chapter 9, case study 3, fuction nn_pred_to_weights is not reasonable

[John87654]

In DQN, the Q-network is an approximation of optimal action value fuction, so that the action with max value is the action agent should choose. If the input to Q-network is only the state, the output should be value of each action, so the output dim should be same as action dim. The implicit policy is the argmax of action value. However, in this code, the action is a complex mapping from all action value to asset weight (in fuction nn_pred_to_weights), which I think is not reasonable.