The paper introduces a multi-task learning approach, called MT-MLCA, to enhance preference elicitation in iterative combinatorial auctions (ICAs) with many participants.
The key insights are:
Existing ICA methods train separate machine learning models for each bidder, which can be inefficient when there are numerous bidders with similar valuation functions.
MT-MLCA applies soft parameter-sharing across bidder models to capture the intrinsic relationships between their valuations. It also utilizes bidder ID features to help the models differentiate task differences.
Experiments on a spectrum auction simulator show that MT-MLCA achieves higher efficiency than the baseline method, especially in scenarios with many bidders and a limited number of queries.
The performance gains are more pronounced in settings with a large number of items (196) and similar bidding patterns, as the increased substitutability among items augments the relatedness of valuation functions, allowing MT-MLCA to better leverage the shared information.
While MT-MLCA-F and MT-MLCA-R (two variants of the multi-task approach) do not consistently outperform each other, they both demonstrate advantages over the baseline in the appropriate settings.
To Another Language
from source content
arxiv.org
Djupare frågor