RA-ABC: Reverse Auction with Assumed Bid Cost

• RA-ABC is a reverse auction incentive mechanism for mobile crowd sensing that uses pre-sensing assumed bid costs to reduce wasteful effort before task execution.
• It employs Bayesian bid revisions and ROI-based participation tracking over multiple rounds to enhance efficiency, cost management, and fairness.
• The RA-ABCDR extension dynamically recruits new users, improving long-term user retention and lowering auction costs compared to conventional approaches.

Reverse Auction with Assumed Bid Cost (RA-ABC) is a reverse-auction–based incentive mechanism for Mobile Crowd Sensing (MCS) in which users bid before doing any sensing work, on the basis of an assumed cost, and only winners actually perform the task and get paid. It is designed to reduce wasted effort, manage auction cost efficiently, and sustain user participation over repeated auction rounds. The mechanism was introduced in "Incentive Mechanism for Mobile Crowd Sensing with Assumed Bid Cost Reverse Auction" (Yangchin et al., 10 Jul 2025), which also defines an extension with dynamic recruitment, RA-ABCDR.

1. Problem setting and motivation

In the MCS setting considered by RA-ABC, three roles are distinguished: task providers, mobile users, and a sensing platform. Task providers need sensing data such as air quality, traffic, or noise data; mobile users sell their sensing service using smartphones and onboard sensors; and the platform publishes tasks, runs the incentive mechanism, collects sensed data from winners, and pays users. What is traded is sensed data from the provider’s perspective and sensing service—effort, device energy, data, and time—from the user’s perspective (Yangchin et al., 10 Jul 2025).

The mechanism is motivated by limitations of conventional reverse-auction bidding in MCS. Standard reverse-auction mechanisms such as RADP, RADP-VPC, and RAIN typically assume that users already performed the task and therefore know their true cost $c_i$ when bidding. This entails a structural inefficiency: users who lose the auction have already expended energy, time, mobility cost, and possibly monetary travel cost. The paper identifies repeated waste of user resources, cumbersome long-horizon ROI tracking, and the common assumption of a fixed user pool as important drawbacks of that design.

RA-ABC reverses the ordering of sensing and bidding. Users estimate an assumed cost $\bar{c}_i$, submit bids before sensing, and incur actual cost only if selected as winners. The stated rationale is that performing tasks only after winning reduces resource consumption relative to performing tasks before bidding. The mechanism also places long-run participation at the center of the design by defining user Return on Investment (ROI) and using it to govern continued participation, dropout, and possible reentry. A further extension, RA-ABCDR, allows new users to join during bidding rounds rather than restricting the system to a fixed initial pool (Yangchin et al., 10 Jul 2025).

2. Formal model and bid construction

The mechanism operates over discrete rounds $r = 1,2,\dots,R$ with $N$ potential users indexed by $i \in \{1,\dots,N\}$. In each round, the platform selects a fixed number of winners $W$. The paper’s typical simulation uses $N=100$, $W=20$, and $R=100$.

For each user $i$ and round $\bar{c}_i$0, the model distinguishes among the true cost $\bar{c}_i$1, the assumed true cost $\bar{c}_i$2, and the bid $\bar{c}_i$3. The true cost is the actual cost of performing the sensing task and is known in practice only after the task is done. The assumed true cost is the user’s pre-sensing estimate of that cost. The bid is the requested payment for performing the task in round $\bar{c}_i$4, and in RA-ABC it is derived from assumed-cost information rather than realized cost.

The initial bid is defined in two equivalent forms in the presentation:

$\bar{c}_i$5

and

$\bar{c}_i$6

where $\bar{c}_i$7 or $\bar{c}_i$8 denotes the mean of historical bids or costs for similar tasks, $\bar{c}_i$9 or $r = 1,2,\dots,R$0 denotes the corresponding variance, and $r = 1,2,\dots,R$1 is a risk factor. Larger $r = 1,2,\dots,R$2 yields a more conservative bid.

During a bid revision window, the bid is refined through a Bayesian-style update:

$r = 1,2,\dots,R$3

where $r = 1,2,\dots,R$4 is a new observation such as current network state, battery status, or environment, $r = 1,2,\dots,R$5 is the variance of prior bid estimates, and $r = 1,2,\dots,R$6 is the variance of newly observed data. The update weights new evidence by the confidence ratio $r = 1,2,\dots,R$7.

To discourage aggressive manipulation between the initial and revised bid, the mechanism defines a deviation penalty

$r = 1,2,\dots,R$8

with $r = 1,2,\dots,R$9 as a sensitivity parameter. It also defines Bid Adjustment Impact (BAI),

$N$0

where $N$1 is the empirical winning probability obtained from historical success rates. The paper uses BAI as a system-efficiency metric and as a way to quantify how bid changes interact with winning probability (Yangchin et al., 10 Jul 2025).

Winner selection combines bid value and participation history. Participation in round $N$2 is represented by $N$3 if the user bids and $N$4 otherwise. Participation frequency is tracked by an Exponentially Weighted Moving Average (EWMA),

$N$5

with $N$6 and $N$7. The platform ranks users primarily by lower bids and, among similar bids, by higher participation frequency $N$8. The exact combined scoring rule is not provided explicitly, but the mechanism is described as favoring recent and regular participation when bid values are comparable.

