Blind Multi-Stage Scoring Auctions with Two-Sided Uncertainty
Published:
Throughout this short study, we explore the problem of blind multi-stage scoring auctions with two-sided incomplete information. We identify the problem of few-shot value function estimation as a key step in the bidding process of such an auction. To that end we introduce a greedy online solver that approximates the structure of an unknown multi-attribute value function using a limited number of observed bids and their associated scores. Our algorithm breaks the approximation problem into tractable sub-problems then returns the shape of each objective’s value function along with their corresponding weights. We apply the proposed method to a case study in public works construction bidding. We first elicit a realistic value function through an expert interview, then simulate a bidder’s bidding behavior using online value function estimation. Our results indicate that the greedy online solver can quickly identify high-value bids and accurately approximate both the shape and weights of the government’s value function within a small number of rounds. However, we also identify certain limitations. Namely that the algorithm tends to plateau due to local non-convexity. Further work may include exploring the efficacy of hybrid methods that combine greedy search with neural network-based meta-learning strategies. Overall, our results suggest that few-shot learning algorithms can substantially mitigate the strategic uncertainty inherent in multi-stage scoring auctions.