TUM researchers present machine learning method capable of learning local equilibria in symmetric auction games using artificial neural networks
Computer scientists have been studying the possibility of applying game theory and artificial intelligence (AI) technologies to chess, the abstract strategy board game, and other games for several decades. Another practical application of game theory is in economics, where it is used to explain strategic market interactions and eventual outcomes.
Auction theory is one of the most frequently used theoretical constructs to apply game theory in economics. Auction theory is based on a game theory that explains how different bidders might behave in auction markets.
Auction theory can be applied to both real and realistic markets. Using this approach, however, calculating equilibrium bidding strategies for bidding games becomes difficult because there are multiple items for sale with values ââthat depend on each other. When no player (or bidder) can improve on their chosen strategy after considering their opponent’s options, Bayesian Nash Equilibrium (BNE) occurs in game theory.
The BNE is considered a predictable outcome of a game or auction because it is a stable consequence. Compared to finished complete information sets like rock-paper-scissors, bids are still difficult to calculate. Because the values ââand offers of opponents are spontaneous and continuous, it happens.
Recent studies have introduced several digital techniques that could be used to master the balances in auction games. These methods are either based on calculations of the best punctual responses in the strategic space, or on the iterative resolution of subsets. Their use was mainly limited to simple auctions of unique items.
Researchers from the Technical University of Munich recently proposed a new machine learning technique that could be used to learn local equilibria in symmetrical auction games. This technique was introduced in an article published in Nature Machine Intelligence and works by representing strategies such as neural networks and then applying policy iteration based on gradient dynamics while a bidder plays against itself.
The research team has been studying auction theory and exploring its applications for several years now. Their recent study was specifically aimed at developing a technique based on artificial neural networks and self-play that can automatically adapt balance bidding strategies in auctions.
When the researchers tested their technique, it was found that the BNEs it approached collided with the analytically derived equilibrium whenever available. The estimated error rate was also insufficient in cases where the analytical balance is still unknown. In the future, the tool they built could be used to study auction efficiency and determine what bidding strategies can be expected from breakeven.
In addition to its crucial contribution to the study of auction theory, the technique created by the researchers and his team could be a valuable tool for auctioneers, as it could help them select auction formats and also help bidders develop their bidding strategies. For example, this could be useful in spectrum auctions, which regulators around the world use to distribute the rights to transmit signals over specific electromagnetic spectrum bands to different mobile network providers.
They mainly adopted the standard learning process in neural networks to deal with the obstacles of utility functions in their auction models and were also able to prove that the method converges to equilibrium in auctions with just one set. moderate assumptions.
The researchers want to test their method under various circumstances to make sure it will generalize well in future studies. They also intend to create tools that can automatically calculate balances across a wider range of game theory challenges, including some that go beyond symmetrical auction games.