Welcome to the Practical Activity of Part 2!
In this part we are going to build and train an AI to play the game of Doom!
First, we will build an AI, which will be the Deep Convolutional Q-Learning model combined to Eligibility Trace.
Then, we will train it to play the game of Doom.
And finally, we will watch a video of our AI playing Doom after it was trained.
And here is the environment of the game on which we will be training our AI to play inside:
Pretty cool isn't it?
The three essential pieces of info you need to know about the environment are the following:
The input states are the frames of the game.
The possible output actions are: ATTACK, MOVE_RIGHT, MOVE_LEFT, MOVE_FORWARD, MOVE_BACKWARD, TURN_RIGHT and TURN_LEFT. Thus 7 actions in total.
The rewards are:
- Plus distance for getting closer to the vest (which is at the end of the corridor).
- Minus distance for getting further from the vest.
- Minus 100 points for getting killed.
Again we will build this AI with PyTorch, a highly advanced Deep Learning & AI library. Every single line of code will be explained in details but I would recommend to have a look at the PyTorch documentation to get even more familiar with PyTorch:
Enjoy AI!