text stringlengths 521 4.24k | source stringclasses 1
value |
|---|---|
See discussions, st ats, and author pr ofiles f or this public ation at : https://www. researchgate. ne t/public ation/335158967 Is Deep Reinforcement Learning Really Su perhu man on Atari? Preprint · August 2019 CITATIONS 0READS 389 3 author s, including: Emilie Wirbel Valeo 23 PUBLICA TIONS 823 CITATIONS SEE P... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Is Deep Reinforcement Learning Really Superhuman on Atari? Leveling the playing field Marin Toromanoff MINES Paris Tech, Valeo DAR, Valeo. ai, name. surname@mines-paristech. fr name. surname@valeo. com Emilie Wirbel Valeo Driving Assistance Research name. surname@valeo. com Fabien Moutarde Center for Robotics, MINES Par... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
of previous work. Specifically, even though ALE is fast at runtime, training an agent on one game takes approximately one week on one GPU and thus the equivalent of more than one year to train on all 61 Atari games. A standardization of the evaluation procedure is needed to make DRL that matters as pointed out by Hender... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
C51 [3]. The idea behind PER is to sample transitions according to their surprise, i. e. the worse the network is at predicting the Q-value of a specific transition, the more we sample it. C51 is the first algorithm in Distributional RL which predicts the full distribution of the Q-function instead of predicting only the... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
linear to the allowed time: the reported score will be 6 times higher if capped at 30 minutes instead of 5 minutes. We argue that the time cap can make the performance comparison non significant. On many games (e. g. Atlantis, Video Pinball) the scores reported of Ape-X [13], Rainbow [12] and IQN [5] are almost exactly ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
of a random player and 100% is the score of the human baseline, which allows to summarize the performance on the whole Atari set in one number, instead of individually comparing raw scores for each of the 61 games. However we argue that this human baseline is far from being representative of the best human player, whic... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
are described in the supplementary material. Both our implementations of Rainbow and Rainbow-IQN are distributed3, following Ape-X [13] and based on the implementation of [4]. IQN is an evolution of the C51 algorithm [3] which is one of the 6 components of the full Rainbow, so this is a natural upgrade. After the imple... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Time 5 min 30 min No limit (SABER) Median Mean Super. Median Mean Super. Median Mean Super. Rainbow 2. 35% 14. 86% 0 2. 61% 17. 09% 1 2. 83% 24. 54% 3 Rainbow-IQN 2. 61% 17. 62% 0 2. 81% 20. 18% 1 3. 13% 30. 89% 4 Table 3: Evolution of performance with evaluation time (mean, median of normalized baseline and number of ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Figure 5: Median performance comparison for DQN, Rainbow and Rainbow-IQN with regards to training frames. Evaluation time is set at 5 minutes to allow a comparison to DQN. 5. 3 Stability of both Rainbow and Rainbow-IQN [15] use 5 different seeds for training to check that the results are reproducible and stable. For th... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
still small on the whole benchmark. We think that this reveals that both Rainbow and Rainbow-IQN are quite stable on Atari and strengthens our confidence on Rainbow-IQN being the new state-of-the-art on the Atari benchmark. In particular, Rainbow-IQN reaches infinite game time on Asteroids on all 5 trials whereas Rainbow... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
We also thank Gabriel de Marmiesse for his assistance to make the open-source implementation on Valeo AI Github4. Finally, most of computing resources were kindly provided by Centre de Calcul Recherche et Tech-nologie (CCRT) of CEA. References [1] Martn Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Martin Wicke, Y... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
