Akshat Gupta’s preprint offers a geometric interpretation of evolution strategies, an optimisation method that samples perturbed versions of a model, scores them and combines their directions into an update. The contribution is an explanation of an existing update rule, rather than a new benchmark-winning optimiser.
Two averages of the same population
The paper treats each perturbed model as a particle. Before scoring, all particles have equal mass, so their centre of mass is the ordinary population centroid. Assigning each particle its reward as mass produces a reward-weighted centre of mass.
For the mean-subtracted evolution-strategy update, the paper shows that the update is exactly a scaled displacement between these two points for a finite population. This makes the effect of reward centring visible in the geometry of one sampled update.
The physical interpretation has conditions
A literal mass interpretation needs nonnegative rewards and a positive total. The paper explains that adding a constant to every reward can arrange those conditions without changing the mean-subtracted update.
The ordinary centroid need not coincide with the current model for one finite, randomly sampled population. Under mirrored, antithetic sampling it does coincide, and it approaches the model as the population grows. That distinction is necessary when interpreting where the displacement starts.
Useful for understanding an implementation
The proposed view separates the model location, sampled centroid and reward-weighted centre. It can help readers understand why subtracting the mean reward changes the reference point used by the update.
For someone implementing or teaching evolution strategies, those quantities offer a concrete way to visualise a step and inspect its assumptions. The interpretation does not itself establish better training efficiency, final task accuracy or reduced compute. Those would require separate experiments beyond the geometric result described here.
Comments