We are a group of friends that, besides having a fun time this weekend, wanted to dive deeper in the technologies we are interested in. Our interests differ from game development to machine learning, so applying both technologies building an AI for a game was the obvious choice.
What it does
Our QDNN reads the inputs from the game logic. Our project is different to every other QDNN project we've seen online since we do so NOT using computer vision, instead we read the wanted inputs from the game state, which we accordingly programmed. Using the game inputs (for positions and rewards) we are able to train two NN to predict the best choice at any given moment of the execution.
How we built it
The whole project is written in Python. The game is backed up by PyGame and the AI by PyTorch.
Challenges we ran into
We ran into many challenges, since we are new to RL and game develpment in Python. The main problem was coming up with the NN architecture we needed, since the original idea was to use a vanilla NN but we quickly realized that it would not be feasible for this RL project (and for most of them)
Accomplishments that we're proud of
While the QDNN training would take a while longer than the project's deadline we are overall very satisfied with having put together a working project for technologies we were absolute begginers a few hours ago.