Change input on the Actor-Critic neural network model

[FanchenBao]

The current model naively takes in the grid coordinates as input, which includes the value 0. This is bad for neural network because no learning is possible when the input is 0. To resolve this issue, we shall use other ways to encode the state. One way is to use on-hot encoding for each cell in the grid. It might also be possible to use a simple labeling system from 1 to 100 for each cell, and then normalize the label as the input. They are both worth the shot. Maybe by changing the input, we can make the on-policy learning work.