Artificially Intelligent Foraging

Bumble bees (bombus spp.) are significant pollinators of many plants, and are particularly attracted to mass-flowering crops such as Oilseed Rape (Brassica Napus), which they cross-pollinate. B. napus is both wind and insect-pollinated, and whilst it has been found that wind is its most significant pollen vector, the influence of bumble bee pollination could be non-trivial when bee densities are large. Therefore, the assessment of pollinator-mediated cross-pollination events could be important when considering containment strategies of genetically modified (GM) crops, such as GM varieties of B. napus, but requires a landscape-scale understanding of pollinator movements, which is currently unknown for bumble bees. I developed an in silico model, entitled HARVEST, which simulates the foraging and consequential inter-patch movements of bumble bees. The model is based on principles from Reinforcement Learning and Individual Based Modelling, and uses a Linear Operator Learning Rule to guide agent learning. The model incoproates one or more agents, or bees, that learn by ‘trial-and-error’, with a gradual preference shown for patch choice actions that provide increased rewards. To validate the model, I verified its ability to replicate certain iconic patterns of bee-mediated gene flow, and assessed its accuracy in predicting the flower visits and inter-patch movement frequencies of real bees in a small-scale system. The model successfully replicated the iconic patterns, but failed to accurately predict outputs from the real system. It did, however, qualitatively replicate the high levels of inter-patch traffic found in the real small-scale system, and its quantitative discrepancies could likely be explained by inaccurate parameterisations. I also found that HARVEST bees are extremely efficient foragers, which agrees with evidence of powerful learning capabilities and risk-aversion in real bumble bees. When applying the model to the landscape-scale, HARVEST predicts that overall levels of bee-mediated gene flow are extremely low. Nonetheless, I identified an effective containment strategy in which a ‘shield’ comprised of sacrificed crops is placed between GM and conventional crop populations. This strategy could be useful for scenarios in which the tolerance for GM seed set is exceptionally low.