Commentary on Altruism and Selfishness, by Howard Rachlin
Abstract: 66 words
Main Text: 1067 words
References: 204 words
Total Text: 1383 words
The new emphasis that Rachlin gives to social learning is welcome, because its role in the emergence of altruism and communication is often underestimated. However, Rachlin's account is underspecified and therefore not satisfactory. I argue that recent computational models of the evolution of language show an alternative approach, and present an appealing perspective on the evolution and acquisition of a complex, altruistic behavior like syntactic language.
In the target article, Rachlin calls attention to the role of social learning in the emergence of altruistic behavior in humans. This shift of emphasis in thinking about altruism has intriguing consequences. Acknowledging the important role of learning leads one to ask at least three new and challenging questions: (1) about the exact mechanisms by which altruistic behavior emerges in learning and development; (2) about the ways in which the existence of learning mechanisms has changed the evolutionary process; and, vice versa, (3) about the ways in which evolution has shaped the learning mechanisms that lead to altruism. We can no longer - as is common in traditional game-theory - ignore the intricate mapping between genotypes (the genes) and phenotypes (the behaviors) and the strong dependence of this mapping on the individual's (cultural) environment.
Rachlin's paper is a welcome effort to underline this point, but I think his explanation for the emergence of altruistic behavior in humans suffers from underspecification: some crucial concepts are too loosely defined to make it possible to really agree or disagree with his analysis. I will discuss Rachlin's answers to the questions above from this perspective, and then try to show that some recent computational models in the related field of the evolution of communication offer a more precise account of the evolution of altruistic behavior.
Rachlin's answer to the first question is a mechanism similar to self-control. Humans discover that choosing for a whole pattern of altruistic activities is in the end more rewarding than repeating alternative, selfish activities, even though the latter offer more short-term benefits. The problem with this account is that it is unclear what constitutes a ``pattern''. Without a theory on how individuals represent and acquire this knowledge, we can never identify the different strategies that individuals can choose from.
A related problem arises for Rachlin's implicit answer to the second question. Rachlin gives the example of a woman that puts her life in danger to rescue someone else's child. His explanation of her brave behavior rests on the crucial assumption that the woman at some point in her development had to choose between life-long altruism or life-long selfishness. If there are only these two choices, and if the first choice is indeed more profitable in the long run, natural selection of course favours the tendency to choose it. However, Rachlin gives no arguments for why the choice would be so constrained. I find it difficult to accept that with all the subtle influences that genes have on our behavior, selectively avoiding life-threatening situations was not a possibility.
Rachlin's implicit answer to the third question is no solution to that objection. Essentially, he explains the evolution of altruistic behavior by claiming that it is not really altruistic after all. Altruistic - at least in its traditional sense in evolutionary game theory (Maynard Smith, 1982) - are those behavioral strategies that benefit others, but harm the individual that employs them even though less harmful strategies are available. A game-theoretic analysis of the evolution of alarm-calls in certain bird species (Maynard Smith, 1982) therefore emphasizes evidence that the calls really are harmful and that other strategies are really available. In contrast, the altruistic strategy in Rachlin's scenario is in the long run advantageous, and better alternatives are not available; it is thus not ``really'' altruistic in the traditional sense.
Rachlin acknowledges this, but he does not mention that the analogy between his explanation and group selection therefore breaks down. Group selection, like kin selection, is a mechanism that is capable of explaining ``real'' altruism. The decrease in the fitness for the individual is explained by assuming a higher or lower level of selection, i.e. that of the group or that of the gene. Thus, the fitness of a worker bee that does not get any offspring really is low (it is 0 by definition), but the fitness of the whole colony or the fitness of the genes that cause her sterility is high. The empirical validity of these explanations remains controversial, but their explanatory power remains appealing.
Researchers in the related field of language evolution have already explored many aspects of the interactions between learning and evolution. Language is a complex behavior, that is at least in some instances used for altruistic purposes (of course, sometimes selfish motives like intimidation, manipulation and encryption can also play a role). The population as a whole benefits from the altruistic use of language, like it does from other altruistic behaviors. In particular, the population benefits from using syntactic language (Pinker and Bloom, 1990), but it is not trivial to explain how an individual that uses syntax can be successful in a non-syntactic population.
By using a methodology of computational modeling that avoids the underspecification of Rachlin's arguments, researchers in this field have shed some new light on how this behavior has emerged (Hurford, 2002; Steels, 1999). E.g., these models have shown that when individuals learn language from each other with rather generic learning mechanisms, a rudimentary syntax can emerge without any genetic change (Batali, 1998; Kirby, 2000). The used learning algorithms, e.g. a recurrent neural network model in (Batali, 1998), provide - although far from final - a fully specified candidate answer to the first question we posed above.
Similarily, in recent work I have explored some provisional answers to the second and third question. In (Zuidema, 2002) I explore the consequences of the fact that language itself can, in the process of learners learning from learners, adapt to be better learnable (Kirby, 2000). As it turns out, this cultural process facilitates the evolutionary process. Evolutionary optimization becomes possible, because the cultural learning process fulfills the preconditions for a coherent language in the population. Moreover, the model also shows that much less of the ``knowledge of language'' needs to be innately specified than is sometimes assumed. Cultural learning thus lifts some of the burden of genetic evolution to explain characteristics of language. Alternatively, Zuidema and Hogeweg (2000) present results of a spatial model of language evolution. These results show that syntax can be selected for through a combined effect of kin selection and group selection.
These answers are for from final, but I believe that such well-defined models present an appealing perspective on how cultural learning can lead to the successful acquisition and creation of a complex, altruistic behavior like syntactic language, and why the learning mechanisms operate the way they do.
Batali, J. (1998). Computational simulations of the emergence of
grammar. In: Hurford, J. and Studdert-Kennedy, M., editors,
Approaches to the evolution of language: social and cognitive
bases. Cambridge University Press.
Hurford, J. R. (2002). Expression / induction models of
language. In: Briscoe, T., editor, Linguistic Evolution through
Language Acquisition: Formal and Computational
Models. Cambridge University Press.
Kirby, S. (2000). Syntax without natural selection: How
compositionality emerges from vocabulary in a population of learners.
In: Knight, C., Hurford, J., and Studdert-Kennedy, M., editors,
The Evolutionary Emergence of Language: Social function and the
origins of linguistic form. Cambridge University
Maynard Smith, J. (1982). Evolution and the Theory of
Games. Cambridge University Press.
Pinker, S. and Bloom, P. (1990). Natural language and natural
selection. Behavioral and brain sciences, 13:707-784.
Steels, L. (1999). The puzzle of language evolution.
Zuidema, W. (2002). Language adaptation helps language acquisition -
a computational model study. In: From Animals to Animats:
Proceedings of the 7th International Conference on the Simulation of
Adaptive Behavior (submitted).
Zuidema, W. and Hogeweg, P. (2000). Selective advantages of syntactic language: a model study. In: Proceedings of the 22nd Annual Meeting of the Cognitive Science Society, pages 577-582. Lawrence Erlbaum Associates. http://arti.vub.ac.be/~jelle/research/cogsci2000.pdf
The author funded by a Concerted Research Action fund (G.O.A.) of the Flemish Government and the Vrije Universiteit Brussel.