Get our free extension to see links to code for papers anywhere online!

Malicious Experts versus the multiplicative weights algorithm in online prediction
Mar 18, 2020
Erhan Bayraktar, H. Vincent Poor, Xin Zhang

We consider a prediction problem with two experts and a forecaster. We assume that one of the experts is honest and makes correct prediction with probability $\mu$ at each round. The other one is malicious, who knows true outcomes at each round and makes predictions in order to maximize the loss of the forecaster. Assuming the forecaster adopts the classical multiplicative weights algorithm, we find upper and lower bounds for the value function of the malicious expert. Our results imply that the multiplicative weights algorithm cannot resist the corruption of malicious experts. We also show that an adaptive multiplicative weights algorithm is asymptotically optimal for the forecaster, and hence more resistant to the corruption of malicious experts.

* Keywords: Adversarial online learning, multiplicative weights algorithm, dynamic programming, partial differential equation approach, viscosity solutions