Networks

Auction theory is a branch of game theory that aims to explain how different bidders could behave in auction marketplaces and the Bayesian Nash Equilibrium (BNE) in auction theory is an outcome of an auction in which no participant can – after having considered all of the other bidders’ selections – better their bidding strategy. It is often used to predict the result of auctions. For auctions particularly, calculating the BNE through numerical techniques has been attempted several times but it is not easy since it is a market with continuous bids, multiple items and value interdependencies. The past successful studies have usually been limited to single item auctions. The article cited explains the machine learning technique introduced by Martin Bichler and his peers to identify the BNE for symmetric auctions which relies on artificial neural networks and finds the optimal policy for a bidding action based on gradient dynamics when the participant bids against himself/herself. The researchers tested their algorithm on various auction models and used certain standard assumptions when checking for the percentage error in their estimation but it turned out that their model almost coincided with the global auction equilibrium derived through analysis every time and across different auction models.

I believe that the technique developed by Bichler and his research colleagues can have various far reaching implications and can help predict the equilibrium using artificial neural networks for problems beyond just multi-item auction games. The technique also outmaneuvers any utility function disconnections or discontinuities by adapting the gradient descent metric to account for it. It is not only proven to be extremely precise but also versatile and can be used by analysts to check for changes in outcomes with small changes in the auction values. The use of supervised learning to train the networks towards obtaining approximately truthful initializations was particularly interesting as I believe this could greatly improve the accuracy of the obtained results and prevent any asymmetries or outliers from hampering the model’s self-learning ability. It would definitely serve as a helpful tool for auction participants to improve their strategies when bidding and also help decide auction arrangements or formats in the future.