---
title: Multi-Constraint Bidding (MCB)
url: https://www.emergentmind.com/topics/multi-constraint-bidding-mcb
type: topic
---

# Multi-Constraint Bidding (MCB)

Searching arXiv for the cited MCB papers to ground the article and verify metadata.
{"query":"2508.03267 HALO Hindsight-Augmented Learning for Online Auto-Bidding", "max_results": 5}
{"query":"Multi-Constraint Bidding HALO auto-bidding 2508.03267", "max_results": 10}
Multi-Constraint Bidding (MCB) denotes bidding problems in which an allocation or bidding policy is optimized subject to multiple simultaneously active constraints. In the cited literature, MCB appears in several distinct auction environments: Real-Time Bidding (RTB) for digital advertising, where total budget and ROI targets are central; Simultaneous Ascending Auctions (SAA), where exposure, own price effect, budget constraints, and eligibility management jointly shape feasible strategies; and procurement auctions, where private unit-cost and private capacity define a bidimensional mechanism-design problem. Across these settings, the common structure is not a single auction rule but the coexistence of multiple operational or strategic limits that must be enforced while maximizing value, utility, conversions, or expected reward [2508.03267] [2307.11428] [1502.06934].

## 1. Scope and domain-specific meanings

The term MCB is used most explicitly in online advertising. In the RTB formulation of HALO, the advertiser seeks to maximize cumulative value subject to a total budget constraint $B$ and an ROI constraint $r_{\text{target}}$ [2508.03267]. In MCMF, the relevant constraints are framed as multiple Key Performance Indicators (KPIs), including eCPC, eCPM, per-conversion cost (PPC), and daily budget, with a unified optimization-and-budget-management objective rather than a separated pacing layer [2206.12147]. In the latent graph diffusion framework, MCB is formulated as constrained maximization over bidding trajectories under multiple KPI constraints such as budget, CPA, ROI, and CVR [2503.05805].

The SAA literature uses the same term differently. There, the MCB framework models spectrum auctions as an $n$-player simultaneous-move game with complete information, and the constraints are the exposure problem, the own price effect, budget constraints, and the eligibility management problem [2307.11428]. In procurement auctions, the bidimensional mechanism of 2D-OPT and 2D-UCB treats private per-unit cost and private capacity as the two operative dimensions, and the exposition explicitly presents this as a natural extension to a general MCB setting [1502.06934].

A common misconception is to treat MCB as a single fixed optimization template. The cited works instead use the term for several related formulations whose shared feature is simultaneous constraint handling, but whose state spaces, action spaces, observability assumptions, and objectives differ materially.

| Setting | Constraints or dimensions | Representative formulation |
|---|---|---|
| RTB digital advertising | total budget, ROI, PPC, CPA, CVR, daily budget | HALO, MCMF, latent graph diffusion |
| Spectrum SAA | exposure, own price effect, budget constraints, eligibility management | $SMS^\alpha$ |
| Procurement auction | unit-cost, capacity, unknown quality | 2D-OPT, 2D-UCB |

## 2. Canonical RTB formulation

A canonical MCB problem in RTB is defined over an impression sequence $\mathcal I=\{1,\dots,N\}$ within one bidding period. For each impression $i$, the advertiser has a private value $v_i\ge 0$, a market-clearing price $c_i>0$ that is unknown until after bidding, and an action $x_i\in\{0,1\}$ indicating whether the impression is won. Under a second-price auction, the advertiser pays $c_i$ if $x_i=1$ [2508.03267].

The optimization target is

$$
\max \sum_{i\in\mathcal I} v_i x_i
$$

subject to

$$
\sum_{i\in\mathcal I} c_i x_i \le B,
\qquad
\frac{\sum_{i\in\mathcal I} v_i x_i}{\sum_{i\in\mathcal I} c_i x_i} \ge r_{\text{target}}.
$$

The special case BCB drops the ROI constraint [2508.03267]. Prior closed-form analyses in the static one-shot setting yield

$$
b_i=\beta v_i
$$

for BCB, with $\beta$ chosen so that $\sum c_i x_i(\beta)=B$, and

$$
b_i=(\beta_0-\beta_1 r_{\text{target}})v_i
$$

for MCB. The stated limitation is that these formulas break down when constraints shift or when one must adapt online over billions of impressions [2508.03267].

