No abstract

No abstract

The use of decentralized reinforcement learning (RL) in the context of multi-agent systems (MAS) poses some difficult problems. The speed of the learning process for example. Indeed, if the convergence of these algorithms has been widely studied and mathematically proven, they suffer from being very slow. In this context, we propose to use RL in MAS in an intelligent way to speed up the learning process in these systems. The idea is to consider the MAS as a new environment to be explored and the communication, between the agents, is limited to the exchange of knowledge about the environment. The last agent to explore the environment has to communicate the new knowledge to the other agents, and the latter have to build their knowledge bases taking into account this knowledge. To validate our method, we chose to evaluate it in a grid environment. Agents must exchange their tables (Qtables) to facilitate better exploration. The simulation results show that the proposed method accelerates the learning process. Moreover, it allows each agent to reach its goal independently.

IT transformation has revolutionized the business landscape and changed most of organizations business model into digital and innovation driven firms. To fully take advantage of this digitalization and the exponential growth of data, organizations need to rely on resilient, scalable, extremely connected, highly available & very performant systems. To meet this need, this paper presents a model of middleware for multi micro-agents system based on reactive programming and designed for massively distributed systems and High-Performance Computing, especially to face big data challenges. This middleware is based on multi-agents systems (MAS) which are known as a reliable solution for High Performance Computing. This proposal framework is built on abstraction and modularity principles through a multi-layered architecture. The design choices aim to ensure cooperation between heterogeneous distributed systems by decoupling the communication model and the cognitive pattern of micro agents. To ensure high scalability and to overcome networks latency, the proposal architecture uses distribution model of data & computing, that allows an adaptation of the grid size as needed. The resilience problem is addressed by adopting the same mechanism as Hazelcast middleware, thanks to his peer-to-peer architecture with no single point of failure.