Index | Archive | Tags | Atom Feed | RSS Feed

Planning in Stochastic Environments with a Learned Model

After extending to arbitrary action spaces and offline RL, we recently published our next step in generalizing MuZero: [cached]Stochastic MuZero, which learns and plans with a stochastic model of the environment that can model not only a single deterministic outcome of our actions, but rather a full set of possibilities.

Previous approaches such as AlphaZero and MuZero have achieved very good results in fully-observed deterministic environments, but this type of environment is insufficiently general to model many problems and real world tasks. Beyond the stochasticity inherent in many domains (think roll of the dice or draw of the cards …


Our lives are measured in memorable moments

The reason our childhood years seemed to pass so much more slowly than our adulthood is that they were filled with novelty and excitement, memorable first experiences: learning to ride a bike, entering a new school, making new friends.

Events, experiences, things are memorable not by some inherent quality, but rather become memorable by virtue of being different from our everyday life. What matters is the magnitude, not the sign: both the best and the worst times in your life will stick out in your memory.

Memorable experiences come in all sizes: cooking a new dish, meeting up with old …


MuZero does YouTube

If you've been watching YouTube lately, the encoding settings for the video you watched might have been selected by MuZero - using less of your bandwidth for the same quality. The task here is rate control: selecting the quantization parameters within the VP9 codec to maximize the quality at a specified bitrate.

This is a constrained RL problem, requiring us to optimize two conflicting objectives of variable difficulty at the same time. To deal with this challenge we introduce a self-competition based reward mechanism, where the reward depends on how successful other recent episodes were at maximizing the quality while staying …


Competitive Programming with AlphaCode

I'm excited to share more on our latest project, AlphaCode - a system to write programs, able to reach roughly median human performance in coding competitions. And that's median competitive programmer performance, not median programmer or median human! In addition to our paper and [cached]official blog post, you can also find my personal take below.

Problem Setup

Coding competitions are difficult even for experienced programmers. Before writing the first character of a program, the first step is to understand the natural language problem description: often spanning several paragraphs in length, problem descriptions do not directly describe the required algorithm, but …


Mastering Atari Games with Limited Data

Another interesting paper based on MuZero was published at NeurIPS 2021: Mastering Atari Games with Limited Data, aka EfficientZero. This paper by Weirui Ye, Shaohuai Liu, Thanard Kurutach, Pieter Abbeel and Yang Gao focuses on the application of MuZero to very low data tasks, such as Atari 100k (only two hours of gameplay!) or DMControl 100k.

To tackle these tasks, the author propose three main techniques:

First they introduce a Self-Supervised Consistency Loss, to ensure that the embeddings produced by MuZero's dynamics function are consistent with the embeddings from the representation function. This loss is insipired by [cached]SimSiam-style …


Online and Offline Reinforcement Learning by Planning with a Learned Model

After extending to arbitrary action spaces, our next step in generalizing MuZero was to work on data efficiency, and to extend it all the way to the offline RL case. We've now published this work as MuZero Unplugged at NeurIPS, below I will give a brief summary of the main ideas.

Environment interactions are often expensive or have safety considerations, while existing datasets frequently already demonstrate good behaviour. We want to learn efficiently from any such data source, without being restricted by off-policy issues or limited to the performance of the policy that generated the data: as always, we want …


Values, Pointers and References in C++

If you've primarily used high level languages like Python, you may not be used to explicitly thinking about the ownership or representation of your values in memory.1 In system languages like C++ or Rust, we have direct control over these aspects, and are able to use the type system to explicitly represent when a function takes ownership of a value, vs when it only takes a (temporary) reference.2

First, different types of ownership, in order of preference:

  • T t. A normal owned value of type T, uniqlue owned. If declared as a variable it is stored on the …

Learning and Planning in Complex Action Spaces

We've been working on making MuZero as easy to use as possible, allowing it to be applied to any RL problem of interest. First and foremost, this required us to make some algorithmic extensions.

The original MuZero algorithm achieved good results in discrete action environments with a small action space - classical board games and Atari, with up to a few thousand different actions. However, many interesting problems have action spaces that cannot be easily enumerated, or are not discrete at all - just think of controlling a robot arm with many degrees of freedom, smoothly moving each joint. So far, MuZero …


Complexity Budget

No matter how smart, there's a limit to how much you can keep in mind at once before you start forgetting something. In the context of software engineering, a common term for this is complexity budget: the maximum complexity of a system you can understand at once.

Just as fitting an entire task into the cache of your CPU is orders of magnitude faster than paging in from RAM (or worse, disk!), fitting the entire problem you are working on into your brain can make you many times more productive.

Every line of code you write, every feature you add …


Rates of Growth

I've been enjoying [cached]A Map that Reflects the Territory, a collection of the best LessWrong essays from 2018. While in general the essays are thought provoking and interesting, one set of essays gave me pause: [cached]Hyperbolic Growth and related chapters on fast vs slow takeoff.

The discussion repeats tropes that are common in the rationalist and futurist community, describing how economic and technological [cached]growth have been accelerating and suggesting that they will soon increase so quickly that we will be unable to follow:

world gdp, growing exponentially ([cached]data source, csv)

However, we have to be careful not to mix up …

© Julian Schrittwieser. Built using Pelican. Theme by Giulio Fidente on github. .