This repository contains the project for the UniTS reinforcement learning course A.Y. 2022-23. It was done in collaboration with @alessimichele and @ElenaRivaroli.
Here you can find the final presentation slides.
The goal is to teach an agent how to play the game flappy bird using the text-flappy-bird-gym.
The agent will have no knowledge of the environment but perfect observability of the states. As such it is phrased as a model-free reinforcement learning problem.
Standard Reinforcement Learning techniques for policy control will be attempted and compared:
For each methods, various set of hyperparameters were tested. The resulting plots are stored in the Plots folder for each hyperparameter configuration:
In Summary folder are stored some statistics for each parameters configuration. The formatting of the files is described in analysis.
Here are final results!
Agent playing before training:
Agent playing after training:
To implement this code we started from the one provided by the course tutor Emanuele Panizon. Moreover, chatGPT was used in the realisation of some small parts of the code, and this is noted accordingly in the code.