Using a half cheetah habitat for random augmentation computing

Reinforcement learning algorithms that depend on physical models are thought to provide more substantial outcomes in comparison to model-free approaches when implemented on dynamic systems. Previous research has mostly focused on addressing self-management difficulties via the use of sophisticated neural network models. In order to get state-of-the-art outcomes, these models need a significant amount of training data. When using model-free policy search techniques, it is often considered that model-free reinforcement learning processes with policies that provide worse outcomes may be avoided by utilizing augmented random search (ARS), model-free. This technique improves the effectiveness and speeds up the training of linear methods for performing control tasks on the virtual physics engine Half-Cheetah Environment. To reach this objective, we use the computational efficiency of Augmented Random Search (ARS) to assess the agent's performance over several random episodes and hyper parameters. The study also showcases the effectiveness of the search approach used for this project via the examination of simulation data, including many events and agents' behaviors. Our simulation reveals that the commonly used metric for assessing the efficiency of RL learning is inadequate in accurately determining the efficacy of specific circumstances. In these instances, Augmented Random Search (ARS) surpasses other algorithms by achieving more rewards.