The Decision Service is a first-in-the-world project making tractable reinforcement learning easily used by developers everywhere. We are hiring for devel opers, data scientist, and a product manager. Please consider joining us to do something interesting this life
At EWRL, I particularly liked the talks from:
- Remi Munos on off-policy evaluation
- Mohammad Ghavamzadeh on learning safe policies
- Emma Brunskill on optimizing biased-but safe estimators (sense a theme?)
- Sergey Levine on low sample complexity applications of RL in robotics.
My talk is here. Overall, this was a well organized workshop with diverse and interesting subjects, with the only caveat being that they had to limit registration
At NIPS itself, I found the poster sessions fairly interesting.
- Allen-Zhu and Hazan had a new notion of a reduction (video).
- Zhao, Poupart, and Gordon had a new way to learn Sum-Product Networks
- Ho, Littman, MacGlashan, Cushman, and Austerwell, had a paper on how “Showing” is different from “Doing”.
- Toulis and Parkes had a paper on estimation of long term causal effects.
- Rae, Hunt, Danihelka, Harley, Senior, Wayne, Graves, and Lillicrap had a paper on large memories with neural networks.
- Hardt, Price, and Srebro, had a paper on Equal Opportunity in ML.
Format-wise, I thought the 2 sessions was better than 1, but I really would have preferred more. The recorded spotlights are also pretty cool.
The NIPS workshops were great, although I was somewhat reminded of kindergarten soccer in terms of lopsided attendance. This may be inevitable given how hot the field is, but I think it’s important for individual researchers to remember that:
- There are many important directions of research.
- You personally have a much higher chance of doing something interesting if everyone else is not doing it also.
During the workshops, I learned about ADAM (a momentum form of Adagrad), testing ML systems, and that even TenserFlow is finally looking into synchronous updates for parallel learning (allreduce is the way).
(edit: added one)
Reinforcement learning is much discussed these days with successes like AlphaGo. Wouldn’t it be great if Reinforcement Learning algorithms could easily be used to solve all reinforcement learning problems? But there is a well-known problem: It’s very easy to create natural RL problems for which all standard RL algorithms (epsilon-greedy Q-learning, SARSA, etc…) fail catastrophically. That’s a serious limitation which both inspires research and which I suspect many people need to learn the hard way.
Removing the credit assignment problem from reinforcement learning yields the Contextual Bandit setting which we know is generically solvable in the same manner as common supervised learning problems. I know of about a half-dozen real-world successful contextual bandit applications typically requiring the cooperation of engineers and deeply knowledgeable data scientists.
Can we make this dramatically easier? We need a system that explores over appropriate choices with logging of features, actions, probabilities of actions, and outcomes. These must then be fed into an appropriate learning algorithm which trains a policy and then deploys the policy at the point of decision. Naturally, this is what we’ve done and now it can be used by anyone. This drops the barrier to use down to: “Do you have permissions? And do you have a reasonable idea of what a good feature is?”
A key foundational idea is Multiworld Testing: the capability to evaluate large numbers of policies mapping features to action in a manner exponentially more efficient than standard A/B testing. This is used pervasively in the Contextual Bandit literature and you can see it in action for the system we’ve made at Microsoft Research. The key design principles are:
- Contextual Bandits. Many people have tried to create online learning system that do not take into account the biasing effects of decisions. These fail near-universally. For example they might be very good at predicting what was shown (and hence clicked on) rather that what should be shown to generate the most interest.
- Data Lifecycle support. This system supports the entire process of data collection, joining, learning, and deployment. Doing this eliminates many stupid-but-killer bugs that I’ve seen in practice.
- Modularity. The system decomposes into pieces: exploration library, client library, online learner, join server, etc… because I’ve seen to many cases where the pieces are useful but the system is not.
- Reproducibility. Everything is logged in a fashion which makes online behavior offline reproducible. Consequently, the system is debuggable and hence improvable.
The system we’ve created is open source with system components in mwt-ds and the core learning algorithms in Vowpal Wabbit. If you use everything it enables a fully automatic causally sound learning loop for contextual control of a small number of actions. This is strongly scalable, for example a version of this is in use for personalized news on MSN. It can be either low-latency (with a client side library) or cross platform (with a JSON REST web interface). Advanced exploration algorithms are available to enable better exploration strategies than simple epsilon-greedy baselines. The system autodeploys into a chosen Azure account with a baseline cost of about $0.20/hour. The autodeployment takes a few minutes after which you can test or use the system as desired.
