Deep RL Bootcamp Lecture 2: Sampling-based Approximations and Function Fitting

[0-10] So...with Q-learning we have made two big assumptions: 1) the underlying dynamics that give us P(s' | s, a), and 2) the states and actions can't be too large because we need to store the probabilities to be able to look them up. We may not always be able to meet these assumptions, so enter sampling-based approximation (starting with tabular Q-learning), where we sample states and keep a running average as our Q-value for a given (s, a).

[10-20] Diving into model-free methods. I LOVED that he used the term "temperature" to describe the shift in weighting that we see in something like a simulated annealing algorithm. I'm starting to see that a lot in RL (especially with the discount function), and I like this somatic layer of temperature. In this case, we are seeking to balance explore-exploit with an \epsilon-greedy algorithm.

[20-30] A brief introduction to on-policy learning (where sampling is done based on a policy, and those samples are used to update that policy) vs. off-policy learning where we sample based on one policy, but are ultimately updating (learning about) another policy. The requirements for q-learning are also given, which are: all states and actions are visited infinitely often, learning rate sum equals infinity as we approach infinity and the sum of squares of the learning rate is defined as we approach infinity. Exploration is a core RL challenge, and there may be situations where even epsilon-greedy doesn't converge in a reasonable amount of time.

[30-40] As long as there is an optimal path pursued towards a rewarded state, we won't need to propagate negative rewards back in time (this was shown in a gridworld example). At some point we converge to the Q-values of each state, so additional iterations won't change their values.

[40-50] Briefly touched upon **Temporal Difference Learning** as a way to take samples and not just expected values from the policy evaluation. But the main purpose of this section was drive the story towards deep RL, where we will use neural networks to to approximate Q values without needing to hand-engineer features. To step in this direction, was saw how tabular Q-learning is a special case of approximate Q-learning.