Memory Bounds for the Experts Problem

Online learning with expert advice is a fundamental problem of sequential prediction. In this problem, the algorithm has access to a set of n "experts" who make predictions on each day. The goal on each day is to process these predictions, and make a prediction with the minimum cost. After making a prediction, the algorithm sees the actual outcome on that day, updates its state, and then moves on to the next day. An algorithm is judged by how well it does compared to the best expert in the set. The classical algorithm for this problem is the multiplicative weights algorithm, which has been well-studied in many fields since as early as the 1950s. Variations of this algorithm have been applied to and optimized for a broad range of problems, including boosting an ensemble of weak-learners in machine learning, and approximately solving linear and semi-definite programs. However, every application, to our knowledge, relies on storing weights for every expert, and uses Omega(n) memory. There is little work on understanding the memory required to solve the online learning with expert advice problem (or to run standard sequential prediction algorithms, such as multiplicative weights) in natural streaming models, which is especially important when the number of experts and number of days are both large. We initiate the study of the learning with expert advice problem in the streaming setting, and show lower and upper bounds. Our lower bound for i.i.d., random order, and adversarial order streams uses a reduction to a custom-built problem with a novel masking technique, to show a smooth trade-off for regret versus memory. Our upper bounds show new ways to run standard sequential prediction algorithms in rounds on small "pools" of experts, thus reducing the necessary memory. For random-order streams, we show that our upper bound is tight up to low order terms. We hope that these results and techniques will have broad applications in online learning, and can inspire algorithms based on standard sequential prediction techniques, like multiplicative weights, for a wide range of other problems in the memory-constrained setting.