OCPC Advertising Bidding Strategy: Problem Modeling, Linear Programming Solution, and PID Control

The document introduces the OCPC advertising product, explaining that advertisers pay media to deliver marketing messages to target users, and outlines the three main entities—advertiser, media, and user—along with their relationships.

It describes the system architecture of the ZZADECPM module that implements the bidding strategy, and provides the business background highlighting the need for budget‑constrained and multi‑constrained bidding solutions.

The project goals are to increase ROI without reducing overall platform revenue and to improve effective fill rates, ASN, revenue, and ROI through A/B testing, with responsibilities divided among product, sales, operations, and technical teams.

Challenges include meeting advertisers' cost and budget constraints, handling cold‑start incentives, aligning algorithmic logic with product design, and managing multiple constrained optimization objectives.

In the solution section, the problem is modeled as a constrained linear programming task (LP1) aiming to maximize GMV under budget and CPC limits, with its dual (LP2) derived via complementary slackness. The optimal bidding formula is presented and analyzed.

To solve the LP, the authors first use the open‑source GLPK/pyomo stack for small‑scale instances, then adopt the commercial Gurobi solver for larger problems, providing installation screenshots and noting licensing considerations.

Parameter analysis examines how the optimal parameters p and q affect budget and CPC, illustrating trade‑offs with plots and discussing reinforcement‑learning and PID feedback control approaches for parameter tuning.

The PID control system is detailed with diagrams and standard PID equations, showing how it regulates p and q in real time.

Offline workflow involves training data and Gurobi to obtain optimal p and q, followed by grid‑search for PID coefficients; online deployment initializes parameters from offline results, collects bidding logs hourly, and adjusts p and q via PID based on budget and cost errors.

The conclusion summarizes the research on OCPC strategies, the collaboration across product, R&D, operations, and sales, and reports successful launch with measurable improvements in revenue, ASN, and ROI.

References to relevant academic papers and tools (e.g., Pyomo, Gurobi, reinforcement learning for budget‑constrained bidding) are listed at the end.