He can surpass humans in two hours! DeepMind's latest AI speedruns 26 Atari games

DeepMind’s AI agent is trying to trick itself again!

Look, this guy named BBF mastered 26 Atari games in just 2 hours. His efficiency is as good as that of a human being, surpassing all his predecessors.

You must know that AI agents have always been effective in solving problems through reinforcement learning, but the biggest problem is that this method is very inefficient and requires a long time to explore.

Picture

The breakthrough brought by BBF is precisely in terms of efficiency.

No wonder its full name can be called Bigger, Better, or Faster.

Moreover, it can complete training on only a single card, and the computing power requirements are also much reduced.

BBF was jointly proposed by Google DeepMind and the University of Montreal. The data and code are currently open source.

Can achieve up to 5 times the performance of humans

The value used to evaluate the performance of BBF games is called IQM.

IQM is a comprehensive score of multi-faceted game performance. The IQM scores in this article are normalized based on humans.

Compared with multiple previous results, BBF achieved the highest IQM score in the Atari 100K test data set containing 26 Atari games.

Moreover, in the 26 games trained, BBF's performance has exceeded that of humans.

Compared with Eff.Zero, which performs similarly, BBF consumes nearly half the GPU time.

The performance of SPR and SR-SPR, which consume similar GPU time, is significantly different from BBF.

Picture

In repeated tests, the proportion of BBF reaching a certain IQM score has always remained at a high level.

Even in more than 1/8 of the total number of tests, it achieved 5 times the performance of humans.

Picture

Even with other Atari games without training, BBF can achieve more than half the IQM score of a human.

If you look at the 29 untrained games alone, BBF's score is 40 to 50% that of humans.

Picture

Modified based on SR-SPR

The problem that promotes BBF research is how to use it when the sample size is sparse Scaling deep reinforcement learning networks.

To study this problem, DeepMind focused its attention on the Atari 100K benchmark.

But DeepMind soon discovered that simply increasing the size of the model did not improve its performance.

Picture

In the design of deep learning models, the number of updates per step (Replay Ratio, RR) is an important parameter.

Specifically for Atari games, the greater the RR value, the higher the model’s performance in the game.

In the end, DeepMind uses SR-SPR as the basic engine, and the RR value of SR-SPR can reach up to 16.

After comprehensive consideration, DeepMind chose 8 as the RR value of BBF.

Considering that some users are unwilling to spend the computing cost of RR=8, DeepMind also developed the RR=2 version of BBF

Picture

After DeepMind modified many contents in SR-SPR, it used self-supervised training to obtain BBF, which mainly includes the following aspects:

• Higher convolution layer reset strength: Increasing the convolution layer reset strength can increase the perturbation amplitude for random targets, allowing the model to perform better and reduce losses. After the reset strength of BBF is increased, The perturbation amplitude increased from 20% of SR-SPR to 50%
• Larger network scale: increasing the number of neural network layers from 3 to 15 layers, and increasing the width by 4 times
• Reduced update range (n): If you want to improve the performance of the model, you need to use non-fixed n values. BBF is reset every 40,000 gradient steps. In the first 10,000 gradient steps of each reset, n decreases exponentially from 10 to 3. The decay phase accounts for 25% of the BBF training process
• Update Large attenuation factor (γ): Some people have found that increasing the γ value during the learning process can improve model performance. The γ value of BBF increases from the traditional 0.97 to 0.997
• Weight attenuation: avoid the occurrence of overfitting, The attenuation of BBF is about 0.1
• Delete NoisyNet: NoisyNet included in the original SR-SPR cannot improve model performance

Ablation experimental results show that when the number of updates per step is 2 and Under the conditions of 8, the above factors have varying degrees of impact on the performance of BBF.

Picture

Among them, the impact of hard reset and reduction of update range is the most significant.

Picture

For NoisyNet, which is not mentioned in the above two figures, the impact on model performance is not significant.

Picture

Paper address: https://arxiv.org/abs/2305.19452GitHub project page: https://github.com /google-research/google-research/tree/master/bigger_better_faster

Reference link: [1]https://www.php.cn/link/69b4fa3be19bdf400df34e41b93636a4

[2]https://www.marktechpost.com/2023/06/12/superhuman-performance-on-the-atari-100k-benchmark-the-power-of-bbf-a-new-value -based-rl-agent-from-google-deepmind-mila-and-universite-de-montreal/

— End —

The above is the detailed content of He can surpass humans in two hours! DeepMind's latest AI speedruns 26 Atari games. For more information, please follow other related articles on the PHP Chinese website!