Maybe the hypothesis that environmental modularity drives agent modularity can be tested empirically using something like Deepmind’s XLand. Fix the size of the space of tasks in the training distribution. Then find some way to define environmental modularity of a subset of tasks. Then test: do RL agents trained on a modular space of tasks have a more modular structure than RL agents trained on an arbitrary space of tasks?
Maybe the hypothesis that environmental modularity drives agent modularity can be tested empirically using something like Deepmind’s XLand. Fix the size of the space of tasks in the training distribution. Then find some way to define environmental modularity of a subset of tasks. Then test: do RL agents trained on a modular space of tasks have a more modular structure than RL agents trained on an arbitrary space of tasks?