Organizer: Merouane Debbah (CentraleSupélec) and Alessio Zappone (University of Cassino and Southern Lazio)
Honggang Zhang (Zhejiang University) - Stigmergic Reinforcement Learning with Multi-Agent Collaboration for 6G Networks
Anticipating the rapid evolution of wireless mobile applications in forthcoming 6G networks, it emerges stronger incentive to design intelligent collaboration mechanisms among the distributed mobile agents. Following their individual observations, multiple intelligent mobile agents could effectively cooperate and efficiently approach the final common objective through collective learning from the environment. Up to date, independent reinforcement learning (IRL) has often been deployed within the multi-agent collaboration to alleviate the dilemma of non-stationary learning environment. However, behavioral strategies of the intelligent mobile agents in IRL could only be formulated upon their local individual observations of the global environment, and appropriate communication mechanisms must be established to reduce their behavioral localities.
In this speech, we tackle the challenging communication problem among the distributed mobile agents in highly mobile networks by jointly adopting two advanced mechanisms with different scales. For the large scale, we introduce the stigmergy mechanism as an indirect communication bridge among the independent learning agents and carefully design a mathematical representation to indicate the impact of digital pheromone. For the small scale, we propose a conflict-avoidance mechanism between adjacent mobile agents by implementing an additionally embedded neural network to provide more opportunities for participants with higher action priorities. Besides, we also design a federal training method to effectively optimize the neural networks within each mobile agent in a decentralized manner. Finally, we establish a testing platform where a number of mobile agents in a certain area move automatically to form a specified target shape, and demonstrate the superiorities of our proposed methods.
Tony Quek (Singapore University of Technology and Design) - Modeling, Learning, and Control in 6G
Claude Shannon founded information theory with a landmark paper “A Mathematical Theory of Communication” that he published in 1948, which is the fundamental cornerstone of today’s communication systems. At the same time, Norbert Wiener published his book “Cybernetics; or, Control and Communication in the Animal and the Machine”, which established the science of cybernetics. Interestingly, these two fields have developed in parallel and we are seeing the possible intersection of these two fields in future wireless systems or 6G. In this talk, we will motivate the need to understand modeling, learning, and control in future wireless systems by sharing some preliminary results in this area.
Modeling Challenges in the Emerging Internet Applications
Organizer: Phuoc Tran-Gia (University of Wuerzburg, Germany)
Onno Boxma(Eindhoven University of Technology) - Redundancy scheduling
Redundancy scheduling has emerged as a powerful strategy for improving response times in parallel server systems. The key feature in redundancy scheduling is replication of a job upon arrival by dispatching replicas to different servers. Redundant copies are removed as soon as the first of these replicas completes service.
Analytical results for redundancy scheduling are scarce. Even for the sta- bility condition only some partial results have been obtained. In this talk we shall consider some performance issues of redundancy scheduling, focussing on a rather general model in which there may be several types of jobs and in which the speeds of different servers for a job type may also differ. We shall in particular pay attention to the learning aspect in case the job types and/or the speeds of the various servers are not a priori known.
Note: This talk is based on joint work with Sem Borst and Youri Raaijmakers.
John Daigle (University of Mississippi) - Role of Classical Methods in Supporting Emerging 5G-based Applications
We begin with a high level overview of networking from the perspective of supporting emerging 5G applications in consideration of emerging technologies including but not limited to network function virtualization and software defined networking. A small number of areas within the core network that appear to be critical to performance at the application level will be identified, and performance modeling requirements for achieving meaningful contributions in those areas will be discussed. The potential role of classical probabilistic and optimization modeling tools in combination with live measurements for developing performance insights that help to stay ahead of the performance curve will be discussed. A perspective on the future role of analytical modeling based on this limited study will be drawn.
Franco Davoli (University of Genoa) - Modelling performance and power consumption for lifecycle management and dynamic control of virtualized networking platforms
The convergence of fixed and mobile networks toward programmable networking platforms based on functionalities implemented for the most part by Virtual Entities (VEs) running on a general-purpose hardware infrastructure is posing new challenges to the designer of control strategies that aim at optimizing the trade-off between performance and power consumption. Indeed, attributing the power consumed by the hardware to specific VEs becomes much more difficult than in networking architectures where the same functionalities are performed by dedicated equipment (e.g., routers or L2 switches). The talk will explore some of the issues that this environment poses on finding suitable traffic and power models to be used for control and lifecycle management of Virtual Network Functions (VNFs).
We have one planet that we can call our home. Yet we are not able to protect it from destruction. We have a desire as humanity to do better but are confounded by boundaries of nations and lack of frameworks to take hard steps. Earth 2.0 is an effort pioneered to build utilities for the planet that help users to confront and solve some of the hard problems that the world is currently encountering; namely, pollution, income-disparity, poverty etc.
For many of these hard problems we have a view from each of our vantage point and corresponding helplessness to handle the scale. We feel like a group of people on an island struggling to survive. To survive and thrive, we need tools to organize resources, share stories of survival and problems, and a to document and distribute actionable knowledge. These utilities for the world are being developed as collaborative effort by many different folks and is organized as Earth 2.0 (earth2.ucsd.edu)
Just as this idea was emerging the world was jolted with the COVID calamity. COVID became the first problem for earth 2.0 to work on. As part of the effort we created three tools, Corespond, Oassis and Homebound. The talk will present these tools and talk about how plugins are being conceived that would allow sophisticated disease modeling tools and behavior prediction/prevention tools would be integrated.
This framework is an innovative way for modern operational scientists to collect data, develop predictive models and then provide actionable knowledge at scale to help humanity. This construct can evolve to many different scenarios. The talk is an invitation for people to join the effort and leverage the situational data to develop useful models and algorithms to help humanity.
Hisashi Kobayashi (Princeton University) - A Stochastic Model of an Infectious Disease
A new stochastic model of an infectious disease, based on the birth-and-death-with immigration process, is presented. The model was originally proposed in the study of population growth or extinction of some biological species, but seems unknown to epidemiologists. We believe this model can capture the essence of infection diseases such as COVID-19, and shed lights on unpredictable stochastic behavior of epidemics, much better than the SIR deterministic model and its variants studied by biologists and epidemiologists.
Our non-homogeneous stochastic model, where the infection rate and recovery rate are arbitrary functions of time, should allow us to characterize infection patterns more accurately than has been possible before. This model should allow us to understand probabilistically the effect of control policies, such as social distancing, on the epidemic pattern.
Use of the saddle-point integration technique is also presented to obtain approximate probability distributions when an exact solution is unavailable.