The evolution of niche width

Abstract

This thesis examines the ultimate and proximate determinants of niche width, with a focus on how cognition and biological information processing may drive the evolution of niche width. Using both field and laboratory experiments I investigate how learning can alter resource use in syrphids. Modelling biological information processing using artificial neural networks I consider how various ecological factors interact and can impact information processing to determine decision accuracy (a proposed factor in the evolution of niche width). Finally the ability of artificial neural networks to overcome evolutionary dead ends due to specialisation and functional loss is examined. I found that syrphids were able to use external, inter-specific cues to alter their resource use. Specialist artificial neural networks decision accuracy was altered by the introduction of the ecological variables they were subjected to and the loss of functionality can create an evolutionary dead end scenario only in very extreme cases or under specific ecological pressures. I studied the syrphid (Episyrphus balteatus) both in the field and under laboratory conditions. There is a huge amount of literature describing how bees use scent marks to aid decision making before landing on flowers but there is currently no work on the syrphids ability to detect and utilise these scent marks. The question I posed was ‘Can syrphids modify their pattern of resource utilisation by using this scent mark information?’ The field work was carried out using motion detection cameras positioned above flowers of knapweed (Centaurea nigra). The flowers had two different treatments: one was bagged overnight to prevent pollinator access and the other was left unbagged allowing foraging insects to deplete the nectar and pollen. Visits from both conditions were recorded and compared. I found that previously bagged flowers received more visits from both bumblebees (Bombus spp.) and syrphids suggesting that syrphids could also detect when a flower was depleted without landing. iii The laboratory tests were conducted in an arena using artificial flowers. The experiment was split into a learning phase and a testing phase. I tested the syrphids ability to recognise and learn an association to two different compounds, bee scent marks or 1-Hexanol. I found that syrphids could learn to associate both bee scent marks and 1-Hexanol with negative rewards and use this information to change their foraging behaviour. I used artificial neural networks to investigate differences between the decision accuracy of specialists and generalists when foraging under ecological pressures. Previous work has shown that specialists had higher decision accuracy when non-host selection carried a mild reward and I was interested to see how ecological variables would impact this advantage. The ecological conditions I considered were search costs, resource availability and starvation. To do this I trained neural networks to recognise different numbers of binary images (hosts) over a range of positive and negative non-host rewards or punishments. The fewer hosts a network had the more specialised it was. I found that both starvation and resource availability reduced the range of non-host values across which specialist networks had a fitness advantage over generalists. Interestingly I found that introducing search costs shifts the range of non-host values where specialist advantage occurs rather than narrowing them as in the previous conditions. Specialists suffering from search costs performed better when non-host selection carried a high to intermediate punishment. Finally, I used artificial neural networks to investigate the evolutionary dead end theory. This theory states that specialist organisms will lose genetic variation and will be unable to respond as effectively to ecological change. I first trained networks as specialists. These networks were then re-trained as generalists. While re-training networks had a percentage of their weights fixed to simulate the suggested reduction in evolutionary potential of specialists. Ecological conditions in these simulations were either non-host penalties, search costs or a combination of the two. I found that networks were relatively robust to loss of evolutionary iv potential. All of the networks performed well even at intermediate (50%) weight fixation. The application of search costs reduced overall network fitness but this effect was not as pronounced as when non-host penalties were introduced. Non-host penalties had the greatest effect on the fitness of networks. These results suggest that specialisation should only become an ‘evolutionary dead end’ under very specific and severe conditions.