Farzana Yusuf

Farzana Beente Yusuf, Vitalii Stebliankin, Giuseppe Vietri et al.

Inspired by the cache replacement problem, we propose and solve a new variant of the well-known multi-armed bandit (MAB), thus providing a solution for improving existing state-of-the-art cache management methods. Each arm (or expert) represents a distinct cache replacement policy, which advises on the page to evict from the cache when needed. Feedback on the eviction comes in the form of a "miss", but at an indeterminate time after the action is taken, and the cost of the eviction is set to be inversely proportional to the response time. The feedback is ignored if it comes after a threshold value for the delay, which we set to be equal to the size of the page eviction history. Thus, for delays beyond the threshold, its cost is assumed to be zero. Consequently, we call this problem with delayed feedback and decaying costs. We introduce an adaptive reinforcement learning algorithm EXP4-DFDC that provides a solution to the problem. We derive an optimal learning rate for EXP4-DFDC that defines the balance between exploration and exploitation and proves theoretically that the expected regret of our algorithm is a vanishing quantity as a function of time. As an application, we show that LeCaR, a recent top-performing machine learning algorithm for cache replacement, can be enhanced with adaptive learning using our formulations. We present an improved adaptive version of LeCaR, called OLeCaR, with the learning rate set as determined by the theoretical derivation presented here to minimize regret for EXP4-DFDC. It then follows that LeCaR and OLeCaR are theoretically guaranteed to have vanishing regret over time.

Sam Ganzfried, Farzana Yusuf

A problem faced by many instructors is that of designing exams that accurately assess the abilities of the students. Typically these exams are prepared several days in advance, and generic question scores are used based on rough approximation of the question difficulty and length. For example, for a recent class taught by the author, there were 30 multiple choice questions worth 3 points, 15 true/false with explanation questions worth 4 points, and 5 analytical exercises worth 10 points. We describe a novel framework where algorithms from machine learning are used to modify the exam question weights in order to optimize the exam scores, using the overall class grade as a proxy for a student's true ability. We show that significant error reduction can be obtained by our approach over standard weighting schemes, and we make several new observations regarding the properties of the "good" and "bad" exam questions that can have impact on the design of improved future evaluation methods.

Sam Ganzfried, Farzana Yusuf

Algorithms for equilibrium computation generally make no attempt to ensure that the computed strategies are understandable by humans. For instance the strategies for the strongest poker agents are represented as massive binary files. In many situations, we would like to compute strategies that can actually be implemented by humans, who may have computational limitations and may only be able to remember a small number of features or components of the strategies that have been computed. We study poker games where private information distributions can be arbitrary. We create a large training set of game instances and solutions, by randomly selecting the information probabilities, and present algorithms that learn from the training instances in order to perform well in games with unseen information distributions. We are able to conclude several new fundamental rules about poker strategy that can be easily implemented by humans.