The MCMF literature generalizes the RTB control perspective by writing the actual bid in eCPM as

$$
\text{eCPM}
=
1000\times pCTR\times pCVR\times PPC_e\times u^{(t)},
$$

where $u^{(t)}$ is a bid adjustment factor decided at time $t$ [2206.12147]. The latent graph diffusion formulation uses a trajectory view: for time horizon $t=1,\dots,T$, agent $i$ chooses bid vectors $b_t$, and the objective is to maximize expected total reward under example constraints

$$
\sum_{t=1}^T \mathbb E[c_t]\le B_{\max},
\qquad
\frac{\sum_t \mathbb E[c_t]}{\sum_t \mathbb E[r_t]}\le \text{CPA}_{\max},
$$

$$
\frac{\sum_t \mathbb E[r_t]}{\sum_t \mathbb E[c_t]}\ge \text{ROI}_{\min},
\qquad
\frac{\sum_t \mathbb E[w_t]}{\sum_t |\mathcal K_t|}\ge \text{CVR}_{\min}.
$$

This RTB line of work therefore spans both single-period constrained optimization and sequential planning over bidding trajectories [2503.05805].

## 3. Online learning and control in advertising MCB

HALO addresses two explicit deficiencies of traditional auto-bidding: severe sample inefficiency and limited generalization under constraint shifts. The paper characterizes the former as “near-zero transfer” across constraint configurations, because most methods perform trial-and-error per $(B,r)$ pair and discard trajectories that overshoot budget or miss ROI. It characterizes the latter as failure to capture the underlying monotonic or derivative relationships in mappings such as $(B\to\beta)$ or $(r\to\beta)$; empirically, about one-third of campaigns adjust constraints mid-flight, causing large KPI volatility in prior methods [2508.03267].

Its core mechanism is hindsight experience with trajectory reorientation. HALO adopts a Fixed Coefficient Strategy (FCS): at decision step $\tau$, it chooses a constant bid multiplier $\hat\beta$ for all remaining slots $t=\tau,\dots,T$ and records realized total cost and value $\hat C_\tau^T(\hat\beta)$ and $\hat V_\tau^T(\hat\beta)$. Each tuple

$$
H_\tau=(s_\tau,\hat\beta,\hat C_\tau^T(\hat\beta),\hat V_\tau^T(\hat\beta))
$$

can be re-labeled as optimal data for budget $B'=\hat C_\tau^T(\hat\beta)$ [2508.03267]. For ROI handling, HALO uses the stated monotonicity of $ROI(\beta)\equiv E[\sum v/\sum c]$ and solves for a shading factor $\alpha<1$ through the integral equation

$$
\beta^* \int_{\alpha}^{1}\bigl[\gamma(\beta^*u)-r_{\text{target}}\eta(\beta^*u)\bigr]\,du
=
\Delta_{ROI},
$$

with $\gamma(\beta)=\partial V/\partial\beta$, $\eta(\beta)=\partial C/\partial\beta$, and
$\Delta_{ROI}=V_1^{\tau-1}+V_\tau^T(\beta^*)-r_{\text{target}}(C_1^{\tau-1}+C_\tau^T(\beta^*))$ [2508.03267].

