Jens Tuyls
@JensTuyls
Followers
783
Following
799
Media
5
Statuses
99
PhD @PrincetonCS . Previously CS & Eng. @UCIrvine . Studying AI, ML, RL, #NLProc .
Silicon Valley, CA
Joined June 2016
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Oprah
• 368291 Tweets
Gus Walz
• 339577 Tweets
الرائد
• 284785 Tweets
Marçal
• 158519 Tweets
Macri
• 122914 Tweets
sabrina
• 112562 Tweets
Beyoncé
• 87451 Tweets
Barron
• 86314 Tweets
#النصر_الرايد
• 82276 Tweets
Ann Coulter
• 73779 Tweets
Happy Anniversary
• 69779 Tweets
Boulos
• 68923 Tweets
MUERO X DECÍRTELO OUT NOW
• 52422 Tweets
رونالدو
• 46204 Tweets
George Bush
• 34325 Tweets
Servette
• 30913 Tweets
Mudryk
• 28996 Tweets
Dotson
• 28333 Tweets
Mike Pence
• 24934 Tweets
CARAJO FURIOSO
• 24487 Tweets
Guiu
• 20089 Tweets
#LUGvBJK
• 19731 Tweets
Dinesh
• 18670 Tweets
Ken Salazar
• 18282 Tweets
Noni
• 16614 Tweets
Jimmy Carter
• 12336 Tweets
対象の3作品
• 12270 Tweets
Masuaku
• 12187 Tweets
エンティーム
• 11890 Tweets
#ابعدوا_رايد_اسماعيل_وغويدو
• 11756 Tweets
Mendy
• 10264 Tweets
Pinned Tweet
Imitation learning is one of the most widely used methods in ML, but how does compute affect its performance?
We explore this question in the challenging game of NetHack and find our scaled-up agent to outperform prior SOTA by 2x!
[1/6]
2
19
107
How can RL agents deal with both sparse rewards and large, dynamic action spaces – a key challenge in text games?
Our method eXploit-Then-eXplore (XTX) tackles these challenges and achieves a more than 2x improvement on Zork!
#ICLR2022
Spotlight
📜[1/5]
6
11
44
I’ll be at
@NeurIPSConf
this week! Feel free to reach out if you’d like to chat about anything scale in RL/IL, language agents (or broadly RL + NLP), or game theory!
0
0
17
0
0
8
See all of this and more in:
Scaling Laws for Imitation Learning in NetHack
by
@JensTuyls
,
@DhruvMadeka
, Kari Torkkola, Dean Foster,
@karthik_r_n
,
@ShamKakade6
Paper:
Project page: coming soon!
0
0
7
More broadly, our results call for work in the larger IL and RL community to more carefully consider the role of scaling laws, which could provide large improvements in many other domains. Also check out prior work by
@openai
: .
[5/6]
1
0
6
We train a suite of neural NetHack agents with different model sizes using Behavioral Cloning (BC) and analyze the loss and mean return isoFLOP profiles. We find both BC loss and mean return to follow clear power law trends with respect to FLOPs.
[3/6]
1
0
6
Using these power laws, we forecast the model and data size needed to train an agent aimed at recovering the underlying expert. While our agent falls short of expert performance, it sets a new SOTA (2.7K) in the unsolved game of NetHack, surpassing the prior best by 2x!
[4/6]
1
0
5
Prior works have found IL to consistently underperform the data-generating policy. However, these works often overlook the role of compute in terms of model and data size. Inspired by work around LLMs, we see if scaling up IL can provide similar performance gains.
[2/6]
1
0
5
See all of this and more in:
Multi-Stage Episodic Control for Strategic Exploration in Text Games
By
@JensTuyls
,
@ShunyuYao12
,
@ShamKakade6
,
@karthik_r_n
Paper:
Project page:
Code:
2
1
4
0
0
4
XTX outperforms several competitive baselines across 12 games in the Jericho benchmark (avg norm. scores across games in fig) in both the deterministic and stochastic setting, showing the strength of our multi-stage approach with strategic exploration at the frontier.
[4/5]
1
0
2
XTX employs a two-stage rollout in each episode to tackle these:
(1) An *exploitation* policy trained on promising past trajectories returns to the frontier.
(2) An *exploration* policy that uses past experience and curiosity explores the frontier.
[3/5]
1
0
2
Text games present unique challenges:
(1) *Sparse rewards:* agents need to quickly learn from only a few rewarding trajectories.
(2) *Large, dynamic action spaces* of up to 50 actions which can differ across states (e.g. “Echo” in fig), requiring clever exploration.
[2/5]
1
0
1
1
2
1