In this paper, we study the design of deep learning-powered iterative combinatorial auctions (ICAs). We build on prior work where preference elicitation was done via kernelized support vector regressions (SVRs). However, the SVR-based approach has limitations because it requires solving a machine learning (ML)-based winner determination problem (WDP). With expressive kernels (like gaussians), the ML-based WDP cannot be solved for large domains. While linear or quadratic kernels have better computational scalability, these kernels have limited expressiveness. In this work, we address these shortcomings by using deep neural networks (DNNs) instead of SVRs. We first show how the DNN-based WDP can be reformulated into a mixed integer program (MIP). Second, we experimentally compare the prediction performance of DNNs against SVRs. Third, we present experimental evaluations in two medium-sized domains which show that even ICAs based on relatively small-sized DNNs lead to higher economic efficiency than ICAs based on kernelized SVRs. Finally, we show that our DNN-powered ICA also scales well to very large CA domains.

翻译：本文研究基于深度学习的迭代组合拍卖（ICA）的设计方法。我们在先前采用核化支持向量回归（SVR）进行偏好诱导的研究基础上展开工作。然而，基于SVR的方法存在局限性，因其需要求解基于机器学习（ML）的胜者决定问题（WDP）。当使用高表达力的核函数（如高斯核）时，基于ML的WDP在大规模领域内无法求解。虽然线性或二次核函数具有更好的计算可扩展性，但其表达力有限。本研究通过采用深度神经网络（DNN）替代SVR来克服上述不足。我们首先论证如何将基于DNN的WDP重新表述为混合整数规划（MIP）；其次，通过实验对比DNN与SVR的预测性能；再次，在两个中等规模领域的实验评估表明，即使基于较小规模DNN的ICA也比基于核化SVR的ICA具有更高的经济效率；最后，我们证明基于DNN的ICA能够很好地扩展至超大规模组合拍卖领域。