A fuzzy reinforcement learning approach for self-optimization of coverage in LTE networks

Abstract

many of the on-site operations. Additionally, about 24 percent of a typical wireless operator's revenue is normally spent on network OPEX, including training, support, power, transmission, and site rental [14]. Self-optimization functions can reduce the workload for site survey and analysis of network performance and thus reduce OPEX. Moreover, energy-saving functions enabled by self-optimization capabilities reduce costs for power consumed by the equipment. Additionally, the improved quality of user experience due to an optimal configuration of network elements and allocation of resources boosts the operator's profile and opens up new avenues for revenue. Figure 1 illustrates the impact of employing self-organizing network (SON) functionalities on various stages of a

Self-optimization of coverage and capacity based on a fuzzy neural network with cooperative reinforcement learning

“…The algorithm stores past successful optimization instances that improved the performance in the memory and applies these instances directly to new situations. In [19][20][21], a fuzzy Q-learning algorithm was used to learn the optimal antenna tilt control policy based on the continuous inputs of current antenna configuration and corresponding performance, and output of the optimized antenna configuration. Yet, the impact on neighboring cells due to such an adjustment was neglected.…”

A Fuzzy Reinforcement Learning Approach for Pre-Congestion Notification Based Admission Control

“…see [15,16] to introduce autonomic capabilities in network control systems, and is a combination of fuzzy logic [17] with Q-learning (type of Reinforcement Learning (RL)) [18] that aims to combine the robustness of a rule based fuzzy system with the adaptability and learning capabilities of Q-learning. In this Section we highlight the main concepts and benefits of this approach and its applicability in the context of PCN-based AC.…”

“…The objective of an agent is to find, by trying out all the possible actions when being in a given state, the action that maximizes its long term reward. The detailed mathematical foundation and formulation of Q-learning can be found in [15,16,18] therefore it is not repeated here, due to space limitations; the core Q-learning algorithm [18] is provided though, so as to highlight the parameters involved in it and consequently in our evaluation in the following Session.…”

“…It can also be easily combined with fuzzy logic and provide the association between the states and the available actions, which is the same as constructing the fuzzy logic "if…then" engine; the only additional step required being the distribution of the reinforcement/reward signal among multiple simultaneously triggered "if…then" rules due to the overlapping of the fuzzy input sets. This distribution can be done proportionally to the strength with which each rule is triggered so that rules (states) triggered with high strength -and contributing therefore more to the output action-are rewarded (or penalized) more compared to rules triggered with lower strength [15,16].…”

Fault compensation algorithm based on handover margins in LTE networks

“…Several optimization algorithms based on antenna tilt modifications can be found in the literature. Many of them are used as COC algorithms [16][17][18]. These works present different methodologies.…”