PyTorch 1.x Reinforcement Learning Cookbook

How it works...

As you can see in Step 2, the action estimator has two layers – input and output layers, followed by a sigmoid activation function, and the loss function is binary cross-entropy.

Step 3 is for training the cross-entropy model. Specifically, for each training episode, we take random actions, accumulate rewards, and record states and actions. After experiencing n_episode episodes, we take the most successful episodes (with the highest total rewards) and extract n_samples of (state, action) pairs as training samples. We then train the estimator for 100 iterations on the training set we just constructed.

Executing the lines of code in Step 7 will result in the following plot:

As you can see, there are +200 rewards for all testing ...