The deployment of edge servers make a significant impact on the service quality of a Mobile Edge Computing (MEC) system. This service quality relies on solving two key sub-problems: 1) interference management between servers 2) the placement of MEC servers. To improve the Quality of Service (QoS), we propose a method based on Graph Partition (GP) and Upper Confidence Bound (UCB) for solving these two sub-problems. Regarding interference management, we use an undirected graph to represent the interference between MEC servers so that the overall graph can be divided into multiple subsets of non-interfering MEC servers. Regarding server placement, we propose a Multiple Choice-Upper Confidence Bound (MC-UCB) algorithm that place an collection of interference aware edge servers in each selection. To evaluate the performance, we define a user's QoS function based on transmission delay, throughput, and user density comprehensively and compared with Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) from previous work. The simulation results show that the performance of the proposed algorithms is improved by more than 4% compared with the GA algorithm and 6% compared with the PSO algorithm.

Massive terminal users have brought explosive need of data residing at edge of overall network. Multiple Mobile Edge Computing (MEC) servers are built in/near base station to meet this need. However, optimal distribution of these servers to multiple users in real time is still a problem. Reinforcement Learning (RL) as a framework to solve interaction problem is a promising solution. In order to apply RL based algorithm into a multi-agent environment, we propose an iterative scheme: select individual users with priorities to interact with the environment iteratively one at a time Furthermore, we tried to optimize the overall system performance based on this scheme. Hence, we construct three objective system performance indicators: average processing cost, delay and energy consumption, improve the existing Deep Q-learning Network (DQN) by using the cost as reward function, changing the fixed exploitation rate into dynamic one that associated with reward and episode time. In order to explore the performance potential of the proposed algorithm, we have simulated the proposed algorithm, DQN algorithm and greedy algorithm under different users and data sizes. The results show that the proposed algorithm had reduced at least 12% of system average processing cost comparing to the greedy algorithm. It also outperform the greedy algorithm and DQN algorithm in delay and energy consumption significantly.