PIRJeu - Machine learning experimentation

PIR is a french acronym, which goes for "Projet d'Initiation à la Recherche". It means, in english, "Initiation to Research Project". This was a school project given to me and my 5 teammates in 2020-2021. There was various subjects, but ours was simply to develop a game agent. 

We were originally meant to use Ludii's API and database to develop an artificial intelligence to play virtual board games as well as possible. But our teachers were really open to any suggestion, so we decided to create our own game, using Unity, and to implement various agents to play our own game as well as possible.

Concerning the game, it is pretty simple : the objective is to dodge incoming asteroids or ships, or detroy them using projectiles or bonuses, to survive as much as possible.

We decided to implement 3 different agents : 
- the first one would be using Unity's ML-Agents toolkit to create a model after many learning time. Any model would be reusable afterwards to see how well the learning strategy worked.
- the second one would be using a homemade neural network, to understand the whole concept better.
- the final one would be an agent specialized to our game, which would calculate everyframe the best position to aim to avoid obstacles for as long as possible. Though, as this is a specialized agent, it would have to be adapted with every new feature added to the game, or with every new mechanic in general.

For now, there are no "automatic" ways to start learning processes, or to load any neural network model. For the ML-Agents agent, a trained model is already loaded. For the homemade neural network agent, you will have to let the agent train for at least 20 minutes to see how well it performs. Finally, the specialized agent can be used immediately and should perform quite well.