significant.

Recent surveys in the field of computational simulations reincarnate a more-than-100-years-old argument in evolutionary study. In 1896, James Mark Baldwin proposed "a new factor in evolution" (Baldwin 1896, Morgan 1896, Osborn 1896, Waddington 1942, Turney, Whitely, and Anderson 1996). He assumed that if an individual is capable of acquiring an adaptive feature postnatally, addition of such a learning process in the context of evolutionary search potentially changes the profile of populational evolution. In a nutshell, the Baldwin effect is an interaction between evolution and learning, where "a behavior that was once learned may eventually become instinctive" (Turney, Whitley, and Anderson 1996). The possibility of this learning-guided evolution has been repeatedly attested in computer simulations by a number of researchers (Hinton & Nowlan 1987, Maley 1996a, 1996b, 1997, Nolfi, Elman, and Parisi 1994, French & Messinger 1994 and more).

It has been a popular idea that the Baldwin effect is a crucial factor of the evolution of language (e.g. Pinker & Bloom, 1990, Briscoe 1997). The learning-guided evolution scenario possibly provides a strikingly attractive solution to a longstanding problem. Preliminary studies suggest that language evolution is out of the scope of natural selection mainly because of its dysfunctional nature. For those researchers, language evolution is a consequence of exaptation or a big leap in evolution (Newmeyer in preparation, Piatelli-Palmarini 1989). This no-intermediate scenario would be, however, explicable by natural selection when it is guided by learning since learning can bridge the gap as Hinton & Nowlan showed (Hinton & Nowlan 1987). There is a further advantage of the Baldwin effect in the evolutionary study of the LAD. Its combination of genetically hardwired features and postnatal learning processes are perfectly compatible with Chomsky’s P&P theory. Together with its "genetic assimilation" process (Waddington 1975), the Baldwin effect may shed a light on the nature of the current relationship between innateness and postnatal learning in language acquisition.

However, it has been a matter of concern for long time that a straightforward representation of the relationship between genotype and phenotype is unrealistic, especially when the expressed phenotype is beyond basic biological expression. It is unlikely that properties of a certain higher cognitive ability are partitioned by independent genes. It might be more plausible to consider that multi-contribution of genes to their expressions or cascade-like reactions of single gene’s expression shape the actual LAD. (Waddington 1942, Deacon 1997, Newmeyer in preparation).

My study examined the real possibility of the effect on the evolution of the LAD under more genetically realistic circumstances based on Turkel’s simulation (Turkel, to appear)

First, a complete replication of Turkel’s simulation (To appear) was tested. A population of agents with genetically represented principles and parameters was evolved on the basis of a basic genetic algorithm. Thus the evolution of such agents reflects the evolution of the LAD itself. At the end of the simulation, as Turkel revealed, all agents were converged into a single genotype. The unified genotype is randomly determined in each run. Regarding the dysfunctional aspect of language, this randomeness is significant. Each configuration was attained regardless of any external factors; only the dynamic aspect of communication within the community contributes to this arbitrariness. This has implications against the "anti natural selection" theories in language evolution (the "no external reality" theory; Fodor 1989). More importantly, however, the Baldwin effect was observed in this simulation; the number of plastic genes is notably decreased at the end of each run (Fig. 1).

Therefore, agents were successful to establish communication with relatively small consumption of learning trials. This result might be compatible with rapid acquisition of a natural language in real world.

Then, to test the effect of a more complex relationship between genotypes and their phenotypic expressions in Turkel’s simulation, I incorporated Stuart Kauffman’s NK-Landscape model (Kauffman 1989). In this model, two or more genes express one feature of the phenotype based on a randomly generated look-up table. This means that our model embodies epistasis and pleiotropy in genes. The modification also puts genotypes and phenotypes on completely independent strata while Turkel’s original simulation partially conflates these two representations. Effectively, this modification blurs the relationship between genes and their phenotypic expressions. The degree of abstractness is controlled by the value of K that specifies how many genes are needed to express one feature of the phenotype. The results were remarkable; although convergence toward a single genotype was still observed, subsequent emergence of the Baldwin effect was severely suppressed under these circumstances. In the worst case, maximum value of K, no plastic genes were replaced by fixed genes even after 200 generations (Fig. 2).

Finally the results of the simulation require reconsideration of the evolutionary scenario of the LAD. Either it has traced a completely different evolutionary path or the LAD is equipped with an extremely robust learning mechanism so that even with high plasticity it can learn a language without failure.

 Conference site: http://www.infres.enst.fr/confs/evolang/