I had access to udemy courses back when I was working at Deloitte Digital. One interesting course I did was the introduction to deep q learning from Pieran Data. Deep Q learning is a way to use neural nets to do reinforcement learning. It was an interesting course and I learned a lot, so I’ll give a recap here.
Classical Q learning is a reinforcement learning algorithm that revolves around making a table of all possible states and actions, and gives an expected reward for each of these combinations. For problems with a continuous state or action space, the states and choices are discretized by bucketing (and bucket distribution is a hyperparameter).
The table is made iteratively. At first it’s filled with zeroes, as no rewards are known, and choices are made at random. Expected reward values are updated using the Bellman Equation. As expected reward values are updated, the randomness of action choices is reduced to fine-tune the reward values (this is called the Epsilon Greedy Strategy).
Classic Q learning is great for a certain class of reinforcement learning problems. Here’s an example we did in the course of classic q learning beating the “hill car challenge.”:
This works great but most realistic problems have too large a space of states and actions, even after bucketing. The key insight of deep Q learning is that similar states will lead to similar actions, so we can model the states and actions using neural networks to make it a tractable problem. It’s been almost a year since I covered this materal , and I feel pretty rusty so I won’t get too deep into it, but here’s an example we did in the course of getting an AI that can play Atari games:
Overall it’s a very cool technique but I doubt I’ll ever get much use out of it.