Balancing stream processing systems using reinforcement learning

This work consists of literature analysis and research. The literature part examines the workings of stream processing systems, way to measure their speed and the ability to tune the performance. Studies by other authors examining the auto–tuning of stream processing systems using reinforcement learning are also analyzed. The reinforcement learning algorithms used in other studies are reviewed and selected for use in the experimental part of this research. The research part defines the model of stream processing systems controlled using reinforcement learning and describes the goal function. The balancing algorithm used to perform the experiments is also defined. Experiments with REINFORCE, DQN and ACER algorithms prove that "Heron" stream processing systems can be balanced using reinforcement learning and that ACER achieves the best result of all analyzed algorithms, lowering the latency by 50 percent