HALO couples this hindsight reuse with a B-spline functional representation. It learns two continuous functions of remaining budget $C$,

$$
\beta(C)\approx f_\theta(C),
\qquad
\hat V(C)\approx f_\phi(C),
$$

where each $f$ is represented as a degree-$k$ B-spline with control points:

$$
f(C)=\sum_{i=0}^{n} N_{i,k}(C)\,\theta_i.
$$

The stated advantages are local control, guaranteed smoothness up to order $k-1$, easy derivative $\partial f/\partial C$ for computing $\alpha$, and more robust extrapolation than pointwise fits or black-box MLPs [2508.03267]. A small MetaModel outputs the control points $\theta,\phi$ from state features such as time left, pacing, and p-value histograms, and the system is trained end-to-end with multi-anchor supervision from hindsight labels.

MCMF addresses a different triad of difficulties: sparse feedback, budget management separated from the optimization, and absence of bidding environment modeling. Its input is a “merging” vector that concatenates, for each KPI $i$, the target value $z_i$, the accumulated feedback $x_i$, and additional accumulated features $\mathbf v$, including posterior features such as realized CTR and realized CVR and prior estimates such as average pCTR and average pCVR [2206.12147]. The model is a two-layer MLP with

$$
\mathbf h=\mathbf W_e \mathbf x,
\qquad
\mathbf U=\sigma(\mathbf W_d \mathbf h).
$$

Its cost function integrates KPI tracking error and control effort over a sliding window:

$$
J=
\int_t^{t+\tau}
\mathbf E^T \mathbf Q \mathbf E
+
\Delta \mathbf U^T \mathbf R \Delta \mathbf U
\,dt.
$$

The unknown auction-response derivative $dx/du$ is approximated by a Hebbian-rule sign term,

$$
\frac{dx}{du}
\approx
\mathrm{sign}\bigl((x^{(t)}-x^{(t-1)})(u^{(t)}-u^{(t-1)})\bigr),
$$

so the updates do not require an explicit auction model [2206.12147]. The paper explicitly states that MCMF never separates utility maximization from pacing, because budget and PPC constraints are folded directly into the same cost.

The latent graph diffusion framework introduces a planning-based latent diffusion model for large-scale auction environments. Each agent constructs a graph whose nodes encode impression opportunities and a special non-exposed node, with edges capturing intra-agent competition and additional cross-agent sharing. An attention-based GNN encoder maps the graph to a latent state, an inverse-dynamics model predicts bid vectors, and self-attention aggregation yields an approximate multi-agent equilibrium representation [2503.05805]. Planning is performed in latent space via a standard Gaussian diffusion process with forward noising

$$
q(x_t\mid x_{t-1})
=
\mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\,\beta_t I\bigr),
$$

and reward alignment fine-tunes the posterior with

$$
\mathcal L_{\mathrm{align}}
=
\mathcal L_{\mathrm{diff}}
-
\lambda\,\mathbb E_{x\sim p_\theta}[V_\phi(x)].
$$

The explicit aim is to generate auto-bidding trajectories that maximize KPI metrics while satisfying constraint thresholds [2503.05805].

## 4. Auction-theoretic and mechanism-design variants

In spectrum auctions, the MCB framework of Pacaud, Bechler, and Coupé’choux models SAA as a finite extensive-form game in which all bidders simultaneously choose subsets of remaining items, subject to hard budget and eligibility constraints [2307.11428]. If bidder $i$ temporarily holds items $Y_i$ and has eligibility $e_i$, its action set must satisfy

$$
|X_i|+|Y_i|\le e_i
$$

and

$$
\sum_{j\in X_i\cup Y_i}\bigl(P_j+\varepsilon\cdot\mathbf 1_{j\notin Y_i}\bigr)\le b_i.
$$

Payoff at auction close is

$$
\sigma_i(W_i,P^f)=v_i(W_i)-\sum_{j\in W_i} P_j^f.
$$

The algorithm $SMS^\alpha$ is a Simultaneous-Move Monte Carlo Tree Search (SM-MCTS) procedure with a risk-averse reward

$$
R^\pi_\alpha
=
\bigl(1+\alpha\,\mathbf 1_{R^\pi<0}\bigr)R^\pi,
$$

a learned closing-price forecast, UCB-style selection, legal-action expansion, point-price-prediction rollouts, and back-propagation of risk-averse returns [2307.11428]. The stated purpose is simultaneous mitigation of the exposure problem, own price effect, budget exhaustion, and eligibility management.

