The influence of bumble bees as a pollen vector in crops has led to concerns about the levels of bee-mediated gene flow in systems containing GM varieties. Assessing potential gene flow requires an understanding of movement between fields and the number of flowers visited in each. Empirically measuring movements of bumble bees at the landscape-scale has proven difficult which means that predictive models are valuable.

Modelling how foraging animals move between patches of resource has long been investigated, but models typically focus on forager distributions. Instead, we wanted to quantify inter-patch traffic in relation to patch sizes and configuration.

We therefore developed a patch choice model for bees based on Reinforcement Learning which originates in Artificial Intelligence. Our model bees learn by trial-and-error sampling.

Results/Conclusions

We tested whether the model was able to (1) predict known patterns of bee behaviour and (2) observations from an empirical study. We found quantitative agreement between the model predictions and empirical data. Also, the model makes a general prediction that systems containing more variable standing crop levels yield less inter-patch movement. We extrapolated the model to the landscape-scale and resultant gene flow calculations predict bee-mediated gene flow levels to be extremely low at the landscape-scale, but increased gene flow is predicted when fields have similar standing crop levels.

We have developed a patch-choice foraging model that begins to predict bumble bee behaviour at the small-patch scale. We propose that our Reinforcement Learning based approach could be extremely useful in understanding foraging behaviours in bumble bees and other animals.