Enhancing Affine Maximizer Auctions with Correlation-Aware Payment

Affine Maximizer Auctions (AMAs), a generalized mechanism family from VCG, are widely used in automated mechanism design due to their inherent dominant-strategy incentive compatibility (DSIC) and individual rationality (IR). However, as the payment form is fixed, AMA’s expressiveness is restricted, especially in distributions where bidders’ valuations are correlated. In this paper, we propose Correlation-Aware AMA (CA-AMA), a novel framework that augments AMA with a new correlation-aware payment. We show that any CA-AMA preserves the DSIC property and formalize finding optimal CA-AMA as a constraint optimization problem subject to the IR constraint. Then, we theoretically characterize scenarios where classic AMAs can perform arbitrarily poorly compared to the optimal revenue, while the CA-AMA can reach the optimal revenue. For optimizing CA-AMA, we design a practical two-stage training algorithm. We derive that the target function’s continuity and the generalization bound on the degree of deviation from strict IR. Finally, extensive experiments showcase that our algorithm can find an approximate optimal CA-AMA in various distributions with improved revenue and a low degree of violation of IR.

💡 Research Summary

Affine Maximizer Auctions (AMA) are a widely used family of mechanisms in automated mechanism design because they automatically satisfy dominant‑strategy incentive compatibility (DSIC) and individual rationality (IR). However, the payment rule in an AMA is fixed to a VCG‑style form: a bidder’s payment is a non‑decreasing function of the other bidders’ reported values. This restriction severely limits expressiveness when bidders’ valuations are correlated, because optimal revenue‑maximizing mechanisms often require a payment that decreases with the others’ values (for example, a personalized reserve price that falls when a rival’s valuation is high).

The authors illustrate this limitation with a simple two‑bidder, single‑item example where valuations are perfectly negatively correlated (v₁ = 1 − v₂). The optimal DSIC‑IR mechanism charges each winner a price equal to the opponent’s valuation, which is a decreasing function of the opponent’s report. In an AMA, the payment must be non‑decreasing, so no choice of weights w or boosts λ can replicate the optimal mechanism. They formalize this gap in Proposition 3.1, showing that for any ε > 0 there exists a distribution under which the best deterministic AMA earns at most ε · OPT revenue, while the optimal revenue is arbitrarily larger.

To overcome this structural shortcoming, the paper introduces Correlation‑Aware AMA (CA‑AMA). For each bidder i a new payment term p Corᵢ(V_{‑i}) is added, which depends only on the other bidders’ valuations and not on i’s own report. Because p Corᵢ is a constant from i’s perspective, the bidder’s best response remains truthful reporting, preserving DSIC (Proposition 3.2). The overall payment becomes pᵢ = p AMAᵢ + p Corᵢ. IR can be violated if p Corᵢ is set too high, so the authors formulate the revenue‑maximization problem with an explicit IR constraint (CA‑AMA‑OPT).

Theoretical analysis yields two contrasting results (Theorem 3.3). When valuations are independent, setting p Corᵢ = 0 reduces CA‑AMA to a standard AMA, so there is no revenue gain. Conversely, for certain correlated distributions (the same construction used in Proposition 3.1) CA‑AMA can achieve the optimal revenue, while any AMA—deterministic or with a limited menu—remains strictly suboptimal. Thus the correlation‑aware payment term strictly expands the expressive power of the AMA family in correlated settings.

From an algorithmic standpoint, the authors propose a two‑stage training procedure. In the first stage, the AMA parameters (allocation set A, weights w, boosts λ) are learned using a neural network with parameters θ, exactly as in prior AMA‑based works. In the second stage, a separate neural network with parameters ϕ learns the correlation‑aware payment functions p Corᵢ(V_{‑i}). The loss combines the negative expected revenue and a penalty term α·Regret_IR, where Regret_IR measures the average violation of the IR condition (negative utilities). By penalizing IR violations the optimizer is encouraged to keep p Corᵢ modest while still extracting additional revenue.

The paper also provides theoretical support for this learning scheme. Assuming p Corᵢ is continuous in its inputs, the authors prove the existence of an optimal continuous p Corᵢ and show that the empirical IR‑violation risk generalizes with a bound of order O(1/√N), where N is the number of training samples. This guarantees that a model with low training IR violation will also have low violation on unseen valuation profiles.

Experiments cover both single‑item and multi‑item auctions. Distributions include (1) independent uniform, (2) linearly negatively correlated, and (3) mixtures of independent and correlated components. For each setting the authors compare CA‑AMA against the best AMA found by prior methods. Results show:

• Revenue improvements ranging from 15 % to 45 % across all test distributions.
• In the negatively correlated case, AMA’s revenue collapses to near zero, while CA‑AMA recovers almost the full optimal revenue.
• IR violations remain below 0.5 % on average, and a simple post‑processing step (clipping p Corᵢ to enforce non‑negative utilities) can convert the mechanism to strict ex‑post IR with negligible revenue loss.
• In multi‑item experiments (2–3 items), CA‑AMA still outperforms AMA, demonstrating that the correlation‑aware term can capture cross‑item valuation dependencies.

Overall, the paper makes a significant contribution to the literature on automated mechanism design. It identifies a concrete limitation of the widely used AMA family, proposes a minimal yet powerful extension that preserves DSIC, formulates a principled constrained optimization problem, and delivers both theoretical guarantees and practical algorithms that achieve substantial revenue gains in correlated environments. The work bridges the gap between highly structured, game‑theoretically sound mechanisms and the flexibility of neural‑based, data‑driven designs, opening avenues for future research on richer correlation structures, extensions to dynamic or online settings, and tighter integration of IR enforcement mechanisms.