The procurement literature frames a different MCB problem: a buyer procures $L$ identical units from heterogeneous strategic agents with private cost and capacity and unknown quality [1502.06934]. In the offline version, with known quality, the mechanism 2D-OPT uses an allocation rule and payments that maximize the auctioneer’s expected utility under Bayesian Incentive Compatibility (BIC) and Individual Rationality (IR). The key quantity is the virtual cost

$$
H_i(c_i,k_i)=c_i+\frac{F_i(c_i\mid k_i)}{f_i(c_i\mid k_i)},
$$

and the allocation rule computes

$$
G_i=R q_i - H_i(\hat c_i,\hat k_i),
$$

sorts agents in decreasing $G_i$, and greedily takes as many units as allowed until $L$ units are chosen or $G_i$ becomes nonpositive [1502.06934]. The payment rule is

$$
t_i=c_i x_i+\int_{c_i}^{\bar c_i} x_i(z,k_i)\,dz.
$$

In the online version, 2D-UCB learns unknown qualities through UCB estimates

$$
q_i^+(t)=\hat q_i(t)+\sqrt{\frac{2\ln t}{n_i(t)}},
$$

combined with self-resampling so that the induced mechanism remains Stochastic-BIC and IR [1502.06934].

These auction-theoretic variants show that MCB is not restricted to ad-delivery platforms. The same label is used for simultaneous-move strategic games, constrained search over legal actions, and truthful multidimensional mechanisms with online learning.

## 5. Structural guarantees and optimality claims

HALO’s analysis is built around the fixed-coefficient paradigm. Lemma 1 gives an FCS error bound: any fixed-coefficient strategy that exactly exhausts budget is within $v_{\max}$ of the omniscient optimum $V^{oracle}$. Lemma 2 states the BCB optimality condition: if $\sum_{t=\tau}^T C_t(\beta^*)=B_\tau$, then $\beta^*$ is optimal on $[\tau,T]$. Lemma 3 states ROI monotonicity: $R(\beta)=E[V/C]$ is non-increasing in $\beta$, which guarantees a unique shading factor $\alpha$ for ROI compliance. Lemma 4 and Theorem 1 provide the closed-form integral equation for $\alpha$ and prove the global optimality of the two-stage rule that deploys $\beta$ if ROI is nonnegative and otherwise shades by $\alpha$ [2508.03267]. The paper’s interpretation is that every rollout can be turned into correct training data for any budget/ROI pair.

The procurement mechanism relies on envelope conditions. A mechanism is BIC and IR if, for each agent, interim allocation is non-increasing in reported cost, interim offered utility is non-decreasing in reported capacity and nonnegative, and the offered utility satisfies the integral identity

$$
\rho_i(\hat c_i,\hat k_i)
=
\rho_i(\bar c_i,\hat k_i)+\int_{\hat c_i}^{\bar c_i} X_i(z,\hat k_i)\,dz.
$$

Under a regularity assumption that $H_i$ is non-decreasing in cost and non-increasing in capacity, the resulting mechanism is DSIC, IR, and utility-maximizing for the auctioneer [1502.06934].

The SAA framework does not present an analogous closed-form optimality theorem in the provided summary, but it does encode feasibility directly in the action sets and embeds risk-aversion into the reward function. Its empirical-game analysis states that every unilateral deviation to $SMS^\alpha$ is strictly profitable against any profile of rivals using the listed baselines, making $(SMS^\alpha,\dots,SMS^\alpha)$ a clear Nash equilibrium [2307.11428].

A plausible implication is that MCB research splits into two broad proof styles. One style proves monotonicity, envelope identities, or near-optimality under simplified policy classes. The other enforces constraints algorithmically through action-space restriction, trajectory re-labeling, or reward alignment and evaluates performance empirically.

