3 skills to master before reinforcement learning (RL)

1. Supervised learning

Modern reinforcement learning is almost entirely focused on deep reinforcement learning . The word in the “ deep ” in the phrase deep reinforcement learning implies the use of a neural network in a core aspect of the algorithm. The neural network does some high-dimensional approximation in the learning process. That being said, the model does not need to have many layers and features, which is a common misconception that deep implies many layers.

Almost all of the courses and tutorials will assume you can fine-tune simple neural networks to approximate state values or create a final policy . These models are historically highly sensitive to all of the following training parameters: learning rate, batch size, model parameters, data normalization, and more. Doubled with tasks that are difficult to solve, debugging RL can be very difficult, and just seem like a binary it works or it doesn’t . Eliminating tails of confusing by knowing that all the sub approximations made are up to par. The best way to do this would be to learn supervised learning, then let an AutoML tool finish the job for you.