Evaluating the Transfer of Co-Evolved Communication from 2D to 3D Simulation

Evaluating the Transfer of Co-Evolved Communication from 2D to 3D Simulation. By Fernando Montes-Gonzalez, from the Universidad Veracruzana. Published on the seventh of October.

Two robotic agents co-evolve a way of communicating inside a discrete, two-dimensional simulator. The question is what happens when that mechanism is moved into a three-dimensional simulator with real physics: does it keep its functional role? The central observation is about what that role is. In both simulators, the social cue functioned primarily as a jam-assistance mechanism rather than as a guide for navigation, and the effect was more pronounced in the two-dimensional setting than in the physics-based one.

To get there, the study examines three factors: the time budget of an episode, the social cue itself, and the asymmetry between the two co-evolved roles.

On time, success rose roughly linearly across the budgets evaluated, with no sign of a plateau between two thousand and six thousand physics steps. The reading offered is that evaluations built on shorter episodes may underestimate how well the trained controllers perform. On asymmetry, eight independent evolutionary runs all showed the same direction of asymmetry between the roles, though how large it was varied from run to run. Controlling the order in which the two agents are processed allowed the author to rule out the possibility that this asymmetry was an artifact of the physics engine.

Several limits come with these results. A performance gap remains after transfer, and the work discusses factors that may contribute to it rather than closing it. The time-budget result covers only the range tested; the absence of a plateau between two thousand and six thousand steps says nothing about what happens beyond that. The asymmetry finding rests on eight runs, and while its direction held, its size did not. And all of it concerns one pair of agents and one co-evolved mechanism, moved between two particular simulators.

Two points carry the weight. The cue's role survived the move between simulators in kind, assisting with jams more than guiding, while its strength did not fully carry over. And the length of an evaluation is itself a variable: with success still climbing at six thousand steps, a shorter episode could understate what these controllers can do.