I think that you have produced a very well organised code but, even if the README file it's very helpful in making your code more understandable, more comments in the code itself would have been very beneficial to the overall lab. Sometimes the intention behind some functions is not very clear so providing the reasoning may have helped.
Strategies
I saw you implemented a very similar idea to the one that the professor used in class, and maybe you could have tried some other methods of Reinforcement Learning (For example Q-learning). An other idea that you could have explored is trying to reduce the state space implementing a way to rotate and reflect a state, and in this way, reduce the number of Montecarlo simulation need in training. In the other hand I think you did a very clever job implementing the benchmark class that allowed you to create a very clear picture of what are the capabilities of the agent that you trained.
Results
The results that you obtained are very good; the agent you trained is able to win in most of the situation and in the worst case draw (which it has to be the worst result), never losing once against any strategy. I think a good improvement could be to also make the agent play as the second player, creating a more complete picture of how much effective this method is. In conclusion I think the overall lab is very well done and I wish you good luck for the exam.
Code Readability
I think that you have produced a very well organised code but, even if the README file it's very helpful in making your code more understandable, more comments in the code itself would have been very beneficial to the overall lab. Sometimes the intention behind some functions is not very clear so providing the reasoning may have helped.
Strategies
I saw you implemented a very similar idea to the one that the professor used in class, and maybe you could have tried some other methods of Reinforcement Learning (For example Q-learning). An other idea that you could have explored is trying to reduce the state space implementing a way to rotate and reflect a state, and in this way, reduce the number of Montecarlo simulation need in training. In the other hand I think you did a very clever job implementing the benchmark class that allowed you to create a very clear picture of what are the capabilities of the agent that you trained.
Results
The results that you obtained are very good; the agent you trained is able to win in most of the situation and in the worst case draw (which it has to be the worst result), never losing once against any strategy. I think a good improvement could be to also make the agent play as the second player, creating a more complete picture of how much effective this method is. In conclusion I think the overall lab is very well done and I wish you good luck for the exam.