Symmetry-Reduced State Spaces Improve Reinforcement Learning Sample Efficiency

This method uses category theory to partition high-dimensional state spaces into equivalence classes via symmetry orbits. By learning on a symmetry-reduced state space rather than treating every orientation as a unique state, agents improve sample efficiency in partially observable environments.