With such a complex setup, you can not hard-code parameters in your python code. Two combinations of HPs lead to very different outcomes, image by the author. The sweep requires knowing the useful range or values of all parameters, creating the agent and its models, and then training them enough to see if learning converges. This is done by training the agent with a set of hyperparameters and evaluating the average reward obtained from validation episodes. The breadth of the search space requires you to systematically sweep the combinations of hyperparameters to find the best combination to optimize the agent's performance. This difficulty is attracting a lot of attention, with recent popular answers to the problem from the likes of Google's research team. Just by looking at a widespread implementation of SAC, from stable-baselines3, they have 25 parameters, most of which depend on your own use case and contribute to the success of optimizing a strategy. There are many levers to make learning more stable, faster, or save some memory. RL algorithms rely on MANY hyperparameters, some for learning, some for the underlying Neural Networks. In a previous post, I go into more depth about the concepts of Reinforcement Learning. They aim to maximize the rewards they accumulate by interacting with the environment and slowly learning the best actions from experience. Agents make sequences of actions, following strategies that dictate how they should react in each situation.

Reinforcement Learning is a type of Machine Learning that involves training agents to interact with an environment.

Photo by Leonel Fernandez on Unsplash Introduction

Tune sweeper 4 activation code