• Citer

Assessing and forecasting the performance of optimization-based design methods for robot swarms: Experimental protocol and pseudo-reality predictors

par Ligot, Antoine
Président du jury Dorigo, Marco
Promoteur Birattari, Mauro
Publication Non publié, 2023-03-31

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Which off-line fully-automatic optimization-based design method produces control software that will yield the best performance once executed on a swarm of physical robots? This fundamental question cannot currently be answered due to the lack of two elements: i) an appropriate experimental protocol for the evaluation and comparison of fully-automatic design methods, and ii) a procedure that reliably predicts the real-world performance of control software. This dissertation focuses on addressing this void.The literature on optimization-based design of robot swarms suffers from the absence of a clearly established state of the art. It is in fact, for the most part, a collection of feasibility studies, and little effort has been devoted to the systematic empirical evaluation and comparison of the methods or ideas proposed. Recent papers have formally characterized two approaches to the optimization-based design that were entangled in the literature: the fully-automatic and the semi-automatic approaches. In light of this novel categorization, we show that the experimental protocols employed so far for the evaluation and comparison of design methods do not respect the tenets of fully-automatic design, and we propose one that does.One of the most challenging issues when designing control software off-line on the basis of a simulation model is the reality gap: the unavoidable discrepancies between the design model and reality. It is generally understood that, because of the reality gap, the design model overestimates the performance that control software eventually yields when executed on physical robots. As a result, conducting expensive and time consuming tests on physical robots is mandatory to reliably assess control software. We introduce the concept of pseudo-reality: a simulation model, different from the one used in the design, whose purpose is to evaluate control software. With this concept, we show via a series of experiments that the reality-gap problem is to be understood as a generalization problem, akin to the one encountered in machine learning. We also use it to conceive several simulation-only predictors of real-world performance, and we assess their accuracy with a large dataset of observations collected from previous studies. Results show that the pseudo-reality predictors we propose are more accurate than the current practice for predicting the expected performance of control software for robot swarms.