Research map / Learning theory and algorithms / Submodular and game algorithms
Prices, Bids, Values: One ML-Powered Combinatorial Auction to Rule Them All
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MLHCA, an ML-powered hybrid iterative combinatorial auction that combines demand queries (CCA-style clock rounds with ML-generated DQs) and value queries with a bridge bid, and a mixed training loss over both query types.
Problem
Iterative combinatorial auctions must balance efficiency and convergence speed under limited bidder communication; prior ML-based ICAs use either value queries (BOCA) or demand queries (ML-CCA), each with efficiency limitations.
Approach
Combine demand and value queries in one ICA: start with ML-powered DQs in the CCA clock-style interaction, then switch to ML-powered VQs after a bridge bid, and train each bidder's mMVNN on both query types with a two-stage loss that provably uses all DQ information.
Key results
- MLHCA (40 DQs + 60 VQs) reaches 0.04% efficiency loss in LSVM versus 0.39% for BOCA with 100 VQs, and 0.00% in GSVM and SRVM.
- In MRVM, MLHCA has 4.81% efficiency loss versus 6.68% to 7.77% for the other mechanisms, and needs 54 queries to statistically outperform BOCA.
- Mixed DQ+VQ training gives nearly three times lower MAE than DQ-only training on utility-maximizing bundles in the test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MLHCA with 40 DQs followed by 60 VQs instead of BOCA | MLHCA vs BOCA · LSVM | efficiency loss (%): 0.39 ± 0.31 -> 0.04 ± 0.07 | Helped |
| Use MLHCA with 40 DQs followed by 60 VQs instead of other tested mechanisms | MLHCA · GSVM | efficiency loss (%): 0.00 ± 0.00; ML-CCA_clock: 1.77 ± 0.68; ML-CCA_raised: 1.07 ± 0.37; CCA: 9.60 ± 1.49 | Helped |
| Use MLHCA with 40 DQs followed by 60 VQs instead of other tested mechanisms | MLHCA · SRVM | efficiency loss (%): 0.00 ± 0.00; BOCA: 0.06 ± 0.02; ML-CCA_clock: 0.41 ± 0.11; ML-CCA_raised: 0.07 ± 0.02; CCA: 0.37 ± 0.11 | Helped |
| Use MLHCA with 40 DQs followed by 60 VQs instead of other tested mechanisms | MLHCA · MRVM | efficiency loss (%): 4.81 ± 0.57 vs 7.77 ± 0.35 | Helped |
| Combine DQ and VQ training rather than train on DQs alone (ablation) | MVNN · Price-driven test set (T_p) | R²: -10.07 -> 0.42 | Helped |
| Combine DQ and VQ training rather than train on VQs alone (ablation) | MVNN · Price-driven test set (T_p) | R²: -3.07 -> 0.42 | Helped |
| Use DQ-only training rather than mixed DQ/VQ training (ablation) | MVNN · Price-driven test set (T_p) | R²: 0.42 -> -10.07 | Hurt |
| Use VQ-only training rather than mixed DQ/VQ training (ablation) | MVNN · Price-driven test set (T_p) | R²: 0.42 -> -3.07 | Hurt |
| Remove the bridge bid when transitioning from DQs to VQs (ablation) | MLHCA · MRVM | efficiency: 7.3% points | Hurt |
| Add the bridge bid when transitioning from DQs to VQs (ablation) | MLHCA · MRVM | efficiency: 7.3% points | Helped |
| Reverse query order: use VQs before DQs (ablation) | MLHCA · All tested domains | efficiency: up to 5 percentage points | Hurt |
| Start the hybrid auction with DQs instead of VQs (ablation) | Hybrid auction · Initial queries | efficiency: up to 20% points | Helped |
Surprising findings
- Introducing a single VQ without a bridge bid can drop efficiency by up to nearly 100% in DQ-based auctions; the bridge bid prevents this.
- DQ-only trained models cannot identify value functions uniquely up to a constant shift, even with all possible DQs.
Limitations
- Switching point from DQs to VQs is fixed by hand (40 DQs) rather than chosen dynamically.
- VQ phase is not strategyproof; incentive properties rely on additional assumptions.
- Evaluation uses synthetic SATS instances rather than real bidder data.
Experimental setup
Datasets: SATS spectrum auction test suite (GSVM, LSVM, SRVM, MRVM domains)
Benchmarks: SATS GSVM SATS LSVM SATS SRVM SATS MRVM
Metrics: efficiency loss number of queries to statistical significance R^2 Kendall Tau scaled MAE
Baselines: BOCA ML-CCA CCA ML-CCA with clock bids raised
Builds on: ML-CCA BOCA MLCA combinatorial clock auction multiset monotone-value neural networks
Closest papers by meaning
- Machine Learning-Powered Combinatorial Clock Auction (AAAI 2024 · similarity 0.95)
- Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition (AAAI 2025 · similarity 0.89)
- BundleFlow: Deep Menus for Combinatorial Auctions by Diffusion-Based Optimization (NeurIPS 2025 · similarity 0.86)
- Learning Optimal Auctions with Correlated Value Distributions (AAAI 2025 · similarity 0.85)
- AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games (NeurIPS 2024 · similarity 0.85)
- Learning Bayesian Nash Equilibrium in Auction Games via Approximate Best Response (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.