A soft actor critic implementation, based on the Haarnoja etc al. paper (https://arxiv.org/pdf/1801.01290), with a learnable alpha value instead of the regular set value.
Utilizes priority replay with a sum-tree approach for faster retrieval. Trialled with normalization pre-processing functions specific for the intended environment.
Designed to work with OpenAI Gym - supports multiple environments to parallelize training and tested on OpenAI Gym Humanoid V4.