I’ve fallen behind on RL literature from the past few months. So, I’ve decided to catch up with a bunch of recent papers.
First Return Then Explore
Let’s start with First Return Then Explore, by Ecoffet et al. This is a continuation and extension of the Go-Explore work from UberAI.
When Go-Explore first came out, I was very excited by its announced results, but got upset by how they were presented. I wrote a post attempting to explain that tension - that I really liked the paper’s ideas, and really disliked its media strategy. The media strategy for First Return Then Explore is comparatively muted. For one, this time they have a draft on arXiv. (Sorry, I’m never going to stop ribbing them for that.) They’ve also been more careful in their claims, and have improved their previous results.
Both First Return Then Explore and Go-Explore aim to first return to a state that has been visited before, then explore from that state. To make this more efficient, states are grouped into “cells” through some encoding. In the original Go-Explore paper, these cells are defined by downsampling by a fixed factor. First Return Then Explore changes this to tune the downsampling factor online, by doing a small search to maximize normalized entropy across a fixed budget of cells. There are also more heuristics on choosing which cell to return to, instead of uniformly at random.
Besides this change, the Atari experiments mostly operate the same way: they assume a simulator or deterministic environment, learn the policy by leveraging the determinism, then do a robustification step where they try to reproduce behavior in a stochastic version of the environment.
The part I care about is the part they call Policy-based Go-Explore. My main criticism of the original Go-Explore paper was that it required access to a deterministic analogue of your final environment. They proposed learning a goal-conditioned policy to return to previous states, instead of following a memorized trajectory, which lets you hand stochastic environments at training time. However, they left it as future work.
Well, now they have results. It worked, but it was only tested on Montezuma’s Revenge with domain-specific features. I view papers through survival bias: if there’s an experiment that’s natural in the paper’s context, but isn’t in the paper, then it probably didn’t work, because if it worked, it’d be in the paper. So for now, I’m assuming it didn’t beat SOTA with domain agnostic features.
My final verdict is that the updated paper improved its strengths, but only mildly improved its weaknesses. The paper is an even stronger case that good exploration can be reduced to learning to quickly return to states you’ve visited before, and exploration algorithms without this capability have failure modes that First Return Then Explore fixes. Learning that return policy, however, is still an open problem for general domains. The reduction is valuable, and I hope it encourages more work on efficiently learning goal-conditioned policies.
The new hotness in RL is data augmentation. Three papers came out on arXiv in the past week: Constrastive Unsupervised Reinforcement Learning (CURL), from Srinivas and Laskin et al, Image Augmentation is All You Need (DrQ) from Kostrikov and Yarats et al, and Reinforcement Learning with Augmented Data (RAD) from Laskin and Lee et al. It also made it to VentureBeat of all places.
These three papers all find that for image-based RL, data augmentation gives very large gains on several tasks. Now at this point, I should mention that CURL and RAD are from people I know from UC Berkeley, and DrQ is from people I know from Google, so I’m going to step very carefully…
CURL learns a representation by contrastive learning. Two randomly sampled data augmentations are applied to the same image, and their representations are encouraged to be close to one another through an InfoNCE loss. (See the SimCLR paper for an ablation showing this contrastive loss does better than other ones.)
RAD compares just using data augmentation, without any contrastive losses, and finds that it outperforms CURL on the DMControl Suite. The theory is that in these environments, RAD beats CURL because it only optimizes for the task reward we care about, while CURL has to balance RL and contrastive learning. An ablation of the data augmentations used finds that random cropping is by far the most important data augmentation.
DrQ also does data augmentation, using random shifts. This is the same as padding the image, then doing a random crop. In an actor-critic framework, they sample data augmentations to estimate , sample other data augmentations to estimate target Q-value , and do a critic update that’s now regularized by the data augmentation.
Now, are these results surprising? Uh, kind of? It isn’t surprising because data augmentation isn’t new. Specifically doing random cropping isn’t new either - the QT-Opt paper I worked on 2 years ago used random cropping. Other groups have used data augmentation as well. The surprising part is the effect size. These papers are the first to carefully design an experimental setup that lets them isolate and measure the gains from data augmentation.
It’s the sort of paper that makes you feel dumb you didn’t write it yourself. I’ve run very similar data augmentation ablations before, with results that were consistent to theirs, but I never did it on standard RL benchmarks and I never dug into it more. If I had, I probably could have written this paper. Ah well, live and learn.
I’m very big on data augmentation. It just seems like the obvious thing to do. You can either view it as multiplying the size of your dataset by a constant factor, or you can view it as decreasing the probability your model learns a spurious correlation, but in either case it usually doesn’t hurt and it often really helps.
Salesforce put out a paper that uses reinforcement learning to design tax policy in a toy economic environment, and they argue their tax policies give better equality-productivity trade-offs, compared to the Saez framework.
I do not understand tax policy very well, but my first instinct is that the economy is really complicated, a model of the economy has to be too simplistic somewhere, and therefore the results should be taken with massive caveats. The authors are aware of this, and the ideas the paper plays with are interesting. I’ve found papers like this are best viewed as idea generators. Within a model, the AI discovers a new strategy, which could be useful in the more complex environment, but you will get better results by asking a human to consider whether the AI’s strategy makes sense, instead of applying the AI’s strategy directly.
Within the simulated economy, the agent preferred higher tax rates for the top brackets and lower tax rates for the middle class. So that’s interesting.
It’s very unlikely this makes it to actual tax policy anytime soon. The real economy is more complicated, the politics is a nightmare to navigate, and the people in charge of economic policy probably care more about the perception of a good economy than the reality of a good economy. Given the ethics questions surrounding economics experiments, perhaps that’s for the best.
Offline Reinforcement Learning
Some colleagues from Google Brain and UC Berkeley have put a tutorial for Offline Reinforcement Learning on arXiv.
By offline reinforcement learning, they mean reinforcement learning from a fixed dataset of episodes from an environment, without doing any additional online data collection during learning. This is to distinguish it from off-policy learning, which can happen in an offline setting, but is commonly used in settings with frequent online data collection.
Offline RL is, in my opinion, a criminally understudied subject. It’s both very important and very difficult, and I’ve been talking about writing a blog post about it for over a year. Suffice it to say that I think this tutorial is worth reading. Even if you do not plan to research offline RL, I feel the arguments for why it’s important and why it’s hard are useful to understand, even if you disagree with them.