Optimal Uncoded Cache Placement Policy by Markov Decision Process

Please enable javascript in your browser.

Rezaei, Elahe | 2014

728 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 45962 (05)
University: Sharif University of Technology
Department: Electrical Engineering
Advisor(s): Hossein Khalaj, Babak
Abstract:
Replicating server information at shorter distances from users and delivering them at more suitable times, decreases average delay of serving requests as well as balancing backhaul link traffic. Due to limited cache size, cache placement strategies play key role in efficiency of such approaches. Well-known heuristic caching mechanisms such as Least Frequently Used (LFU) and Adaptive Replacement Cache (ARC), while quite efficient in terms of complexity, are not optimal in terms of hit rate which is the key caching performance index. In this paper, we address the optimal cache placement problem from an analytic Markov Decision Process (MDP)-based view to maximize cache hit rate. In this framework, the a-priori popularity distribution of files is exploited to achieve optimal cache hit rate performance. As shown in the paper, even in the case that such distribution is not known a-priori, learning mechanisms such as Q-Learning can be adopted to outperform LFU and ARC at the cost of modest additional complexity
Keywords:
Reinforcement Learning ; Markov Decision Making ; Cache Placement ; Popularity Profile