Reinforcement Learning for Real Time Bidding

When an internet user opens a web page containing an advertising slot, how is it determined which ad is shown? Today, the most common software-based approach to trading advertising slots is real time bidding: as soon as the user begins to load the web page, an auction for the slot is held in real time, and the highest bidder gets to display their advertisement of choice. Auction bidding is performed by different demand side platforms (DSPs). Emerse AB, where this master's thesis work was carried out, owns and operates such a DSP. Each bidder (Emerse and competing DSPs) has a limited advertising budget, and strives to spend it in a manner that maximizes the value of the advertisement slots bought. In this thesis, we formalize this problem by modelling the bidding process as a Markov decision process. To find the optimal auction bid, two different solution methods are proposed: value iteration and actor–critic policy gradients. The effectiveness of the value iteration Markov decision process approach (versus other common baselines methods) is demonstrated on real-world auction data