Hands-On Reinforcement Learning for Games

Training a neural network – computational graph

In order to train a network or computational graph, we need to first feed it the input data, determine what the graph thinks is the answer, and then correct it iteratively using backpropagation. Let's go back to the Chapter_6_1.py code example, and follow the next exercise to learn how training works:

1. We will start at the beginning of the training loop that starts with the for loop, as shown in the following code:

for t in range(500):  h = x.mm(layer1)  h_relu = h.clamp(min=0)  y_pred = h_relu.mm(layer2)

2. So, 500 in this example denotes the total number of training iterations or epochs. In each iteration, we calculate the predicted output using the next three lines. This step is called the forward ...