Prices, Bids, Values: One ML-Powered Combinatorial Auction to Rule Them All

A team of researchers from the University of Zurich has unveiled MLHCA, a groundbreaking machine learning algorithm designed for iterative combinatorial auctions (ICAs). The core challenge in ICAs is the exponential growth of possible item bundles, making it impractical to ask bidders about every potential combination. While recent AI approaches used value queries (asking "What's this bundle worth to you?"), real-world auctions rely on demand queries (asking "Which bundle do you want at these prices?"). MLHCA is the first algorithm to provably integrate the full information from both query types, leading to vastly more efficient preference learning.

In practical experiments, this hybrid approach delivered staggering performance gains. MLHCA reduced the efficiency loss—the gap between the auction's outcome and the theoretically optimal allocation—by up to a factor of 10 compared to previous state-of-the-art methods. Crucially, it achieved these superior economic outcomes while asking up to 58% fewer questions. This dual benefit means auctioneers can achieve near-optimal results faster and with less burden on bidders, who no longer face an overwhelming questionnaire. The work, which received an Oral Presentation at ICML 2025, establishes a new benchmark for applying machine learning to complex market design problems where both computational efficiency and human usability are critical.