Chapter 6: Conclusion and discussion 6.1 Conclusion The leading question of this thesis is whether dialect formation in songbird populations can be a cause for complexity to emerge in their song. This question is intriguing, as it is counterintuitive: the complexity of a communication signal is rather expected to decrease than to increase as simple signals will be easier to learn. We propose a view on the emergence of complexity in human language as a side effect of the selection for diversification between dialects. This view on language evolution is radically different from more common instrumentalist views that focus on human language as a tool for survival. With our model we have tried to explore the viability of this idea. To start with, we focussed on the local interactions involved in the model, investigating whether a difference in complexity of dialects influences the settling chance of the agents. We validated this idea by showing that the fitness of agents dispersing from a complex dialect is higher than the fitness for agents dispersing from a simple dialect. However, a trade-off exists between getting extra fitness and losing signal complexity due to the symmetry of the interaction between the agents. As a consequence, the population neither diversified, nor started to use more complex signals. We chose two ways of exploration by adjustments to the model. We added predators to the population, in order to create more empty space, which causes a local lowering of the selection pressure. As a consequence, improvisation becomes cheaper, which favours the diversification of the signal. The predators indeed provoked diversification of the signal in the population, but did not cause the signal to become more complex. When the improvisation chance is made density dependent, agents that lengthen their songs seem have a fitness advantage. However, in an evolutionary setting higher improvisation rates hardly were selected for. Possibly the advantage for improvisers is frequency dependent. A second adjustment to the model was a change in the local interaction rules, in order to make the interaction less symmetric. We gave the agents the possibility to have a unique shared song with every neighbour. This individual song learning not only causes diversification of the signal, it also causes a strong selection pressure for longer songs. We showed that the song length as an evolutionary parameter increased only if dialect structure was present. 52 6.2 Limitations When addressing a problem by investigating the behaviour of a simulation model, the first question to ask is whether the problem at hand can be answered by such means. A model is a simplification and can never prove how the phenomenon it represents came about in real life. However, models can provide us with valuable insights on the consequences of the assumptions we make. Models enable us to explore what minimal set of assumptions is sufficient to produce certain behaviour. If these assumptions stick close to the constraints and properties observed in nature, a valuable analogy could be drawn, reaching us a possible cause for a natural phenomenon. Our model was simplified to formulate a minimal set of properties for the agents to show increasing signal complexity as a consequence of spatial patterns in the signal. Some remarks about the simplifications to the model: One of the simplifications we made for our first model was eliminating the differences between the songs a bird sings with different neighbours. Most of the songbirds in nature share songs individually with their neighbours. However, as we wanted to focus on the difference between groups of birds rather than on individual differences, it seemed a useful simplification to rule out individual interaction. Nevertheless, the first simulations made clear that exactly the individual nature of the interactions makes selection for complexity possible. Individual song sharing turned out to be important after all. A perhaps somewhat crude simplification is the discrete nature of the song elements. The song element of songbirds in nature and the words and their denotation in human language both change gradually over time and space. However, incorporating continuous signals in the model would force us to implement a categorisation mechanism in the agents, as no fixed barriers would separate the different song elements anymore. As object categorisation methods generally cause the behaviour of the model to become less tractable, we chose to make the signals discrete. As a direct consequence of this choice, problems turned up with introducing variation in the population in a proper way, which was solved by introducing an explicit improvisation mechanism. We decided to keep the agents in the model asexual. Sexual selection is known to select sometimes for signal complexity as for example can be observed in the tail of the peacock. Many argue, and do so quite convincingly, that the same accounts for bird song and human language. By incorporating female choice in the model the cause for the resulting complexity would become less clear. However, it remains obscure whether the influence of sexual selection on the complexity of bird song interferes with the effect we are investigating in the model. A choice rather than a simplification was the definition of complexity by song length. When the 53 emergence of complexity is investigated, a (local) definition of complexity has to be introduced. In our model, we chose the length of the bird’s song as a measure for complexity, because longer songs are more difficult to learn. We do not argue that the length of the song is the only way songs can get more complex, but we narrowed our model down to this option. Even though all our simplifications can be accounted for, the question remains whether the outcome of the model was not already encapsulated in the interaction rules. Although we stated that in a non-spatial model there is a strong tendency for the song to stay small due to learning costs. However, in our last model we showed that above certain song lengths, mixing the population did not cause small song lengths to be selected for anymore. Adding the cost to the model might solve this problem. 6.3 Extensions This model investigates the selection for complexity as a consequence of spatial pattern formation. However, we only intended to get a hunch of what is going on. We think that the construction of more sophisticated models as well as a more structural analysis of the present models can improve our insight in the proposed idea. It might also be interesting to look for experimental setups that might shed more light on the issue. An obvious direction for extending this research will be to investigate other possibilities to decrease the symmetry in the local interactions. As bird populations frequently show dominance hierarchies in the interaction with their neighbours, adding such a structure to the model population might be an extension to take in consideration. In Hogeweg & Hesper (1983), bumble bee civilisations are modelled by imposing a dominance interaction rule. Through self-organisation, a dominance hierarchy emerged with asymmetric relations between the members. One of the most promising extensions might be to make the used signal continuous instead of discrete. Continuous signals mean a more natural introduction of new song elements, as two signals drifting apart eventually become separate signals. However, with continuous elements a categorisation method is necessary to discriminate between various signals. For this purpose, a Kohonen map could be introduced. The Kohonen map is specific kind of neural network that enables self-organisation of the provided examples, so that they can be categorised without defining the categories. In order to get input that the network can handle, each element should then be constructed from a fixed set of properties like tone and pitch. A disadvantage of neural nets is the reduced tractability. Another possible extension is the addition of females to the model by letting them choose a mate 54 by his song. Female choice is known for its influence on the complexity of signals (see chapter 2), so it will certainly make the outcome more ambiguous. However, the assumption that dialect formation can be seen apart from female choice is questionable as females use dialect information for their choice. It would be interesting to see whether females would be able to take advantage from a correlation between dialects and fitness by favouring certain songs, whether longer or shorter. Theories always have to be falsifiable by experimental results. The predictions made by the model presented here regard the link between dialect formation and complexity of the signal. To our knowledge, no comparative research has been done on the level of song complexity between bird species living in a population with high or low variation in song. This might be interesting to look at. The same might be interesting to look at in the human population under the influence of globalisation. Cheap travelling and widespread Internet access cause an unprecedented intensity of contact between people from different language groups in the world, comparable to the repositioning procedure we used in the model. Already today it is observable that the immense linguistic diversity in the human species is disappearing with great speed, as a large portion of the languages is in danger of extinction. Probably within a few centuries, the people will all speak the same language. It would be interesting to look at the change of complexity of human language under the influence of this homogenising process. 55