As the very large-scale integrated circuit designs enter the deep sub-micron era, many-core processors are regarded as promising architectures to keep up with the Moore’s law. To provide effective communications between the on-chip components, network-on-chip was proposed as a new paradigm that exhibits better scalability than the traditional buses. There have been previous researches on application mappings to reduce the power consumption, the network latency and the network area overhead. However, some of the previous proposed algorithms such as the Kernighan–Lin algorithm (KL) and some genetic algorithms (GA) have the problem of finding the local best result instead of a global optimal solution. In this paper, we propose a novel application mapping algorithm for the mesh-of-tree network topology, called KL_GA algorithm. Our proposed algorithm takes the advantage of both the Kernighan–Lin algorithm and genetic algorithms to reduce the overall communication cost. Our KL_GA algorithm first generates a mapping solution using a KL-based method. In order to avoid the appearance of premature phenomena, we next apply a GA-based algorithm to get rid of the population trapped in the local optimum and re-generate a new population. Our evaluations show that, compared to the random mapping algorithm, our KL_GA algorithm saves the power by 21.6 % and reduces the network latency by 16.3 % on the average.

In the current mobile edge computing (MEC) system, the user dynamics, diversity of applications, and heterogeneity of services have made cooperative service caching and task offloading decision increasingly important. Service caching and task offloading have a naturally hierarchical structure, and thus, hierarchical reinforcement learning (HRL) can be used to effectively alleviate the dimensionality curse in it. However, traditional HRL algorithms are designed for short-term missions with sparse rewards, while existing HRL algorithms proposed for MEC lack delicate a coupling structure and perform poorly. This article introduces a novel HRL-based algorithm, named hierarchical service caching and task offloading (HSCTO), to solve the problem of the cooperative optimization of service caching and task offloading in MEC. The upper layer of HSCTO makes decisions on service caching while the lower layer is in charge of task offloading strategies. The upper-layer module learns policies by directly utilizing the rewards of the lower-layer agent, and the tightly coupled design guarantees algorithm performance. Furthermore, we adopt a fixed multiple time step method in the upper layer, which eliminates the dependence on the semi-Markov decision processes (SMDPs) theory and reduces the cost of frequent service replacement. We conducted numerical evaluations and the experimental results show that HSCTO improves the overall performance by 20%, and reduces the average energy consumption by 13% compared with competitive baselines.

Mobile edge computing (MEC) can enhance application performance effectively by offloading computation tasks to the edge server hosting corresponding service via multi-access wireless networks. However, existing service placement policies are more based on the assumption that application services can be used without restrictions, while ignoring the constraints of software usage by license, which always limits user number, usage periods, etc. To address these challenges, in this paper, we introduce an architecture of digital twin-empowered MEC with good scalability and reliability, formulate an optimal problem by jointly optimizing task offloading and service placement, which takes into account the usage number limitation for one service at the same time. To tackle this mix integer non-linear programming problem (MINLP), we propose a DRL-based algorithm JTOSP to deal with high-dimensional input data from MEC system, and to cope with the stochastic nature in dynamic underlying network efficiently. Numerical experiments are conducted and the simulation results show that our algorithm outperform other algorithms and reduce energy consumption efficiently.

Intelligent water platform is based on cloud computing and big data to build a comprehensive water management system. The system can analyze and process massive water information in a timely manner through cloud management, and make corresponding processing results to assist decision-making suggestions, so as to manage the whole production, management and service process of water system in a precise and dynamic way, so as to reach the intelligent state.

Concurrent multipath transmission combining FEC is an effective approach to achieve ultra-reliable low latency communication in modern network. Rate allocation algorithm which distributing packets over multiple parallel paths plays a critical role in this architecture. However, traditional heuristic algorithm becomes less applicable due to great computational burdens and increasing demands for real-time. In this paper, we propose a deep reinforcement learning based algorithm NeuroCMT to solve this problem. Moreover, in order to deal with large discrete action space problem in NeuroCMT, we introduce a method with low complexity and high generalization ability. Furthermore, we implement prioritized experience replay technique in DDPG to attain better sample efficiency and enhanced convergence speed utilizing prioritized sampling. Extensive experiments are performed and the evaluation results demonstrate clearly that our approach achieves high reliability as well as low latency effectively.