[18] Open AI. Open AI Five. https://openai. com/five/, 2018. [19] Georg Ostrovski, Marc G Bellemare, A ¨aron van den Oord, and R ´emi Munos. Count-based exploration with neural density models. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2721-2730. JMLR. org, 2017. [20] Adam ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
we implemented a distributed version of Rainbow following the paper Distributed Prioritized Expe-rience Replay (Ape-X) [13]. Ape-X [13] is a distributed version of Prioritized Experience Replay (PER) but which can be adapted on any value-based RL algorithm including PER, e. g. Rainbow. There is no study of this in the ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
2 Nvidia V100 (in a remote supercomputer) 4 Nvidia Tesla V-100 (DGX station) 4 Nvidia Quadro M2000 (local workstations) B. 3 Rainbow-IQN Ape-X To ascertain our distributed implementation of Rainbow-IQN was functional, 3 experiments were conducted with multiple actors (10 actors instead of one). All locks and synchroniz... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
D Detailed experimental figures In this section, we provide more detailed versions of the figures in the main article. The structure of this section follows the one of Section 5 of the main article. As a reminder, all normalized world record baseline scoressare reported according to the following equation, where we note ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Figure 8: Performance comparison per game between the original Rainbow [12] versus Rainbow trained with [15] guidelines (30 minutes evaluation time to align with original conditions) Figure 9: Evolution of agents performance classification with evaluation time: Rainbow-IQN, 200M training frames, evaluation time ranging ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Figure 10: Performance comparison per game between Rainbow and Rainbow-IQN on SABER conditions (200M training frames) Figure 11: Rainbow-IQN normalized with regards to a Rainbow baseline for each game E Raw scores For verification purposes, we provide tables containing all relevant agent scores used to build the figures ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
infinite. They are marked as infinite gameplay in the table, and capped at 200% of the world record baseline for the mean computation. Evolution of scores with time Table 7 compares agents scores with increasing evaluation times for Rainbow and Rainbow-IQN, at 200M training frames. Evolution of scores with training frame... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Agent Category Game Name Random [16] World Record air raid 579. 25 NA 23050. 0 alien 211. 9 7127. 7 251916. 0 amidar 2. 34 1719. 5 104159. 0 assault 283. 5 742. 0 8647. 0 asterix 268. 5 8503. 3 1000000. 0 asteroids 1008. 6 47388. 7 10506650. 0 atlantis 22188. 0 29028. 1 10604840. 0 bank heist 14. 0 753. 1 82058. 0 batt... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Training frames Game name 10M 50M 100M 200M air raid 7765. 25 11690. 0 13434. 25 12289. 75 alien 2740. 6 1878. 1 5223. 0 7046. 4 amidar 347. 13 1554. 84 2129. 27 3092. 05 assault 966. 87 2783. 49 4443. 03 6372. 7 asterix 3467. 0 9280. 0 16344. 5 28015. 0 asteroids 1194. 16 (98. 25) 3261. 88 (2602. 48) Infinite gameplay ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Rainbow Rainbow-IQN Game name 5 minutes 30 minutes SABER 5 minutes 30 minutes SABER air raid 10549 12308. 25 12308. 25 11107. 25 12289. 75 12289. 75 alien 3458. 5 3458. 5 3458. 5 7046. 4 7046. 4 7046. 4 amidar 2835. 53 2952. 43 2952. 43 2601. 82 3092. 05 3092. 05 assault 3779. 98 3986. 1 3986. 1 5178. 41 6372. 7 6372. ... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Training frames Game name 10M 50M 100M 200M air raid 7549. 0 9168. 75 10272. 75 11107. 25 alien 2740. 6 1878. 1 5223. 0 7046. 4 amidar 347. 13 1554. 84 2129. 27 2601. 82 assault 966. 87 2783. 49 4103. 89 5178. 41 asterix 3467. 0 9280. 0 16344. 5 28015. 0 asteroids 1194. 16 (98. 25) 3251. 64 (2582. 07) 12261. 36 (12251.... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
Training frames Game name 10M 50M 100M 200M air raid 7765. 25 11690. 0 13434. 25 12289. 75 alien 2740. 6 1878. 1 5223. 0 7046. 4 amidar 347. 13 1554. 84 2129. 27 3092. 05 assault 966. 87 2783. 49 4443. 03 6372. 7 asterix 3467. 0 9280. 0 16344. 5 28015. 0 asteroids 1194. 16 (98. 25) 3261. 88 (2602. 48) 48027. 06 (56599.... | Is_Deep_Reinforcement_Learning_Really_Superhuman_o.pdf |
README.md exists but content is empty.
- Downloads last month
- 4