The authors consider a setting where a supplier needs to allocate limited divisible resources among self-interested agents without the use of payments. They aim to design a mechanism that maximizes social welfare, as measured by Nash Social Welfare (NSW), while ensuring approximate incentive compatibility (IC) - i.e., minimizing the maximum utility gain agents can obtain by misreporting their preferences.
The key contributions are:
They propose a novel neural network architecture called ExS-Net that simulates "money-burning" (intentional withholding of resources) to align agents' incentives with the supplier's objective of maximizing NSW.
They provide generalization bounds for the ExS-Net mechanism, showing that the empirical objective converges to the true objective as the training dataset size increases.
Extensive experiments demonstrate that ExS-Net significantly outperforms the state-of-the-art hand-designed mechanisms, Proportional Fairness (PF) and Partial Allocation (PA), in terms of achieving a superior trade-off between NSW and exploitability (a measure of approximate IC).
They show that ExS-Net is robust to distribution mismatch between the training and test data, maintaining good performance even when the training data contains untruthful agent reports.
翻譯成其他語言
從原文內容
arxiv.org
深入探究