A few questions on the project: 1. Do we get more marks for more complex agents? Or does it not make a difference as long as we implement and analyse them properly? It seems to me that a reinforcement learning agent would take a lot more in terms of code and understanding compared to a rule-based agent. 2. Is it actually feasible to make a normal Q-learning agent that uses a Q-table for this? The state space of the display is massive. I am attempting to simplify the contents of the display instead of using all the pixels but working based on the full dimensions of the screen still looks to be too big. I was thinking of limiting Mario's vision further, but I just want to know if this has been implemented successfully before, since everything I've read online just tells me Q-tables are no good for mario and to use DQN instead. 3. How much of the implementation can we leave to outside libraries? If, say, I end up using neural networks, the Keras library can do a lot of the heavy-lifting while you mainly just feed it data and run training (I think?). It seems to simplify things a bit too much. Tell me if I have any misconceptions about the agents, I know we have a reinforcement learning lecture further down the line but we have a basic rule-based agent going so I want to start planning for the second.