This thesis presents a new tuning strategy for the node-level auto-tuned particle simulation library AutoPas. The strategy uses reinforcement learning to predict the best configuration for the simulation to use to achieve the fastest calculation time. An implementation of a modified version of the SARSA algorithm is shown. Furthermore, the hyperparameters: learning rate, discount factor, and exploration rate are fine-tuned trough grid search to produce the best possible results. The reinforcement learning tuning strategy is then tested according to different criteria. These criteria are then used to compare it against the full search and predictive tuning strategies. The reinforcement learning tuning interface shows considerable improvement compared to the already implemented options.     «

This thesis presents a new tuning strategy for the node-level auto-tuned particle simulation library AutoPas. The strategy uses reinforcement learning to predict the best configuration for the simulation to use to achieve the fastest calculation time. An implementation of a modified version of the SARSA algorithm is shown. Furthermore, the hyperparameters: learning rate, discount factor, and exploration rate are fine-tuned trough grid search to produce the best possible results. The reinforcemen...     »