Abstract: Throughout modern society, human users interact with large-scale engineered systems, e.g., road-traffic networks, electric power grids, wireless communication networks. As the performance of such systems greatly depends on the decisions made by the users – often leading to undesirable system behaviour – a natural question arises: how can we design incentives to promote efficient use of the existing infrastructure? In this talk, I answer this question in relation to the well-studied class of atomic congestion games, used to model a variety of problems arising in theory and practice, including traffic routing. More precisely, I show how to design incentives (congestion-pricing schemes) that utilize only local information, and robustly maximize the system efficiency. The proposed approach reduces this difficult combinatorial problem to a tractable linear program, recovering and generalizing a number of well-known results in the computer science, economics, and transportation literature. Surprisingly, optimal incentives designed using only local information perform closely to those designed using full information, while marginal cost tolls, known to be optimal in the continuous-flow model, have unexpectedly lower efficiency than that encountered levying no toll at all.