General Track Winner (Dsat2) Presentation

Team DSAT2 proposed a training framework for auto-bidding agents which combines imitation learning with elements of operations research. They used a greedy algorithm to compute a near-optimal bidding strategy assuming perfect information about the whole advertisement campaign. The bids computed by this “oracle” algorithm were used as a supervision signal for an agent only having access to the real-time campaign data. They called this framework “Oracle Imitation Learning”. DSAT2 ranked 1st in the official phase of the challenge and 6th in the final phase.