Mechanism Design Via Machine Learning: Overfitting, Incentives, and Privacy
summary
Research shows how to tune a parameterized auction family—most simply the second-price (Vickrey) auction with reserve rho—by optimizing empirical revenue on a training set drawn from an unknown distribution D and then asks whether that parameter overfits or generalizes to true expected revenue. Methods are given to estimate ex-ante approximate incentive compatibility from samples (EC 19 with Nina Balkan and Thomas Sandholm), coping with jumpy, non‑Lipschitz utility functions, and to train mechanisms while preserving training‑set privacy and masking user signals in ad auctions.
AI-written summary · about
chapters (9)
- 0:00Mechanism design via machine learning Machine learning is increasingly used for mechanism design
- 0:27Auction for 1 item
- 1:44Parameterized mechanisms Many other types of parameterized mechanisms
- 3:18Overfitting
- 3:33Broader picture A mechanism is a special type of algorithm
- 4:57Algorithm design via machine learning
- 5:34Manipulable mechanisms
- 6:09Estimating apx incentive compatibility Our goal: Estimate how far a manipulable mechanism is from IC
- 7:21Privacy and societal impact Mechanism design via ML uses a lot of sensitive information
related talks
01h 07mMulti-Dimensional and Non-Linear Mechanism Design and Approximation I
32mLearning Reserve Prices in Second-Price Auctions
35mThe Sample Complexity of Revenue Maximization
40mOnline Ad Allocation: Robust Optimization for Repeated Auctions
27mEconometrics for Learning Agents
19mMarket Algorithms for Autobidding
01h 10mMulti-Dimensional and Non-Linear Mechanism Design and Approximation II
45mAGT and Machine Learning