by OnAnyPoint
This is the source for the final report of my master's project at EPFL. In the report, the work I have done over 6 months at NEC Laboratories Europe is presented.
A PDF version is available here.
This exploratory work looks at a decentralised multi-agent reinforcement learning problem in which each agent is trying to maximise a global utility. Agents having only access to partial observations of the environment, a key challenge is coordination. To mitigate the effect of the partial observability, the agents must learn to communicate and share the information required to solve the task. The main objective of this work is to provide an in-depth look at the quality of the learning, the behaviour of the agents and the learned communication protocols.