Distributed caching based on decentralized learning automata

In this paper we propose a novel distributed caching scheme in Heterogeneous Cellular Networks (HCN). We are interested in optimizing the content placement in order to minimize the downloading latency. We achieve this in a decentralized manner, based on a game of independent learning automata (LA). First, we propose a faster-converging discrete generalist pursuit algorithm (DGPA) for a single LA based on the concept of conditional inaction (CI), referred to as CI-DGPA. Then we develop a framework for a game of LA based on CIDGPA defining the information exchange between learners and the environment. Within this framework, we design a reward function that approaches the performance of a greedy algorithm and show that a smart partition of the search space can double the game convergence speed, thereby halving the overhead due to signalling. Simulations show that our scheme can approach the greedy algorithm with a very small performance gap while providing a much lower computational complexity.