Abstract of doctoral thesis - Athanasios Psaltis

Federated, Multi-agent, deep reinforcement learning 

Within this Thesis, novel Deep Learning techniques will be developed for efficiently identifying the areas of the space where an agent does not have enough knowledge and query those data which bridge its knowledge gap. In particular, this work will exploit Active Reinforcement Learning algorithms (e.g. Deep Q Learning, Actor-Critic, Policy Gradient) in order to increase the performance and accelerate the training (by reducing the required number of training samples) of the developed machine learning algorithms (e.g. object detection, action recognition, etc.). Communication among agents will be achieved through a Federated Learning (FL) framework i.e. a decentralised training approach that allows processing nodes to contribute, which in principle is effective in building high-quality policies for agents under the condition that training data are not shared between agents. Additionally, FL will act as an umbrella term for centralized coordination strategies in a multi-agent environment, while leveraging enhanced local resources on each device. Moreover, a set of Federated analytics algorithms (e.g. differentially private federated GANs) will be exploited to reveal patterns in decentralized datasets that are difficult for the existing models to recognize.