The benefit of off-policy learning is fundamentally limited by the fact that data from ineffective early exploration isn't that useful for improving on later more refined policies. It's clear if you think of a few examples: chess blunders, spasmodic movement, or failing to solve a puzzle. This becomes especially clear once you realize that data only becomes off-policy when it describes something the policy would not do. I think the solution to this problem is (unfortunately) related to the need for better generalization / sample efficiency.

While I like the blogpost, I think the use of unexplained acronyms undermines the opportunity of this blogpost to be useful to the wider audience. Small nit: make sure acronyms and jargon is explained.

the magic thing about off-policy techniques such as Q-Learning is that they will converge on an optimal result even if they only ever see sub-optimal training data.

For example, you can use a dataset of chess games from agents that move totally randomly (with no strategy at all) and use that as an input for Q-Learning, and it will still converge on an optimal policy (albeit more slowly than if you had more high-quality inputs)

The article mentions AlphaGo/Mu/Zero was not based on Q-Learning - I'm no expert but I thought AlphaGo was based on DeepMind's "Deep Q-Learning"? Is that not right?

This blog post is unfortunately missing what I consider the bigger reason why Q learning is not scalable:

As horizon increases, the number of possible states (usually) increases exponentially. This means you require exponentially increasing data to have a hope of training a Q that can handle those states.

This is less of an issue for on policy learning, because only near policy states are important, and on policy learning explicitly only samples those states. So even though there are exponential possible states your training data is laser focused on the important ones.

Q learning is great as a hello world RL project for teaching undergraduates.

Humans actually do both. We learn from on-policy by exploring consequences of our own behavior. But we also learn off-policy, say from expert demonstrations (but difference being we can tell good behaviors from bad, and learn from a filtered list of what we consider as good behaviors). In most, off-policy RL, a lot of behaviors are bad and yet they get into the training set and hence leading to slower training.

The stated problem is getting Off-policy RL to work, aka discover a policy smarter than the one it was shown in its dataset.

If I understand correctly, they show random play, and expect perfect play to emerge from the naive Q-learning training objective.

In layman's term, they expect the algorithm to observe random smashing of keys on a piano, and produce a full-fledge symphony.

The main reason it doesn't work is because it's fundamentally some Out Of Distribution training.

Neural networks works best in interpolation mode. When you get into Out Of Distribution mode, aka extrapolation mode, you rely on some additional regularization.

One such regularization you can add, is to trying to predict the next observations, and build an internal model whose features help make the decision for the next action. Other regularization may be to unroll in your head multiple actions in a row and use the prediction as a training signal. But all these strategies are no longer the domain of the "model-free" RL they are trying to do.

Other regularization, can be making the decision function more smooth, often by reducing the number of parameters (which goes against the idea of scaling).

The adage is "no plan survive first contact with the enemy". There needs to be some form of exploration. You must somehow learn about the areas of the environment where you need to operate. Without interaction from the environment, one way to do this is to "grok" a simple model of the environment (fitting perfectly all observation (by searching for it) so as to build a perfect simulator), and learn on-policy from this simulation.

Alternatively if you have already some not so bad demonstrations in your training dataset, you can get it to work a little better than the policy of the dataset, and that's why it seems promising but it's really not because it's just relying of all the various facets of the complexity already present in the dataset.

If you allow some iterative gathering phase of information from the environment, interlaced with some off-policy training, it's the well known domain of Bayesian methods to allow efficient exploration of the space like "kriging", "gaussian process regression", multi-arm bandits and "energy-based modeling", which allow you to trade more compute for sample efficiency.

The principle being you try model what you know and don't know about the environment. There is a trade-off between the uncertainty that you have because you have not explored the area of space yet and the uncertainty because the model don't fit the observation perfectly yet. You force yourself to explore unknown area so as not to have regrets (Thomson Sampling) ) but still sample promising regions of the space.

In contrast to on-policy learning, the "bayesian exploration learning" learn in an off-policy fashion all possible policies. Your robot doesn't only learn to go from A to B in the fastest way. Instead it explicitly tries to learn various locomotion policies, like trotting or galloping, and other gaits and use them to go from A to B, but spend more time perfecting galloping as it seems that galloping is faster than trotting.

Possibly you can also learn adaptive strategy like they do in sim-to-real experiments where your learned policy is based on unknown parameters like how much weight your robot carry, and your learned policy will estimate on-the-fly these unknown parameters to become more robust (aka filling in the missing parameters to let the optimal "Model Predictive Control" work).

Completely agree and think it’s a great summary. To summarize very succinctly; you’re chasing a moving target where the target changes based on how you move. There’s no ground truth to zero in on in value-based RL. You minimise a difference in which both sides of the equation have your APPROXIMATION in them.

I don’t think it’s hopeless though, I actually think RL is very close to working because what it lacked this whole time was a reliable world model/forward dynamics function (because then you don’t have to explore, you can plan). And now we’ve got that.

How can we eliminate Q-learning’s bias in long-horizon, off-policy tasks?....

This post mischaracterizes AlphaZero/MuZero.

AlphaZero/MuZero are not model based, and aren't on policy either. They train at a significantly higher temperature, and thus intentional suboptimal play, than they use when playing to win. LeelaChessZero has further improvements to reduce the bias on training on suboptimal play.

There's a well known tradeoff in TD learning based on how many steps ahead you look- 1-step TD converges off policy, but can give you total nonsense/high bias when your Q function isn't trained. Many-step can't give you nonsense because it scores based on the real result, but that real result came from suboptimal play so your own off-policyness biases the results, plus it's higher variance. It's not hard to adjust between these two in AlphaZero training as you progress to minimize overall bias/variance. (As in, AlphaZero can easily do this- I'm not saying the tuning of the schedule of how to do it is easy!)

Probably a dumb/obvious question: if the bias comes from selecting the Q-maximum, can't this simply be replaced by sampling from a PDF?

The thing that worked for them was reducing the horizon. From my limited and dated understanding, I thought that's what the gamma term was for: exponentially you discount the value of stuff in the future to the point of being negligible (or less than the epsilon of differences you can even represent). So ... why/when is exponential discounting not enough?

I guess it's worth pointing that when humans learn the long-horizon tasks that we learn by repetitive training, we segment them in tasks with a shorter horizon and compose them hierarchically later.