I now turn a critical eye on simulation work to ask the question. How has it helped us in solving the enigma of the origin of language? The first part of the section deals with problems with a totally self-organisational approach and the last part deals with simulation issues in general.

One of the key aspects of self-organising systems is the bottom-up development of structure. One of the prototypical examples of this phenomenon is the formation of near optimal paths to a food source in a species of ant, Linapithema humile (see Bonabeau and Theraulaz, 2000 for a fuller description). Though it is clear that an ant has no concept of optimising the time and energy spent getting to a food source, in spite of this, the optimal path is (often) obtained. The mechanism works in the following way. On returning to the nest from a plentiful food source an ant drops pheromone. Other ants are attracted to this pheromone trail, and so to the food source, but with a very weak trail, the ants will initially take very many different paths back to the nest. The number of ants increases and so does the amount of pheromone between the nest and the food source. The paths that are most direct to travel will, on average, be those ones that are reinforced the most and will, after a time, begin to have significantly more pheromone. Eventually ants will begin to use these paths exclusively. So it appears that an optimal path, sometimes a very long one through even the immensely varied jungle terrain is reached. This sort of result looks almost intelligent, though once we look at the mechanism we see clearly it is not. It is in this way that an increasing number of researchers wish to investigate the origins and evolution of language. In many ways, however, this is very strong claim. If it is an advantage to be a successful communicator (and most take this point for granted, though there are powerful arguments against this position, cf. Power, 2000; Dessalles, 2000) and all that is required for this are a few basic all-purpose learning mechanisms and a modicum of habitual association then how is it that only humans have language? We know even lower primates (e.g., Vervet monkeys, Cheney and Seyfarth, 1990) are capable of maintaining a digital lexicon, and capable of storing and processing intricate tapestries of social information. And what of the chimpanzees of Savage-Rumbaugh, who have shown the capability of acquiring at least a rudimentary form of what we might call a "protolanguage" (in the sense of Bickerton, 1990)? We must ask why this has not developed into full-blown language if the principles of self-organisation obtain in the real world.

The urge to over interpret results should be resisted. If it is a mistake to anthropomorphise the behaviour of the higher primates, who share so much with us in the real world, then it is surely a faux-pas of the worst kind to read anything into rudimentary simulations.

It is here we see a key problem in taking the self-organisation approach. Artificial agents are not even ants, let alone vervets or proto-humans. While this approach appears to have been fruitfully applied in the economic sphere (see Arthur et al., 1997 for examples) the fact that many aspects, such as differential advantage in communicating between the sexes - an aspect some claim to be of vital importance (e.g., Power, 2000) - are utterly ignored. It seems to me that, in essence, what is happening is that the lessons learnt over the past decade or two in the EOL debate are abandoned - for the sake of constructing manageable models with clear-cut results. Though this might bring simulation into the mainstream, it seems an intellectually questionable move.

It was noted above that there is a distinct tendency in current simulation experiments to deny any role within the model to genetic evolution. The rationale behind the move seems innocuous enough - the less we need to explain in terms of physical genetic evolution the better. We obviously have no way of knowing the exact phenotypic/genotypic progression in the hominid line, let alone the sequence of specific selective pressures that were present. Therefore, all structure that can be explained in terms of the interaction between low-level elements decreases the load on genetic explanations and should strengthen our theory. However, to ignore sexual (genetic) selection entirely is to fly in the face of established facts about how biological organisms interact with their environments over time.

It also seems highly unlikely that the physical prerequisites (short-term and long-term memory; vocal apparatus - serial gesture coordination, lowered larynx, etc) could have developed in the precise ways they did without any selective pressure. While it may be conceivable that general purpose mechanisms could be responsible for the development of 5-, 10-, or even, 50-item lexicons, modern humans have lexicons of tens of thousands, sometimes hundreds of thousands of items. The hardware necessary for this sort of processing simply could not have been left lying idle until one day the group suddenly started speaking to each other. By the same token, if we prefer the theory that the necessary hardware was already instated (as in the "sudden-change" theories e.g., Bickerton, 1990) and was simply coopted for speaking then at the very least it needs to be shown that the increased processing load was bearable. This "monster-mutation" theory of Bickerton is losing favour in the EOL debate however, and even Bickerton seems to be tempering his stance (see Bickerton, 2002). At the very least, the assumption of physical prerequisites is not unproblematic.

There are a number of language-specific adaptations, such as our sound-producing capabilities which have no analogue in the higher primate world. Why would humans develop sometimes dangerous adaptations, such as a lowered larynx, if not under pressure to increase the range of sounds in an already-present repertoire? None of these issues are adequately dealt with by the simulationists (though see the work of Daniel Livingstone and Colin Fyfe (e.g., 2000) for an attempt to address physiological issues).

It is not that none of the scholars are trying to address these issues - many are - but too often, it seems, we are told "future models will include" this or that. Luc Steels, for example, has done a good deal of recent work on robots, which he feels are more realistic, as they will introduce physical and temporal constraints not present in completely "soft" simulations. I wholeheartedly agree that this is a positive more, but it forces us to ask whether even the most sophisticated "soft" simulation is lacking in the vital ingredient of physical context/situation. While the robot's world is not the savannah of East Africa, it is considerably more open to the sort of indeterminacies that would have been present for our distant ancestors. Surely, also, this indeterminacy played a role in the development of language. Just as surely, it is lacking from the sterile cyber-worlds.

We might go even further, however, and ask whether this sort of modelling is appropriate for modern, rational, free agents such as humans. Simulationists assume that the essential properties of human rationality, action and agency have been formalised. This point leads into the issue of the computability of real world phenomena. There is an implicit assumption that the relevant aspects of a situation can all be translated into a finite bit string, and that this translation is unbiased and complete. While this is an area of debate in computer science and well out of the scope of this essay, suffice it to say that assuming that continuous, real-world phenomena can be represented by finite binary bit strings is not uncontentious. There is even a school of thought (social constructionism) that would claim that all of these things are by nature incomputable because their interpretation is "up for grabs" (see Shotter, 1993).

Can we really model just one aspect of human communication? (e.g., vowel systems?). Doing this suggests that these areas are either separate/independent from other parts of the system or, alternatively, that we know all of the other relevant factors and can accurately determine how and how much they affect the functioning and development of the whole. As is evidenced in the variety of theories to be found in any of the main texts coming out of the series of Evolution of Language conferences (Hurford et al, 1998, 2000; Wray, 2002), there is certainly no consensus on what the important factors were, nor how they might have interrelated.

MacLennan and Burghardt (1993) make the following comment, which is clearly germane: