Abstract
This dissertation investigates the effects of the distribution of Deep Reinforcement
Learning (DRL) by developing novel vertically and horizontally distributed
actor-model-base architectures. The study examines the impact of vertical and
horizontal distribution on the performance and stability of DRL algorithms, and
validates the proposed actor-model-based architectures through the development
and testing of model specific prototypes in a suitable context. Experimental evaluations
suggest that there may be a correlation between batch size and learning
rate requirements alongside possible improvements in training speed and stability
facilitated by the distribution of DRL.