Evolving neural networks for optimal foraging
If a neural network produces an output that switches at random, can evolution be used to shape this noise? Can this help explain the random movement patterns of animals foraging for food?
There is ongoing debate regarding the strategies animals use to explore their environment, namely how they divide their time and energy between long and short-range exploration. Recent work using evolutionary algorithms (Campeau et al. 2024) has found that the optimal distribution of steps is bimodal (i.e. an intermittent process), rather than a long-tailed curve (e.g. Levy flights). However, the mechanism by which such a probability distribution can evolve is unclear. There are several neuron-like network models that can allow for time-varying output with complex noise distributions, such as random recurrent networks (Pontes-Filho et al. 2020) and heteroclinic networks (Horchler et al. 2015). This project will perform experiments with such models adapted to controlling foraging agents. Parameters will form a genotype, evolving to produce more successful foragers via natural selection, reproduction and mutation. The resulting dynamics will be compared with eye-tracking data obtained by the Virtual Eye project members at Simula and OsloMet.
Goal
The goal of the thesis is to answer the following research questions:
- Can a single neural network model produce a variety of noise distributions seen in e.g. eye-tracking data and animal movements?
- How does evolution act on the network under selective pressure to forage efficiently?
Learning outcome
The student will become intimately familiar with dynamical systems modelling, stochastic processes, and evolutionary algorithms. This will benefit any student interested in bio-inspired AI and robotics, as well as computational neuroscience.
Qualifications
The student should be comfortable with differential equations and time series analysis, as well as coding in Python with numpy. Students undertaking cybernetics or robotics programs will have an advantage.
Supervisors
- Alexander Szorkovszky
Collaboration partners
- OsloMet AI Lab
References
- Campeau, W., Simons, A. M., & Stevens, B. (2024). Intermittent Search, Not Strict Lévy Flight, Evolves under Relaxed Foraging Distribution Constraints. The American Naturalist, 203(4).
- Horchler, A. D., Daltorio, K. A., Chiel, H. J., & Quinn, R. D. (2015). Designing responsive pattern generators: stable heteroclinic channel cycles for modeling and control. Bioinspiration & biomimetics, 10(2), 026001.
- Pontes-Filho, S., Lind, P., Yazidi, A., Zhang, J., Hammer, H., Mello, G. B., ... & Nichele, S. (2020). A neuro-inspired general framework for the evolution of stochastic dynamical systems: Cellular automata, random Boolean networks and echo state networks towards criticality. Cognitive Neurodynamics, 14(5), 657-674.