## 6. Empirical evidence, recurring difficulties, and research directions

The HALO evaluation uses AuctionNet with 21 days, approximately 500K impression events per day, and 48 advertisers per day, split into 14 days train and 7 days test. In MCB mode, the reported metrics are average conversions when the ROI constraint is satisfied (“Conv”) and compliance rate (C.R.); in BCB mode, they are conversions and cost-to-budget ratio (C/B). Across budgets from 50% to 150% of a normalized baseline, HALO attains the highest Conv in MCB while keeping $\text{C.R.}\ge 0.92$ in all budgets, and in BCB it reaches $\text{C/B}\approx 99\%$ with the highest Conv in every scenario. The reported final realized ROI/target is $1.0204$ for HALO versus $0.9265$ for BC, $0.7657$ for IQL, and $0.7632$ for LP [2508.03267].

MCMF is evaluated on the iPinYou open dataset and in Alibaba RTB production. On iPinYou, under Single/Adequate conditions, MCMF achieves 33 conv at PPC approximately 786 FEN versus RL’s 33 conv at PPC approximately 1519 FEN; under Multi/Tight, it achieves 26 conversions with PPC approximately 707 FEN versus RL’s 3 conversions at PPC approximately 1724 FEN. Across all four setups, the paper reports that MCMF maximizes conversions, never exceeds the PPC target, and spends exactly the allotted budget. In the reported online A/B test with 7 K SKUs, 22 K advertisers, 4 M bids per day, and $10^{10}$ requests per day, over 15 days MCMF produced $+2.96\%$ conversions, $+5.55\%$ ROI, and $-2.46\%$ cost [2206.12147].

In SAA experiments on realistic instances with $n\in\{2,3,4\}$ and $m\in\{7,9,11\}$, when all four bidders play $SMS^\alpha$, the average per-bidder utility is roughly $21.3$, versus $10.8$ under EPE and negative under SB. Exposure frequency falls below $0.5\%$ for $SMS^\alpha$, compared with $30$–$60\%$ for the baselines, and expected losses are $10$–$90\times$ lower. Average price paid per licence is $1.5$–$3\times$ lower than under SB or SCPD and $10$–$15\%$ lower than under EPE when facing the same opponents [2307.11428].

The latent graph diffusion framework reports results on AuctionNet and a synthetic auction dataset. On AuctionNet, against DiffBid, the reported metrics are Return $489.04$ versus $353.19$, CPA $0.992$ versus $1.041$, ROI $0.030$ versus $0.012$, CVR $0.152$ versus $0.071$, budget adherence approximately $0.995$ for both, and social welfare $4.09$ versus $3.63$. On the synthetic dataset, the corresponding results are Return $2123.10$ versus $982.59$, CPA $0.592$ versus $0.842$, ROI $0.055$ versus $0.026$, CVR $0.203$ versus $0.089$, budget adherence approximately $1.0$, and social welfare $69.20$ versus $50.43$ [2503.05805].

Several recurring technical difficulties emerge directly from these studies. The RTB literature emphasizes sparse and delayed feedback, severe sample inefficiency, mid-flight constraint changes, and catastrophic extrapolation outside the training range [2508.03267] [2206.12147]. MCMF explicitly notes that the choice of $\mathbf Q$, $\mathbf R$, $\tau$, and $\eta$ requires careful tuning, and that the sign-based gradient approximation can be noisy under extremely erratic auction environments [2206.12147]. The procurement formulation identifies multidimensional ironing, multi-dimensional critical-value integrals, and regret control when multiple dimensions affect ranking gaps as key technical challenges for general MCB extensions [1502.06934]. Taken together, these results indicate that the central research problem in MCB is not merely constrained optimization, but robust adaptation under shifting constraints, partial feedback, strategic interaction, and domain-specific feasibility rules.

Source: https://www.emergentmind.com/topics/multi-constraint-bidding-mcb