New Math Shows How AI Bidders Stay Sharp in Endless Auctions

Two researchers, Ashkan Soleymani and Georgios Piliouras, have published a paper describing a new learning recipe called MORM — short for Multiplicatively Optimistic Regret Matching. It's a set of instructions for how self-interested players adjust their choices round after round. Crucially, each player learns entirely on their own, with no referee or central coordinator telling them what to do. And it works in 'general-sum' games — situations like real life, where one person's win isn't automatically another person's loss.

The magic word is 'regret.' In game theory, regret is the gap between what you actually earned and what you would have earned with perfect hindsight. Most learning methods let that gap pile up: the longer the game drags on, the further behind you fall. This new method keeps the gap essentially constant, growing only with the number of choices available, not with time. The secret ingredient is 'optimism' — a small built-in bias that assumes the next round will be slightly better than the last, which keeps AI from wildly overcorrecting.

Why should you care? These algorithms quietly run real markets. Ad auctions that set the price of online advertising, electricity markets, stock-trading bots, and delivery routing all rely on software that learns as it goes. When those bidders fail to track the best strategy, prices wobble and everyone pays for the inefficiency — higher ad costs, higher delivery fees, stranger energy bills. The paper also shows the method holds up even when other players act unpredictably or hostilely, which matters when the 'players' are competitors trying to game each other.

The honest catch: this is mathematics, not a shipping product. It assumes all players learn at the same time and can see the full outcome of each round. Real markets have hidden information, very few players, and humans who change their minds. So expect ideas like this to shape AI trading and bidding tools over years, not overnight — but it's a real step toward AI agents that can share a market without destabilizing it.