The protocol in each round is: task publication; pre-sensing bidding based on assumed cost; winner selection; post-selection task execution and payment; and ROI update followed by the user’s participation decision for future rounds (Yangchin et al., 10 Jul 2025).

3. ROI, participation dynamics, and RA-ABCDR

RA-ABC uses ROI as the central state variable governing retention. Average earnings are tracked by another EWMA:

$N$9

where $i \in \{1,\dots,N\}$0 weights recent wins more heavily.

For an active user, ROI in round $i \in \{1,\dots,N\}$1 is defined as

$i \in \{1,\dots,N\}$2

where $i \in \{1,\dots,N\}$3 is average earnings, $i \in \{1,\dots,N\}$4 is participation frequency, $i \in \{1,\dots,N\}$5 is true per-task cost, and $i \in \{1,\dots,N\}$6 is a personal tolerance parameter expressing how much loss the user can tolerate. The platform sets a satisfaction threshold $i \in \{1,\dots,N\}$7. If $i \in \{1,\dots,N\}$8, the user becomes dropped, that is, inactive.

For a dropped participant considering reentry in the next round, the estimated ROI is

$i \in \{1,\dots,N\}$9

where $W$0 replaces $W$1 because the user has not performed the task in the current round. If $W$2, the user is inclined to rejoin. The paper states that a user’s initial ROI is set slightly above $W$3 when first joining so as to avoid immediate dropout after one loss.

This ROI machinery creates a repeated-game participation process rather than a one-shot procurement event. Users with sufficiently favorable average earnings relative to participation-weighted costs remain active; users whose ROI falls below the threshold leave; and dropped users may later return after re-estimating profitability. The paper does not specify a closed-form bid-update rule directly as a function of ROI, but it states that users with good ROI are more likely to continue participating and may slightly increase bids, whereas deteriorating ROI may induce bid reduction or exit.

RA-ABCDR extends the mechanism by introducing dynamic recruitment. In RA-ABC, the participant pool is fixed at initialization, and only active-versus-dropped status changes. In RA-ABCDR, new users may join at any round. A new entrant $W$4 is initialized with $W$5, $W$6, and an initial ROI slightly above $W$7, then computes an assumed cost $W$8 and an initial bid $W$9. From that point onward, the entrant follows the same bidding, selection, and ROI-update process as other active users. The paper explicitly links dynamic recruitment to improved stability, fairness, and cost-efficiency over long horizons (Yangchin et al., 10 Jul 2025).

4. Auction properties, fairness measures, and mechanism-theoretic interpretation

The paper proves a formal result for the adaptive bid cost function $N=100$0: it is non-negative, monotone (non-decreasing), and submodular. The argument is based on the conditions $N=100$1, $N=100$2, $N=100$3, the assumption that updates only increase bids when new observations suggest higher cost, and an aggregate diminishing-marginal-increase condition for combined bid functions. These properties are presented as supporting stable long-term auction behavior (Yangchin et al., 10 Jul 2025).

Auction cost per round is defined as

$N=100$4

and is used as the principal cost-efficiency measure. The paper states explicitly that there is no explicit platform budget constraint in the presented formulas, even though cost-efficiency is evaluated through $N=100$5 across rounds. This is an important distinction: reduced auction cost is not formal budget feasibility.

The mechanism evaluates fairness using two explicit indices. The Monopoly Prevention Index (MPI) is

$N=100$6

where $N=100$7 is the win frequency of user $N=100$8. Higher MPI indicates less monopolization and a more balanced distribution of wins. The Bid Accuracy Ratio (BAR) is

$N=100$9

which measures how closely bids align with actual costs. Lower BAR indicates more accurate bid–cost alignment.

The paper does not characterize RA-ABC as truthful in the classical auction-theoretic sense. Instead, it states that the mechanism is not explicitly designed as a truthful mechanism and encourages approximate cost-revealing bids through Bayesian cost refinement, penalties for large bid deviations, and BAR-based evaluation. It also treats individual rationality qualitatively: winners are paid their bid and only perform tasks after learning that they will be paid. A common misconception would be to equate this with full dominant-strategy truthfulness; the paper does not make that claim.

A useful contrast appears in later reverse-auction research on truthful adaptive procurement. "Truthful Reverse Auctions for Adaptive Selection via Contextual Multi-Armed Bandits" formalizes ex-post monotone allocation, critical-threshold payments, and truthfulness in expectation for reverse auctions with contextual learning, using virtual cost $W=20$0, a reverse self-resampling procedure (ROSA), and the generic transformation mechanism REV-GTM (Patra et al., 16 Feb 2026). This later framework is directly relevant because it supplies a mechanism-design vocabulary—EPIC, EPIR, monotonicity, critical payments, and virtual costs—that RA-ABC itself does not fully formalize. This suggests that RA-ABC is best understood as an incentive mechanism centered on pre-sensing cost estimation and participation sustainability rather than as a fully truthful procurement mechanism.

