Mastering Atari Games with Limited Data · HackerPapers discussions

Mastering Atari Games with Limited Data — mentioned by Isinlor

Fri, 19 Aug 2022 22:04:53 GMT

Have you heard about EfficientZero? This is the first algorithm that achieved super-human performance on Atari 100k actions benchmark. EfficientZero's performance is also close to DQN's performance at 200 million frames while we consume 500 times less data. DQN was published in 2013, EfficientZero in 2021. That's 8 years with 500 times improvement. So data efficiency was doubling roughly every year for the past 8 years. Side note: EfficientZero I think still may not be super-human on games like Montezuma's Revenge. https://arxiv.org/abs/2111.00210 Reinforcement learning has achieved great…

Paper and discussion history. A mention is not an endorsement.

Mastering Atari Games with Limited Data — mentioned by Isinlor

Sun, 06 Feb 2022 14:53:47 GMT

Roughly half/half, skewed more towards improvements in algorithms. Measuring the Algorithmic Efficiency of Neural Networks: https://arxiv.org/abs/2005.04305 In this work, we argue that algorithmic progress has an aspect that is both straightforward to measure and interesting: reductions over time in the compute needed to reach past capabilities. We show that the number of floating-point operations required to train a classifier to AlexNet-level performance on ImageNet has decreased by a factor of 44x between 2012 and 2019. This corresponds to algorithmic efficiency doubling every 16 months…

Paper and discussion history. A mention is not an endorsement.

Mastering Atari Games with Limited Data — mentioned by andreyk

Sat, 11 Dec 2021 22:22:34 GMT

I don't believe so, as it's more about competitive games and about incomplete information. They did release Agent 57 last year, and that got better than human results on all Atari games (though still not sample efficiently). More recently there was EfficientZero, an extension to MuZero, which is pretty cool as it is both sample efficient and quite performant, though I don't think it can do Montezuma. https://arxiv.org/abs/2111.00210

Paper and discussion history. A mention is not an endorsement.

Mastering Atari Games with Limited Data — HN submission

Sun, 07 Nov 2021 03:03:59 GMT

EfficientZero: Human ALE sample-efficiency with MuZero+self-supervised

4 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Mastering Atari Games with Limited Data — HN submission

Tue, 02 Nov 2021 03:29:15 GMT

Mastering Atari Games with Limited Data (EfficientZero)

5 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.