Generative Auto-Bidding Framework with Value-Guided Explorations

This technical report introduces the Generative Auto-bidding framework with Value-Guided Explorations (GAVE), designed to optimize bidding strategies in dynamic online advertising environments. The report outlines the limitations of existing auto-bidding strategies, which often rely on rule-based or Reinforcement Learning (RL) methods that struggle to adapt to changing market conditions. GAVE addresses these issues by implementing a score-based Return-To-Go (RTG) module for accommodating various advertising objectives and an action exploration mechanism to ensure stability in updates. These innovations enable the framework to explore new actions while maintaining performance across diverse scenarios. Additionally, the report presents experimental results demonstrating GAVE's effectiveness based on evaluations from two offline datasets and real-world applications. The findings indicate that GAVE provides reliable adaptability for modern advertising platforms, as well as a learnable value function to tackle Out-of-Distribution (OOD) challenges. Overall, this comprehensive framework represents a novel approach to enhancing auto-bidding performance.