AIGB Track Winner (KGAB) Presentation

The existing DT baseline grapples with two main issues: (i) misaligning its return-to-go function to maximize total conversions, and reliance on behavior cloning, which constrains it to average strategies; (ii) performance constraints due to dataset quality and bidding complexities. To solve them, our proposed AIGB solution involves (i) enhancing DT structures for optimal bidding strategies on the original dataset by optimizing objectives and exploring effective actions during training and (ii) enhancing bidding performance by learning DT from an expert dataset. These steps significantly boost baseline performance by refining the DT structure and regularizing with the expert dataset.