Itai Bear, 325839710, itai.bear1@gmail.com
Noam Manaker Morag, 214726002, noam.school.only@gmail.com
We took off our learning period by going over the RL Course by David Silver: https://www.davidsilver.uk/teaching/
We also started the Deep Reinforcement Learning course by Huggingface, which had practical programming assignments: https://huggingface.co/deep-rl-course/unit0/introduction?fw=pt
We finished the course by David Silver.
Due to the exam period, we weren't able to finish the Huggingface course, but we are on our way to doing so.
We finished the Hugginface course and began to search for what game to solve. We are searching for a game that is sufficiently complex to be interesting and also has a gym environment.
Found an existing Tetris emulator with a prepared gym-environment. Began training on a Deep Q Learning model to play the game.
Tried using some simple reward functions:
- In-Game score
- Number of lines cleared
- Penalization for board height But none of them yielded significant results.
Began using Optuna, a library for hyper-parameter optimization. Optuna tries many different combinations of hyper-parameters and selects the best one.
Tried more advanced reward functions:
- Penalization for empty cells
- "Bumpiness"
- Different combinations of the different rewards
Unfortunately, none of these worked either.
We began searching for different ideas to improve the model.