Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Ermis Soumalias, Jakob Heiss, Jakob Weissteiner, Sven Seuken

TL;DR

The paper tackles the challenge of efficiently allocating large combinatorial bundles in ICAs by learning bidder preferences through iterative queries. It introduces MLHCA, a hybrid ML-powered auction that combines demand and value queries with a mixed-training learning algorithm, augmented by a bridge bid to preserve efficiency when transitioning between query types. The authors provide a theoretical framework showing when DQs or VQs are most effective and demonstrate empirically that MLHCA achieves up to 10x reductions in efficiency loss and up to 58% fewer queries across SATS domains, while reducing bidders' cognitive load. The work highlights practical applicability to real-world spectrum-like settings by offering strong efficiency guarantees, compatibility with standard payment rules, and mechanisms to detect misreports, marking a significant advance in practical, high-widelity ML-driven ICAs.

Abstract

We study the design of iterative combinatorial auctions (ICAs). The main challenge in this domain is that the bundle space grows exponentially in the number of items. To address this, recent work has proposed machine learning (ML)-based preference elicitation algorithms that aim to elicit only the most critical information from bidders to maximize efficiency. However, while the SOTA ML-based algorithms elicit bidders' preferences via value queries, ICAs that are used in practice elicit information via demand queries. In this paper, we introduce a novel ML algorithm that provably makes use of the full information from both value and demand queries, and we show via experiments that combining both query types results in significantly better learning performance in practice. Building on these insights, we present MLHCA, a new ML-powered auction that uses value and demand queries. MLHCA substantially outperforms the previous SOTA, reducing efficiency loss by up to a factor 10, with up to 58% fewer queries. Thus, MLHCA achieves large efficiency improvements while also reducing bidders' cognitive load, establishing a new benchmark for both practicability and efficiency.

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

TL;DR

The paper tackles the challenge of efficiently allocating large combinatorial bundles in ICAs by learning bidder preferences through iterative queries. It introduces MLHCA, a hybrid ML-powered auction that combines demand and value queries with a mixed-training learning algorithm, augmented by a bridge bid to preserve efficiency when transitioning between query types. The authors provide a theoretical framework showing when DQs or VQs are most effective and demonstrate empirically that MLHCA achieves up to 10x reductions in efficiency loss and up to 58% fewer queries across SATS domains, while reducing bidders' cognitive load. The work highlights practical applicability to real-world spectrum-like settings by offering strong efficiency guarantees, compatibility with standard payment rules, and mechanisms to detect misreports, marking a significant advance in practical, high-widelity ML-driven ICAs.

Abstract

We study the design of iterative combinatorial auctions (ICAs). The main challenge in this domain is that the bundle space grows exponentially in the number of items. To address this, recent work has proposed machine learning (ML)-based preference elicitation algorithms that aim to elicit only the most critical information from bidders to maximize efficiency. However, while the SOTA ML-based algorithms elicit bidders' preferences via value queries, ICAs that are used in practice elicit information via demand queries. In this paper, we introduce a novel ML algorithm that provably makes use of the full information from both value and demand queries, and we show via experiments that combining both query types results in significantly better learning performance in practice. Building on these insights, we present MLHCA, a new ML-powered auction that uses value and demand queries. MLHCA substantially outperforms the previous SOTA, reducing efficiency loss by up to a factor 10, with up to 58% fewer queries. Thus, MLHCA achieves large efficiency improvements while also reducing bidders' cognitive load, establishing a new benchmark for both practicability and efficiency.

Paper Structure

This paper contains 63 sections, 14 theorems, 20 equations, 8 figures, 11 tables.

Table of Contents

  1. Introduction
  2. ML-Powered Iterative Combinatorial Auctions
  3. Our Contributions
  4. Further Related work
  5. Practical Considerations and Incentives
  6. Preliminaries
  7. Formal Model for ICAs
  8. Benchmark ICAs
  9. CCA
  10. ML-CCA
  11. BOCA
  12. ML Framework
  13. A Theoretical Framework for Effectively Combining DQs and VQs
  14. VQ-based approaches rely on cognitively complex random VQs.
  15. DQs Offer Superior Efficiency Gains in Initial Rounds.
  16. ...and 48 more sections

Key Result

Proposition 3.1

The expected social welfare of an auction that uses a single random demand query can be arbitrarily larger than that of an auction that uses any constant number ($k \ll 2^m$) of random value queries.

Figures (8)

  • Figure 1: Efficiency loss paths (i.e., regret plots) of MLHCA compared to BOCA, ML-CCA and CCA. Shown are averages over 50 instances with 95% CIs.
  • Figure 2: Efficiency of MLHCA with and without the bridge bid (\ref{['def:bridge_bid']}) in the MRVM domain.
  • Figure 3: NextQueries$(I,R)$(Brero et al. 2021)
  • Figure 4: Mlca($Q^{\textrm{\tiny init}},Q^{\textrm{\tiny max}},Q^{\textrm{\tiny round}}$) (Brero et al. 2021)
  • Figure 5: MixedTraining
  • ...and 3 more figures

Theorems & Definitions (48)

  • Definition 2.1: Demand Query
  • Definition 2.2: Value Query
  • Definition 2.3: Inferred Value
  • Proposition 3.1
  • Theorem 3.2
  • Proposition 3.3
  • Definition 1.1: Indirect Utility and Revenue
  • Definition 1.2: Clearing Prices
  • Theorem 1.3: Soumalias2024MLCCA
  • Theorem 1.4: Soumalias2024MLCCA
  • ...and 38 more