Profile picture
Eiso Kant @eisokant
, 9 tweets, 3 min read Read on Twitter
It's been 9 months since I became aware of the work @EntropyFarmer & @Miau_DB were doing in obscurity on tackling Artificial Intelligence.

This public release today sets new world records on many Atari games.

This thread tells part of the story:

github.com/FragileTheory/…
This release of FractalAI, shows a vastly more efficient and completely different way to do imagination based planning (deepmind.com/blog/agents-im…). Providing a highly efficient way to balance exploration of state space, of RAM or pixels, with exploitation.
The inspiration and foundation of this work couldn't have happened without @alexwg's paper on Causal Entropic Forces (alexwg.org/publications/P…)

FractalAI leverages the information contained in the entropy of any state space, together with any reward function.
It's important to note that this work allows efficient exploration of future states and their rewards, often using 1000x less samples than Monte-Carlo Tree Search.

But it does not do any learning, FractalAI purely asses the potential states through simulation.
Since FractalAI is able to set record scores on the Atari benchmark using a consumer laptop. It opens up an avenue of research for Reinforcement Learning hybrids such as AlphaGo to use the FractalAI module instead of MCTS.
Equally exciting is to treat traditional RL problems as supervised learning problems where FractalAI generates high quality samples.
FractalAI is just a small part that hopefully can contribute to an incredibly complex puzzle. What has amazed me, and humbled me to have had a small involvement, is that entire theory and experimental work was built from first principles by @EntropyFarmer & @Miau_DB
Both Sérgio and Guillem are purely self taught, with no background or experience in academia or industry. This meant that for years on end, people ignored their work and refused to take the time to understand it.
Now there is experimental evidence on a widely accepted benchmark, code to reproduce the experiments and GitHub to critique the work, the hope is that others in the industry will take the time to give constructive feedback. So please share this.
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to Eiso Kant
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member and get exclusive features!

Premium member ($3.00/month or $30.00/year)

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!