Our project focuses on microcircuits, particularly examining the effects of a swimmer’s physics and learning how to interact with new environments of varying viscosities. We are interested in understanding how robustness is influenced by learning paradigms and physical architecture. Ultimately, we aim to determine the optimal joint length and number for achieving robust target performance.
To investigate the performance of agents with different physical abilities in varying environments, we need to address the following questions:
- We will train several models of the same swimmer with a constant length but varying numbers of joints or joint lengths in both high and low viscosity environments.
- We will measure the agent’s performance (e.g., time to reach a target, distance from the optimal trajectory) in new environments to assess robustness.
- We hypothesize that optimal dexterity in generalization is implied by the model’s architecture.
- We will work with model connectome features and environmental properties under our control to plan the details of our experiment.
- We will investigate the effects of changing model and environment parameters on the trained networks.
- Finally, we will test the models' performance in new environments to establish the most robust strategy.
This structured approach will allow us to explore how different physical architectures impact the learning and robustness of swimmer agents in varying viscosity environments.