A summary of Sergey's understanding of Bottlenecks, tips, and tricks to resolve them.
approximation error is not actually a major problem
Sergey managed to deduce –with low dimensional "oracle based" unit tests like mazes– a series of empirical benchmarks that did not previously exists. By using pre-computed answers from a "master" training set and gradually swapping out master examples for actual training, Sergey and co were able to measure which aspects of DQNs get stuck, and document methods & techniques that contribute to fixing them.
Q-learning will converge (depending on the Q values being tabular, linear, etc.). If it does not, it is implied that the action space is employing approximation operators, as infinite action spaces cannot be tabulated. Even when does not converge, divergence only occurs with a 0.9% probability and can be remediated with early-stop protocols.
returns an evaluative quantity given a state + action. Whereas standard is not necessarily good for all .
Big Neural Networks do it Better.
Imitation learning constrains the agent to being good at strictly what it observed from the experts. If the agent where to find itself out of the narrowly defined 'expertly good' action space, it would have no idea what to do.
Replay Buffer - stores all tuples and is used to track how it got good. However, step-wise action spaces with disparate environments per step (or moving targets) make it difficult for the agent to act consistently. By sampling the replay buffer, we can average previous actions to find, not necessarily , but . This is important as at state might be drastically different than at state .
Temporal Difference Error - Roughly measures the "amount of surprise at reward" and is used in weighting the replay buffer distribution sampling method. Broader entropy distributions are almost always better than any other distribution-sampling techniques.
The Bellman Backup - An operator you pass to your function which returns a better based on predictions from futures . Bellman backups focus on a single policy , rather than a branching structure. Bellman Backups are used to refine the way an agent takes action since it is typically too expensive to sample and evaluate all possible actions.
essentially euclidean distance.
the maximum element such that
minimize and you're guaranteed to find
In continuous domains, is really hard to solve, so you make another Neural Network to observe the function: the critic.
Probably Approximately Correct (lol)
Sergey Levin is poor and therefore his network architecture sucks
Q-learning is prone to overfitting - Replay buffers are effective because they handle big picture evaluations and get the agent good at the broadest distribution of state spaces whereas will get you lost faster based on the current state. The solution: train on previous actions, actuate on current situations.
Larger architectures remove error bias whereas networks with like 6 nodes get derailed by one error.
Early Stopping - once increases, stop training
- use fewer steps on many samples
- don't use many steps on few samples otherwise your network could fall victim of oversampling
High Entropy Distribution Sampling
Makes sure can't game the system by getting familiar with easy samples from the replay buffer. As increases, we need to make sure we have a robust sampling process via High Entropy Distribution Sampling. The solution is creating another network to fight in a bellman curve weighting "fitness game" such that they both become really good and when the outcome of the game is highly unpredictable. Then we can map the graph of the outcome to the replay buffer sampling to ensure both random, and surprising samples such that we aren't surprised in production. High variance in outcome --> entropy --> good samples.
To do this is difficult in theory, so we add human interference. This process is akin to walking in on two people and giving them swords to accelerate the process. Just a very complicated overkill-method of generating random numbers to sample from the replay buffer.