5. Empirical evaluation and reported results

The evaluation reported for RA-ABC and RA-ABCDR uses Python, specifically Spyder IDE in Anaconda, on AMD Ryzen 5 hardware with 8 GB RAM. The simulation parameters are an initial bidder population of $W=20$1, $W=20$2 winners per round, $W=20$3 rounds, assumed cost $W=20$4 generated from a Gaussian with mean $W=20$5 and in the range $W=20$6–$W=20$7 of actual cost $W=20$8, and ROI satisfaction threshold $W=20$9. The experiments are repeated over $R=100$0 different scenarios, and each plotted point is the average over $R=100$1 runs (Yangchin et al., 10 Jul 2025).

Three mechanisms are compared: RA-ABC, RA-ABCDR, and the Tullock Optimal Prize Function. The metrics are grouped as system efficiency, system stability, and system fairness. Efficiency uses ROI and BAI. Stability uses the number of active participants per round and auction cost. Fairness uses MPI and BAR.

The principal reported findings are as follows.

• Participant retention: RA-ABCDR consistently maintains the highest number of active participants across rounds. At round 100, it retains 54.6\% more users than Tullock.
• Auction cost: Tullock exhibits the highest auction costs across rounds; RA-ABC is moderate; RA-ABCDR has the lowest auction cost on average. RA-ABCDR reduces auction cost by about 22.2\% compared to Tullock.
• Effect of the satisfaction threshold $R=100$2: as $R=100$3 increases, fewer users remain active in both RA-ABC and RA-ABCDR. At the same $R=100$4, RA-ABCDR retains more users and has lower cost than RA-ABC.
• Efficiency trade-off: for a given user-utility level, RA-ABCDR achieves that level at lower auction cost than RA-ABC. The paper states that RA-ABCDR reaches near maximum utility of approximately $R=100$5 at a total auction cost around $R=100$6, whereas RA-ABC requires significantly higher cost to reach the same utility.
• Fairness: RA-ABC shows more stable BAR with fewer fluctuations, whereas RA-ABCDR has higher BAR volatility because dynamically arriving users may estimate costs less accurately. By contrast, RA-ABCDR exhibits more stable MPI over rounds, indicating less monopolization and more stable long-term fairness.

The paper therefore presents a trade-off rather than an unconditional dominance result. RA-ABCDR is strongest on long-term participation, auction cost, and MPI stability; RA-ABC is more stable on BAR; and Tullock is less suitable when repeated tasks and long-run sustainability are central design requirements (Yangchin et al., 10 Jul 2025).

6. Position within MCS incentive research and related directions

Within the broader MCS literature summarized in the paper, incentive mechanisms are grouped into auction-based approaches, including reverse auctions and dynamic pricing such as RADP, RADP-VPC, and RAIN, and contest-based approaches such as Tullock contests and lotteries. RA-ABC differs from these families in four explicitly stated ways: bid-before-acting rather than act-before-bid; assumed-cost estimation with adaptive updating; ROI-centered long-term participation management; and, in the RA-ABCDR extension, dynamic recruitment rather than a fixed user pool (Yangchin et al., 10 Jul 2025).

The most distinctive conceptual move is the separation of bidding from sensing execution. In conventional reverse auctions for MCS, sensing may occur before winner determination, so non-winning users bear real cost without compensation. In RA-ABC, only winners expend sensing effort. A plausible implication is that the mechanism is particularly relevant for settings with strong battery, data-plan, or mobility constraints, because it directly targets wasted effort rather than only payment efficiency.

The mechanism also adopts a deliberately longitudinal view of participation. ROI, participation frequency, and average earnings are all tracked through EWMA updates, so user state depends on repeated outcomes rather than only the current round. This differs from contest mechanisms, where losing users may repeatedly invest effort but receive nothing, and from one-shot auction analyses, which do not explicitly model dropout and reentry.

At the same time, RA-ABC leaves several classical mechanism-design issues open. The paper itself states that truthfulness is not formalized, that budget feasibility is not explicitly modeled, and that fairness is evaluated empirically through MPI and BAR rather than derived from dominant-strategy or Bayesian incentive constraints. The contextual reverse-auction framework of (Patra et al., 16 Feb 2026) addresses a different domain—adaptive selection of LLMs—but contributes formal tools that are germane to future RA-ABC-like designs: monotone reverse-auction allocation, threshold payments, ex-post monotonicity, resampling-based truthful transformation, and regret guarantees of order $R=100$7 for contextual learning. This suggests a possible research trajectory in which pre-sensing assumed-cost procurement, dynamic recruitment, and explicit truthfulness guarantees are unified within a single reverse-auction model.

RA-ABC and RA-ABCDR therefore occupy a specific position in the MCS incentive landscape. They are mechanisms for repeated procurement under uncertain pre-sensing costs, aimed at reducing wasted sensing effort and sustaining participation over many rounds. Their reported gains—especially the $R=100$8 improvement in user retention and the $R=100$9 reduction in auction cost for RA-ABCDR relative to Tullock—derive from that design orientation rather than from a classical truthful-auction formulation (Yangchin et al., 10 Jul 2025).