Evolution of Alibaba's AI-Driven Advertising Decision Technologies

This article reviews the evolution of advertising intelligent decision technologies at Alibaba's Alimama platform, highlighting how deep learning, reinforcement learning, and generative AI have been applied to automated bidding and auction mechanisms.

It begins with an overview of the advertising ecosystem, distinguishing advertising from search recommendation and emphasizing the multi‑party optimization challenges involving media, advertisers, and the platform.

The core of the discussion is the progression of automated bidding strategies across four generations: classic control‑based methods, linear programming (LP) planning, reinforcement learning (RL) approaches, and finally generative model‑based bidding (AIGB). Each stage addresses increasing complexity and dynamic environments, moving from static budget‑control algorithms to RL‑based bidding that learns from large‑scale interaction data, and ultimately to generative models that directly generate bidding policies conditioned on constraints.

Key technical contributions include the Simulation RL‑based Bidding (SRLB) framework, Offline RL‑based Bidding that learns from logged decisions to avoid simulation bias, and Sustainable Online RL (SORL) that safely explores online environments. The article also describes the transition to generative bidding, where a large language model learns the joint distribution of bidding actions, objectives, and constraints, enabling more stable long‑sequence decision making.

In parallel, the paper surveys the evolution of auction mechanism design, from classic GSP/VCG auctions to learning‑based mechanisms such as Deep GSP and Neural Auction, and to integrated designs that jointly optimize bidding and auction rules. It discusses challenges like externalities, fairness across advertisers, multi‑agent coordination, and the need for incentive‑compatible mechanisms under private constraints.

Finally, the article reflects on future directions, including online‑offline hybrid RL, generative bidding models, and the joint optimization of bidding and auction mechanisms to achieve better performance, fairness, and scalability in large‑scale e‑commerce advertising.