This system is open source and there are many ways for people to help if they are interested. For example, support for the client-side library in more languages, support of other learning algorithms & systems, better documentation, etc… are all obviously useful.
However, some of the discussion around this seems like giddy overstatement. Wired says Machines have conquered the last games and Slashdot says We know now that we don’t need any big new breakthroughs to get to true AI. The truth is nowhere close.
For Go itself, it’s been well-known for a decade that Monte Carlo tree search (i.e. valuation by assuming randomized playout) is unusually effective in Go. Given this, it’s unclear that the AlphaGo algorithm extends to other board games where MCTS does not work so well. Maybe? It will be interesting to see.
Delving into existing computer games, the Atari results (see figure 3) are very fun but obviously unimpressive on about ¼ of the games. My hypothesis for why is that their solution does only local (epsilon-greedy style) exploration rather than global exploration so they can only learn policies addressing either very short credit assignment problems or with greedily accessible polices. Global exploration strategies are known to result in exponentially more efficient strategies in general for deterministic decision process(1993), Markov Decision Processes (1998), and for MDPs without modeling (2006).
The reason these strategies are not used is because they are based on tabular learning rather than function fitting. That’s why I shifted to Contextual Bandit research after the 2006 paper. We’ve learned quite a bit there, enough to start tackling a Contextual Deterministic Decision Process, but that solution is still far from practical. Addressing global exploration effectively is only one of the significant challenges between what is well known now and what needs to be addressed for what I would consider a real AI.
This is generally understood by people working on these techniques but seems to be getting lost in translation to public news reports. That’s dangerous because it leads to disappointment. The field will be better off without an overpromise/bust cycle so I would encourage people to keep and inform a balanced view of successes and their extent. Mastering Go is a great accomplishment, but it is quite far from everything.
I had a chance to attend UAI this year, where several papers interested me, including:
- Hoifung Poon and Pedro Domingos Sum-Product Networks: A New Deep Architecture. We’ve already discussed this one, but in a nutshell, they identify a large class of efficiently normalizable distributions and do learning with it.
- Yao-Liang Yu and Dale Schuurmans, Rank/norm regularization with closed-form solutions: Application to subspace clustering. This paper is about matrices, and in particular they prove that certain matrices are the solution of matrix optimizations. I’m not matrix inclined enough to fully appreciate this one, but I believe many others may be, and anytime closed form solutions come into play, you get 2 order of magnitude speedups, as they show experimentally.
- Laurent Charlin, Richard Zemel and Craig Boutilier, A Framework for Optimizing Paper Matching. This is about what works in matching papers to reviewers, as has been tested at several previous NIPS. We are looking into using this system for ICML 2012.
In addition I wanted to comment on Karl Friston‘s invited talk. At the outset, he made a claim that seems outlandish to me: The way the brain works is to minimize surprise as measured by a probabilistic model. The majority of the talk was not actually about this—instead it was about how probabilistic models can plausibly do things that you might not have thought possible, such as birdsong. Nevertheless, I think several of us in the room ended up stuck on the claim in questions afterward.
My personal belief is that world modeling (probabilistic or not) is a useful subroutine for intelligence, but it could not possibly be the entirety of intelligence. A key reason for this is the bandwidth of our senses—we simply take in too much information to model everything with equal attention. It seems critical for the efficient functioning of intelligence that only things which might plausibly matter are modeled, and only to the degree that matters. In other words, I do not model the precise placement of items on my desk, or even the precise content of my desk, because these details simply do not matter.
This argument can be made in another way. Suppose for the moment that all the brain does is probabilistic modeling. Then, the primary notion of failure to model is “surprise”, which is low probability events occurring. Surprises (stumbles, car wrecks, and other accidents) certainly can be unpleasant, but this could be correct if modeling is a subroutine as well. The clincher is that there are many unpleasant things which are not surprises, including keeping your head under water, fasting, and self-inflicted wounds.
Accounting for the unpleasantness of these events requires more than probabilistic modeling. In other words, it requires rewards, which is why reinforcement learning is important. As a byproduct, rewards also naturally create a focus of attention, addressing the computational efficiency issue. Believing that intelligence is just probabilistic modeling is another example of simple wrong answer.