Dynamic Computation Offloading in Ultra-Dense Networks based on Mean Field Games
Abstract : In ultra-dense networks, the increasing popularity of computation intensive applications imposes challenges to the resource-constrained smart mobile devices (SMDs), which may be solved by offloading these computation tasks to the nearby mobile edge computing centers. However, when massive SMDs offload computation tasks in a dynamic wireless environment simultaneously, the joint optimization of their offloading decisions becomes prohibitively complex. In this paper, we firstly model the joint optimization problem as a multi-user non-cooperative dynamic stochastic game, then propose a mean field game based algorithm to solve it with a drastically reduced complexity. We derive the two partial differential equations ruling the optimal strategies of the mean field game, namely the Hamilton-Jacobi-Bellman and Fokker-Planck-Kolmogorov equations, which are solved in an iterative manner in our proposed algorithm. Numerical results demonstrate that the proposed mean field game-based offloading algorithm requires a lower cumulated cost than the conventional strategies under the latency constraints of computation tasks, with perfect prediction of future channel states. It also appears that the performance of the mean field game-based offloading strategy depends on the accuracy of the future channel knowledge provided to the system, as the uncertainty may compromise its cumulated cost performance.
? Existing works requiring predefined stochastic dynamics of channels for learning strategies, this study adopted a model-free reinforcement learning mechanism to design an offloading policy for players.
? For the multi-user and multi-MEC scenario, most existing works do not provide an efficient approach that reduces energy consumption and price under latency constraints.
? The existence of binary variable xij changes the optimization problem to a mixed integer programming problem which is non-convex and NP-hard.
? In consideration of problem conditions and different factors, such as the deadline, energy consumption, distance of MDs, and lack of user awareness of others decisions, the problem is solved via the school choice.
? We think that the MFG will find wider applications in ultra-dense networks, and more types of MFG will also be implemented for various kinds of novel problems.
? This calls for large-scale convex optimization methods to deal with design problems with discrete and continuous variables.
? Game models can provide distributed solutions to the resource allocation problems for D2D communication.
? The inter-cell interference management was formulated as two coupled problems in , where the mean field theory was exploited to help decouple a complex large-scale optimization problem into a family of localized optimization problems.
• The proposed algorithm aims to minimize the price and energy consumption of the user task while meeting its deadline.
• The authors demonstrated that the proposed mechanism was individually-rational and returned envy-free allocations.
• In fog computing, they modeled the resource allocation problem as the SPA game, in which lecturers propose projects and students request these projects.
• In the proposed method, when MEC servers have depleted their capacity and there are still MDs without resources, the MD’s decision is to run tasks locally.
• Moreover, firstly in the proposed system, users can make offloading decisions to execute the task before the deadline and so save mobile device energy.
? In this article, we take network densification as the most important feature and dimension to improve network capacity and save energy, thus the 5G spectrum and energy efficiency.
? Both intra-tier and inter-tier interference degrade performance, which results by full frequency reuse among different nodes.
? Energy efficiency is a critical performance requirement for green communications, especially when small cells are densely deployed to enhance the quality of the users’ experience.
? Different players have various actions of parameters to adjust, and different preference and performance metrics always exist.
? However, it is known that the performance metrics in the 5G era should be various, which should include the perceived delay, reliability, and cost, thus leading to the QoE.
|